Upload 5 files
Browse files- configuration_kosmos2.py +331 -0
- modeling_kosmos2.py +1747 -0
- processing_kosmos2.py +498 -0
- tokenization_kosmos2.py +413 -0
- tokenization_kosmos2_fast.py +250 -0
configuration_kosmos2.py
ADDED
@@ -0,0 +1,331 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# coding=utf-8
|
2 |
+
# Copyright 2023 Microsoft Research and The HuggingFace Inc. team. All rights reserved.
|
3 |
+
#
|
4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
5 |
+
# you may not use this file except in compliance with the License.
|
6 |
+
# You may obtain a copy of the License at
|
7 |
+
#
|
8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
9 |
+
#
|
10 |
+
# Unless required by applicable law or agreed to in writing, software
|
11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
+
# See the License for the specific language governing permissions and
|
14 |
+
# limitations under the License.
|
15 |
+
""" KOSMOS-2 model configuration"""
|
16 |
+
|
17 |
+
import copy
|
18 |
+
import os
|
19 |
+
from typing import Union
|
20 |
+
|
21 |
+
from ...configuration_utils import PretrainedConfig
|
22 |
+
from ...utils import logging
|
23 |
+
|
24 |
+
|
25 |
+
logger = logging.get_logger(__name__)
|
26 |
+
|
27 |
+
BEIT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
|
28 |
+
"microsoft/kosmos-2-patch14-224": (
|
29 |
+
"https://huggingface.co/microsoft/kosmos-2-patch14-224/resolve/main/config.json"
|
30 |
+
),
|
31 |
+
# See all KOSMOS-2 models at https://huggingface.co/models?filter=kosmos-2
|
32 |
+
}
|
33 |
+
|
34 |
+
|
35 |
+
class Kosmos2TextConfig(PretrainedConfig):
|
36 |
+
r"""
|
37 |
+
This is the configuration class to store the configuration of a [`Kosmos2TextModel`]. It is used to instantiate a KOSMOS-2 text decoder
|
38 |
+
according to the specified arguments, defining the model architecture. Instantiating a configuration with the
|
39 |
+
defaults will yield a similar configuration to that of the text decoder of the KOSMOS-2
|
40 |
+
[microsoft/kosmos-2-patch14-224](https://huggingface.co/microsoft/kosmos-2-patch14-224) architecture.
|
41 |
+
|
42 |
+
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
43 |
+
documentation from [`PretrainedConfig`] for more information.
|
44 |
+
|
45 |
+
Args:
|
46 |
+
vocab_size (`int`, *optional*, defaults to 65037):
|
47 |
+
Vocabulary size of the Kosmos2 model. Defines the number of different tokens that can be represented by the
|
48 |
+
`inputs_ids` passed when calling [`Kosmos2Model`].
|
49 |
+
embed_dim (`int`, *optional*, defaults to 2048):
|
50 |
+
Dimensionality of the layers and the pooler layer.
|
51 |
+
layers (`int`, *optional*, defaults to 24):
|
52 |
+
Number of hidden layers in the Transformer encoder.
|
53 |
+
attention_heads (`int`, *optional*, defaults to 32):
|
54 |
+
Number of attention heads for each attention layer in the Transformer encoder.
|
55 |
+
ffn_dim (`int`, *optional*, defaults to 8192):
|
56 |
+
Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
|
57 |
+
activation_function (`str` or `function`, *optional*, defaults to `"gelu"`):
|
58 |
+
The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
|
59 |
+
`"relu"`, `"silu"` and `"gelu_new"` are supported.
|
60 |
+
dropout (`float`, *optional*, defaults to 0.1):
|
61 |
+
The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
|
62 |
+
attention_dropout (`float`, *optional*, defaults to 0.1):
|
63 |
+
The dropout ratio for the attention probabilities.
|
64 |
+
activation_dropout (`float`, *optional*, defaults to 0.0):
|
65 |
+
The dropout ratio for activations inside the fully connected layer.
|
66 |
+
max_position_embeddings (`int`, *optional*, defaults to 2048):
|
67 |
+
The maximum sequence length that this model might ever be used with. Typically set this to something large
|
68 |
+
just in case (e.g., 512 or 1024 or 2048).
|
69 |
+
layerdrop (`float`, *optional*, defaults to 0.0):
|
70 |
+
The LayerDrop probability for the decoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
|
71 |
+
for more details.
|
72 |
+
layer_norm_eps (`float`, *optional*, defaults to 1e-5):
|
73 |
+
The epsilon used by the layer normalization layers.
|
74 |
+
scale_embedding (`bool`, *optional*, defaults to `True`):
|
75 |
+
Scale embeddings by diving by sqrt(embed_dim).
|
76 |
+
use_cache (`bool`, *optional*, defaults to `True`):
|
77 |
+
Whether or not the model should return the last key/values attentions (not used by all models).
|
78 |
+
|
79 |
+
Example:
|
80 |
+
|
81 |
+
```python
|
82 |
+
>>> from transformers import Kosmos2TextConfig, Kosmos2TextModel
|
83 |
+
|
84 |
+
>>> # Initializing a Kosmos2TextConfig microsoft/kosmos-2-patch14-224 style configuration
|
85 |
+
>>> configuration = Kosmos2TextConfig()
|
86 |
+
|
87 |
+
>>> # Initializing a Kosmos2TextModel (with random weights) from the microsoft/kosmos-2-patch14-224 style configuration
|
88 |
+
>>> model = Kosmos2TextModel(configuration)
|
89 |
+
|
90 |
+
>>> # Accessing the model configuration
|
91 |
+
>>> configuration = model.config
|
92 |
+
```"""
|
93 |
+
model_type = "kosmos_2_text_model"
|
94 |
+
keys_to_ignore_at_inference = ["past_key_values"]
|
95 |
+
attribute_map = {"num_attention_heads": "attention_heads", "hidden_size": "embed_dim"}
|
96 |
+
|
97 |
+
def __init__(
|
98 |
+
self,
|
99 |
+
vocab_size=65037,
|
100 |
+
max_position_embeddings=2048,
|
101 |
+
embed_dim=2048,
|
102 |
+
layers=24,
|
103 |
+
ffn_dim=8192,
|
104 |
+
attention_heads=32,
|
105 |
+
activation_function="gelu",
|
106 |
+
dropout=0.1,
|
107 |
+
attention_dropout=0.1,
|
108 |
+
activation_dropout=0.0,
|
109 |
+
layerdrop=0.0,
|
110 |
+
layer_norm_eps=1e-5,
|
111 |
+
scale_embedding=True,
|
112 |
+
use_cache=True,
|
113 |
+
pad_token_id=1,
|
114 |
+
bos_token_id=0,
|
115 |
+
eos_token_id=2,
|
116 |
+
**kwargs,
|
117 |
+
):
|
118 |
+
super().__init__(
|
119 |
+
pad_token_id=pad_token_id,
|
120 |
+
bos_token_id=bos_token_id,
|
121 |
+
eos_token_id=eos_token_id,
|
122 |
+
**kwargs,
|
123 |
+
)
|
124 |
+
|
125 |
+
self.vocab_size = vocab_size
|
126 |
+
self.max_position_embeddings = max_position_embeddings
|
127 |
+
self.embed_dim = embed_dim
|
128 |
+
self.layers = layers
|
129 |
+
self.ffn_dim = ffn_dim
|
130 |
+
self.attention_heads = attention_heads
|
131 |
+
self.activation_function = activation_function
|
132 |
+
self.dropout = dropout
|
133 |
+
self.attention_dropout = attention_dropout
|
134 |
+
self.activation_dropout = activation_dropout
|
135 |
+
self.layerdrop = layerdrop
|
136 |
+
self.layer_norm_eps = layer_norm_eps
|
137 |
+
self.scale_embedding = scale_embedding
|
138 |
+
self.use_cache = use_cache
|
139 |
+
|
140 |
+
@classmethod
|
141 |
+
def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
|
142 |
+
cls._set_token_in_kwargs(kwargs)
|
143 |
+
|
144 |
+
config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
|
145 |
+
|
146 |
+
# get the text config dict if we are loading from Kosmos2Config
|
147 |
+
if config_dict.get("model_type") == "kosmos-2":
|
148 |
+
config_dict = config_dict["text_config"]
|
149 |
+
|
150 |
+
if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
|
151 |
+
logger.warning(
|
152 |
+
f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
|
153 |
+
f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
|
154 |
+
)
|
155 |
+
|
156 |
+
return cls.from_dict(config_dict, **kwargs)
|
157 |
+
|
158 |
+
|
159 |
+
class Kosmos2VisionConfig(PretrainedConfig):
|
160 |
+
r"""
|
161 |
+
This is the configuration class to store the configuration of a [`Kosmos2VisionModel`]. It is used to instantiate a
|
162 |
+
KOSMOS-2 vision encoder according to the specified arguments, defining the model architecture. Instantiating a
|
163 |
+
configuration with the defaults will yield a similar configuration to that of the vision encoder of the KOSMOS-2
|
164 |
+
[microsoft/kosmos-2-patch14-224](https://huggingface.co/microsoft/kosmos-2-patch14-224) architecture.
|
165 |
+
|
166 |
+
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
167 |
+
documentation from [`PretrainedConfig`] for more information.
|
168 |
+
|
169 |
+
Args:
|
170 |
+
hidden_size (`int`, *optional*, defaults to 1024):
|
171 |
+
Dimensionality of the encoder layers and the pooler layer.
|
172 |
+
intermediate_size (`int`, *optional*, defaults to 4096):
|
173 |
+
Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
|
174 |
+
num_hidden_layers (`int`, *optional*, defaults to 24):
|
175 |
+
Number of hidden layers in the Transformer encoder.
|
176 |
+
num_attention_heads (`int`, *optional*, defaults to 16):
|
177 |
+
Number of attention heads for each attention layer in the Transformer encoder.
|
178 |
+
image_size (`int`, *optional*, defaults to 224):
|
179 |
+
The size (resolution) of each image.
|
180 |
+
patch_size (`int`, *optional*, defaults to 14):
|
181 |
+
The size (resolution) of each patch.
|
182 |
+
hidden_act (`str` or `function`, *optional*, defaults to `"quick_gelu"`):
|
183 |
+
The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
|
184 |
+
`"relu"`, `"selu"` and `"gelu_new"` ``"quick_gelu"` are supported.
|
185 |
+
layer_norm_eps (`float`, *optional*, defaults to 1e-5):
|
186 |
+
The epsilon used by the layer normalization layers.
|
187 |
+
attention_dropout (`float`, *optional*, defaults to 0.0):
|
188 |
+
The dropout ratio for the attention probabilities.
|
189 |
+
initializer_range (`float`, *optional*, defaults to 0.02):
|
190 |
+
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
|
191 |
+
initializer_factor (`float`, *optional*, defaults to 1):
|
192 |
+
A factor for initializing all weight matrices (should be kept to 1, used internally for initialization
|
193 |
+
testing).
|
194 |
+
|
195 |
+
Example:
|
196 |
+
|
197 |
+
```python
|
198 |
+
>>> from transformers import Kosmos2VisionConfig, Kosmos2VisionModel
|
199 |
+
|
200 |
+
>>> # Initializing a Kosmos2VisionConfig with microsoft/kosmos-2-patch14-224 style configuration
|
201 |
+
>>> configuration = Kosmos2VisionConfig()
|
202 |
+
|
203 |
+
>>> # Initializing a Kosmos2VisionModel (with random weights) from the microsoft/kosmos-2-patch14-224 style configuration
|
204 |
+
>>> model = Kosmos2VisionModel(configuration)
|
205 |
+
|
206 |
+
>>> # Accessing the model configuration
|
207 |
+
>>> configuration = model.config
|
208 |
+
```"""
|
209 |
+
|
210 |
+
model_type = "kosmos_2_vision_model"
|
211 |
+
|
212 |
+
def __init__(
|
213 |
+
self,
|
214 |
+
hidden_size=1024,
|
215 |
+
intermediate_size=4096,
|
216 |
+
projection_dim=512,
|
217 |
+
num_hidden_layers=24,
|
218 |
+
num_attention_heads=16,
|
219 |
+
num_channels=3,
|
220 |
+
image_size=224,
|
221 |
+
patch_size=14,
|
222 |
+
hidden_act="quick_gelu",
|
223 |
+
layer_norm_eps=1e-5,
|
224 |
+
attention_dropout=0.0,
|
225 |
+
initializer_range=0.02,
|
226 |
+
initializer_factor=1.0,
|
227 |
+
**kwargs,
|
228 |
+
):
|
229 |
+
super().__init__(**kwargs)
|
230 |
+
|
231 |
+
self.hidden_size = hidden_size
|
232 |
+
self.intermediate_size = intermediate_size
|
233 |
+
self.projection_dim = projection_dim
|
234 |
+
self.num_hidden_layers = num_hidden_layers
|
235 |
+
self.num_attention_heads = num_attention_heads
|
236 |
+
self.num_channels = num_channels
|
237 |
+
self.patch_size = patch_size
|
238 |
+
self.image_size = image_size
|
239 |
+
self.initializer_range = initializer_range
|
240 |
+
self.initializer_factor = initializer_factor
|
241 |
+
self.attention_dropout = attention_dropout
|
242 |
+
self.layer_norm_eps = layer_norm_eps
|
243 |
+
self.hidden_act = hidden_act
|
244 |
+
|
245 |
+
@classmethod
|
246 |
+
def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
|
247 |
+
cls._set_token_in_kwargs(kwargs)
|
248 |
+
|
249 |
+
config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
|
250 |
+
|
251 |
+
# get the vision config dict if we are loading from Kosmos2Config
|
252 |
+
if config_dict.get("model_type") == "kosmos-2":
|
253 |
+
config_dict = config_dict["vision_config"]
|
254 |
+
|
255 |
+
if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
|
256 |
+
logger.warning(
|
257 |
+
f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
|
258 |
+
f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
|
259 |
+
)
|
260 |
+
|
261 |
+
return cls.from_dict(config_dict, **kwargs)
|
262 |
+
|
263 |
+
|
264 |
+
class Kosmos2Config(PretrainedConfig):
|
265 |
+
r"""
|
266 |
+
This is the configuration class to store the configuration of a [`Kosmos2Model`]. It is used to instantiate a KOSMOS-2
|
267 |
+
model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
|
268 |
+
defaults will yield a similar configuration to that of the KOSMOS-2
|
269 |
+
[microsoft/kosmos-2-patch14-224](https://huggingface.co/microsoft/kosmos-2-patch14-224) architecture.
|
270 |
+
|
271 |
+
Args:
|
272 |
+
text_config (`dict`, *optional*):
|
273 |
+
Dictionary of configuration options used to initialize [`Kosmos2TextConfig`].
|
274 |
+
vision_config (`dict`, *optional*):
|
275 |
+
Dictionary of configuration options used to initialize [`Kosmos2VisionConfig`].
|
276 |
+
latent_query_num (`int`, *optional*, defaults to 64):
|
277 |
+
The number of latent query tokens that represent the image features used in the text decoder component.
|
278 |
+
kwargs (*optional*):
|
279 |
+
Dictionary of keyword arguments.
|
280 |
+
|
281 |
+
Example:
|
282 |
+
|
283 |
+
```python
|
284 |
+
>>> from transformers import Kosmos2Config, Kosmos2Model
|
285 |
+
|
286 |
+
>>> # Initializing a Kosmos-2 kosmos-2-patch14-224 style configuration
|
287 |
+
>>> configuration = Kosmos2Config()
|
288 |
+
|
289 |
+
>>> # Initializing a model (with random weights) from the kosmos-2-patch14-224 style configuration
|
290 |
+
>>> model = Kosmos2Model(configuration)
|
291 |
+
|
292 |
+
>>> # Accessing the model configuration
|
293 |
+
>>> configuration = model.config
|
294 |
+
```"""
|
295 |
+
model_type = "kosmos-2"
|
296 |
+
is_composition = True
|
297 |
+
|
298 |
+
def __init__(
|
299 |
+
self,
|
300 |
+
text_config=None,
|
301 |
+
vision_config=None,
|
302 |
+
latent_query_num=64,
|
303 |
+
**kwargs,
|
304 |
+
):
|
305 |
+
super().__init__(**kwargs)
|
306 |
+
|
307 |
+
if text_config is None:
|
308 |
+
text_config = {}
|
309 |
+
logger.info("`text_config` is `None`. Initializing the `Kosmos2TextConfig` with default values.")
|
310 |
+
|
311 |
+
if vision_config is None:
|
312 |
+
vision_config = {}
|
313 |
+
logger.info("`vision_config` is `None`. Initializing the `Kosmos2VisionConfig` with default values.")
|
314 |
+
|
315 |
+
self.text_config = Kosmos2TextConfig(**text_config)
|
316 |
+
self.vision_config = Kosmos2VisionConfig(**vision_config)
|
317 |
+
|
318 |
+
self.latent_query_num = latent_query_num
|
319 |
+
|
320 |
+
def to_dict(self):
|
321 |
+
"""
|
322 |
+
Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
|
323 |
+
|
324 |
+
Returns:
|
325 |
+
`Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
|
326 |
+
"""
|
327 |
+
output = copy.deepcopy(self.__dict__)
|
328 |
+
output["text_config"] = self.text_config.to_dict()
|
329 |
+
output["vision_config"] = self.vision_config.to_dict()
|
330 |
+
output["model_type"] = self.__class__.model_type
|
331 |
+
return output
|
modeling_kosmos2.py
ADDED
@@ -0,0 +1,1747 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# coding=utf-8
|
2 |
+
# Copyright 2023 Microsoft Research and The HuggingFace Inc. team. All rights reserved.
|
3 |
+
#
|
4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
5 |
+
# you may not use this file except in compliance with the License.
|
6 |
+
# You may obtain a copy of the License at
|
7 |
+
#
|
8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
9 |
+
#
|
10 |
+
# Unless required by applicable law or agreed to in writing, software
|
11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
+
# See the License for the specific language governing permissions and
|
14 |
+
# limitations under the License.
|
15 |
+
""" PyTorch KOSMOS-2 model."""
|
16 |
+
|
17 |
+
|
18 |
+
import math
|
19 |
+
from dataclasses import dataclass
|
20 |
+
from typing import List, Optional, Tuple, Union
|
21 |
+
|
22 |
+
import torch
|
23 |
+
import torch.utils.checkpoint
|
24 |
+
from torch import nn
|
25 |
+
|
26 |
+
from ...activations import ACT2FN
|
27 |
+
from ...modeling_outputs import (
|
28 |
+
BaseModelOutput,
|
29 |
+
BaseModelOutputWithPastAndCrossAttentions,
|
30 |
+
BaseModelOutputWithPooling,
|
31 |
+
CausalLMOutputWithCrossAttentions,
|
32 |
+
)
|
33 |
+
from ...modeling_utils import PreTrainedModel
|
34 |
+
from ...utils import (
|
35 |
+
ModelOutput,
|
36 |
+
add_start_docstrings,
|
37 |
+
add_start_docstrings_to_model_forward,
|
38 |
+
logging,
|
39 |
+
replace_return_docstrings,
|
40 |
+
)
|
41 |
+
from .configuration_kosmos2 import Kosmos2Config, Kosmos2TextConfig, Kosmos2VisionConfig
|
42 |
+
|
43 |
+
|
44 |
+
logger = logging.get_logger(__name__)
|
45 |
+
|
46 |
+
_CHECKPOINT_FOR_DOC = "microsoft/kosmos-2-patch14-224"
|
47 |
+
_CONFIG_FOR_DOC = Kosmos2Config
|
48 |
+
_EXPECTED_OUTPUT_SHAPE = None
|
49 |
+
|
50 |
+
|
51 |
+
# Copied from transformers.models.bart.modeling_bart._expand_mask
|
52 |
+
def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
|
53 |
+
"""
|
54 |
+
Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
|
55 |
+
"""
|
56 |
+
bsz, src_len = mask.size()
|
57 |
+
tgt_len = tgt_len if tgt_len is not None else src_len
|
58 |
+
|
59 |
+
expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
|
60 |
+
|
61 |
+
inverted_mask = 1.0 - expanded_mask
|
62 |
+
|
63 |
+
return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
|
64 |
+
|
65 |
+
|
66 |
+
# Copied from transformers.models.bart.modeling_bart._make_causal_mask
|
67 |
+
def _make_causal_mask(
|
68 |
+
input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
|
69 |
+
):
|
70 |
+
"""
|
71 |
+
Make causal mask used for bi-directional self-attention.
|
72 |
+
"""
|
73 |
+
bsz, tgt_len = input_ids_shape
|
74 |
+
mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min, device=device), device=device)
|
75 |
+
mask_cond = torch.arange(mask.size(-1), device=device)
|
76 |
+
mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
|
77 |
+
mask = mask.to(dtype)
|
78 |
+
|
79 |
+
if past_key_values_length > 0:
|
80 |
+
mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
|
81 |
+
return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
|
82 |
+
|
83 |
+
|
84 |
+
# Copied from transformers.models.roberta.modeling_roberta.create_position_ids_from_input_ids
|
85 |
+
def create_position_ids_from_input_ids(input_ids, padding_idx, past_key_values_length=0):
|
86 |
+
"""
|
87 |
+
Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding symbols
|
88 |
+
are ignored. This is modified from fairseq's `utils.make_positions`.
|
89 |
+
|
90 |
+
Args:
|
91 |
+
x: torch.Tensor x:
|
92 |
+
|
93 |
+
Returns: torch.Tensor
|
94 |
+
"""
|
95 |
+
# The series of casts and type-conversions here are carefully balanced to both work with ONNX export and XLA.
|
96 |
+
mask = input_ids.ne(padding_idx).int()
|
97 |
+
incremental_indices = (torch.cumsum(mask, dim=1).type_as(mask) + past_key_values_length) * mask
|
98 |
+
return incremental_indices.long() + padding_idx
|
99 |
+
|
100 |
+
|
101 |
+
KOSMOS2_START_DOCSTRING = r"""Kosmos-2"""
|
102 |
+
KOSMOS2_VISION_INPUTS_DOCSTRING = r"""Kosmos-2"""
|
103 |
+
KOSMOS2_TEXT_INPUTS_DOCSTRING = r"""Kosmos-2"""
|
104 |
+
KOSMOS2_INPUTS_DOCSTRING = r"""Kosmos-2"""
|
105 |
+
|
106 |
+
|
107 |
+
@dataclass
|
108 |
+
class Kosmos2ModelOutput(ModelOutput):
|
109 |
+
"""
|
110 |
+
Base class for text model's outputs that also contains a pooling of the last hidden states.
|
111 |
+
|
112 |
+
Args:
|
113 |
+
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
|
114 |
+
Sequence of hidden-states at the output of the last layer of the model.
|
115 |
+
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
|
116 |
+
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
|
117 |
+
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
|
118 |
+
|
119 |
+
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
|
120 |
+
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
|
121 |
+
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
|
122 |
+
sequence_length)`.
|
123 |
+
|
124 |
+
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
|
125 |
+
heads.
|
126 |
+
image_features (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*, returned when being computed by the model):
|
127 |
+
Sequence of hidden-states at the output of `Kosmos2ImageToTextConnector`.
|
128 |
+
image_connector_attention (`tuple(torch.FloatTensor)`, *optional, returned when being computed by the model):
|
129 |
+
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
|
130 |
+
sequence_length)`.
|
131 |
+
|
132 |
+
Attentions weights given by `Kosmos2ImageToTextConnector`, after the attention softmax, used to compute the weighted average in the self-attention
|
133 |
+
heads.
|
134 |
+
vision_model_output(`BaseModelOutputWithPooling`, *optional*, returned when being computed by the model):
|
135 |
+
The output of the [`Kosmos2VisionModel`].
|
136 |
+
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
|
137 |
+
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
|
138 |
+
`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and optionally if
|
139 |
+
`config.is_encoder_decoder=True` 2 additional tensors of shape `(batch_size, num_heads,
|
140 |
+
encoder_sequence_length, embed_size_per_head)`.
|
141 |
+
|
142 |
+
Contains pre-computed hidden-states (key and values in the self-attention blocks and optionally if
|
143 |
+
`config.is_encoder_decoder=True` in the cross-attention blocks) that can be used (see `past_key_values`
|
144 |
+
input) to speed up sequential decoding.
|
145 |
+
"""
|
146 |
+
|
147 |
+
last_hidden_states: torch.FloatTensor = None
|
148 |
+
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
|
149 |
+
hidden_states: Optional[Tuple[torch.FloatTensor]] = None
|
150 |
+
attentions: Optional[Tuple[torch.FloatTensor]] = None
|
151 |
+
image_features: Optional[torch.FloatTensor] = None
|
152 |
+
image_connector_attention: Optional[Tuple[torch.FloatTensor]] = None
|
153 |
+
vision_model_output: BaseModelOutputWithPooling = None
|
154 |
+
|
155 |
+
|
156 |
+
@dataclass
|
157 |
+
class Kosmos2ForConditionalGenerationModelOutput(ModelOutput):
|
158 |
+
"""
|
159 |
+
Model output class for `Kosmos2ForConditionalGeneration`.
|
160 |
+
|
161 |
+
Args:
|
162 |
+
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
|
163 |
+
Language modeling loss (for next-token prediction).
|
164 |
+
logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
|
165 |
+
Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
|
166 |
+
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
|
167 |
+
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
|
168 |
+
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
|
169 |
+
|
170 |
+
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
|
171 |
+
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
|
172 |
+
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
|
173 |
+
sequence_length)`.
|
174 |
+
|
175 |
+
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
|
176 |
+
heads.
|
177 |
+
image_features (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*, returned when being computed by the model):
|
178 |
+
Sequence of hidden-states at the output of `Kosmos2ImageToTextConnector`.
|
179 |
+
image_connector_attention (`tuple(torch.FloatTensor)`, *optional, returned when being computed by the model):
|
180 |
+
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
|
181 |
+
sequence_length)`.
|
182 |
+
|
183 |
+
Attentions weights given by `Kosmos2ImageToTextConnector`, after the attention softmax, used to compute the weighted average in the self-attention
|
184 |
+
heads.
|
185 |
+
vision_model_output(`BaseModelOutputWithPooling`, *optional*, returned when being computed by the model):
|
186 |
+
The output of the [`Kosmos2VisionModel`].
|
187 |
+
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
|
188 |
+
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
|
189 |
+
`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and optionally if
|
190 |
+
`config.is_encoder_decoder=True` 2 additional tensors of shape `(batch_size, num_heads,
|
191 |
+
encoder_sequence_length, embed_size_per_head)`.
|
192 |
+
|
193 |
+
Contains pre-computed hidden-states (key and values in the self-attention blocks and optionally if
|
194 |
+
`config.is_encoder_decoder=True` in the cross-attention blocks) that can be used (see `past_key_values`
|
195 |
+
input) to speed up sequential decoding.
|
196 |
+
"""
|
197 |
+
|
198 |
+
loss: Optional[torch.FloatTensor] = None
|
199 |
+
logits: torch.FloatTensor = None
|
200 |
+
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
|
201 |
+
hidden_states: Optional[Tuple[torch.FloatTensor]] = None
|
202 |
+
attentions: Optional[Tuple[torch.FloatTensor]] = None
|
203 |
+
image_features: Optional[torch.FloatTensor] = None
|
204 |
+
image_connector_attention: Optional[Tuple[torch.FloatTensor]] = None
|
205 |
+
vision_model_output: BaseModelOutputWithPooling = None
|
206 |
+
|
207 |
+
|
208 |
+
# Copied from transformers.models.clip.modeling_clip.CLIPVisionEmbeddings with CLIP->Kosmos2
|
209 |
+
class Kosmos2VisionEmbeddings(nn.Module):
|
210 |
+
def __init__(self, config: Kosmos2VisionConfig):
|
211 |
+
super().__init__()
|
212 |
+
self.config = config
|
213 |
+
self.embed_dim = config.hidden_size
|
214 |
+
self.image_size = config.image_size
|
215 |
+
self.patch_size = config.patch_size
|
216 |
+
|
217 |
+
self.class_embedding = nn.Parameter(torch.randn(self.embed_dim))
|
218 |
+
|
219 |
+
self.patch_embedding = nn.Conv2d(
|
220 |
+
in_channels=config.num_channels,
|
221 |
+
out_channels=self.embed_dim,
|
222 |
+
kernel_size=self.patch_size,
|
223 |
+
stride=self.patch_size,
|
224 |
+
bias=False,
|
225 |
+
)
|
226 |
+
|
227 |
+
self.num_patches = (self.image_size // self.patch_size) ** 2
|
228 |
+
self.num_positions = self.num_patches + 1
|
229 |
+
self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
|
230 |
+
self.register_buffer("position_ids", torch.arange(self.num_positions).expand((1, -1)), persistent=False)
|
231 |
+
|
232 |
+
def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
|
233 |
+
batch_size = pixel_values.shape[0]
|
234 |
+
patch_embeds = self.patch_embedding(pixel_values) # shape = [*, width, grid, grid]
|
235 |
+
patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
|
236 |
+
|
237 |
+
class_embeds = self.class_embedding.expand(batch_size, 1, -1)
|
238 |
+
embeddings = torch.cat([class_embeds, patch_embeds], dim=1)
|
239 |
+
embeddings = embeddings + self.position_embedding(self.position_ids)
|
240 |
+
return embeddings
|
241 |
+
|
242 |
+
|
243 |
+
# Copied from transformers.models.clip.modeling_clip.CLIPAttention with CLIP->Kosmos2Vision
|
244 |
+
class Kosmos2VisionAttention(nn.Module):
|
245 |
+
"""Multi-headed attention from 'Attention Is All You Need' paper"""
|
246 |
+
|
247 |
+
def __init__(self, config):
|
248 |
+
super().__init__()
|
249 |
+
self.config = config
|
250 |
+
self.embed_dim = config.hidden_size
|
251 |
+
self.num_heads = config.num_attention_heads
|
252 |
+
self.head_dim = self.embed_dim // self.num_heads
|
253 |
+
if self.head_dim * self.num_heads != self.embed_dim:
|
254 |
+
raise ValueError(
|
255 |
+
f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
|
256 |
+
f" {self.num_heads})."
|
257 |
+
)
|
258 |
+
self.scale = self.head_dim**-0.5
|
259 |
+
self.dropout = config.attention_dropout
|
260 |
+
|
261 |
+
self.k_proj = nn.Linear(self.embed_dim, self.embed_dim)
|
262 |
+
self.v_proj = nn.Linear(self.embed_dim, self.embed_dim)
|
263 |
+
self.q_proj = nn.Linear(self.embed_dim, self.embed_dim)
|
264 |
+
self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
|
265 |
+
|
266 |
+
def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
|
267 |
+
return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
|
268 |
+
|
269 |
+
def forward(
|
270 |
+
self,
|
271 |
+
hidden_states: torch.Tensor,
|
272 |
+
attention_mask: Optional[torch.Tensor] = None,
|
273 |
+
causal_attention_mask: Optional[torch.Tensor] = None,
|
274 |
+
output_attentions: Optional[bool] = False,
|
275 |
+
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
276 |
+
"""Input shape: Batch x Time x Channel"""
|
277 |
+
|
278 |
+
bsz, tgt_len, embed_dim = hidden_states.size()
|
279 |
+
|
280 |
+
# get query proj
|
281 |
+
query_states = self.q_proj(hidden_states) * self.scale
|
282 |
+
key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
|
283 |
+
value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
|
284 |
+
|
285 |
+
proj_shape = (bsz * self.num_heads, -1, self.head_dim)
|
286 |
+
query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
|
287 |
+
key_states = key_states.view(*proj_shape)
|
288 |
+
value_states = value_states.view(*proj_shape)
|
289 |
+
|
290 |
+
src_len = key_states.size(1)
|
291 |
+
attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
|
292 |
+
|
293 |
+
if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
|
294 |
+
raise ValueError(
|
295 |
+
f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
|
296 |
+
f" {attn_weights.size()}"
|
297 |
+
)
|
298 |
+
|
299 |
+
# apply the causal_attention_mask first
|
300 |
+
if causal_attention_mask is not None:
|
301 |
+
if causal_attention_mask.size() != (bsz, 1, tgt_len, src_len):
|
302 |
+
raise ValueError(
|
303 |
+
f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is"
|
304 |
+
f" {causal_attention_mask.size()}"
|
305 |
+
)
|
306 |
+
attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + causal_attention_mask
|
307 |
+
attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
|
308 |
+
|
309 |
+
if attention_mask is not None:
|
310 |
+
if attention_mask.size() != (bsz, 1, tgt_len, src_len):
|
311 |
+
raise ValueError(
|
312 |
+
f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
|
313 |
+
)
|
314 |
+
attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
|
315 |
+
attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
|
316 |
+
|
317 |
+
attn_weights = nn.functional.softmax(attn_weights, dim=-1)
|
318 |
+
|
319 |
+
if output_attentions:
|
320 |
+
# this operation is a bit akward, but it's required to
|
321 |
+
# make sure that attn_weights keeps its gradient.
|
322 |
+
# In order to do so, attn_weights have to reshaped
|
323 |
+
# twice and have to be reused in the following
|
324 |
+
attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
|
325 |
+
attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
|
326 |
+
else:
|
327 |
+
attn_weights_reshaped = None
|
328 |
+
|
329 |
+
attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
|
330 |
+
|
331 |
+
attn_output = torch.bmm(attn_probs, value_states)
|
332 |
+
|
333 |
+
if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
|
334 |
+
raise ValueError(
|
335 |
+
f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
|
336 |
+
f" {attn_output.size()}"
|
337 |
+
)
|
338 |
+
|
339 |
+
attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
|
340 |
+
attn_output = attn_output.transpose(1, 2)
|
341 |
+
attn_output = attn_output.reshape(bsz, tgt_len, embed_dim)
|
342 |
+
|
343 |
+
attn_output = self.out_proj(attn_output)
|
344 |
+
|
345 |
+
return attn_output, attn_weights_reshaped
|
346 |
+
|
347 |
+
|
348 |
+
# Copied from transformers.models.clip.modeling_clip.CLIPMLP with CLIP->Kosmos2Vision
|
349 |
+
class Kosmos2VisionMLP(nn.Module):
|
350 |
+
def __init__(self, config):
|
351 |
+
super().__init__()
|
352 |
+
self.config = config
|
353 |
+
self.activation_fn = ACT2FN[config.hidden_act]
|
354 |
+
self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
|
355 |
+
self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
|
356 |
+
|
357 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
358 |
+
hidden_states = self.fc1(hidden_states)
|
359 |
+
hidden_states = self.activation_fn(hidden_states)
|
360 |
+
hidden_states = self.fc2(hidden_states)
|
361 |
+
return hidden_states
|
362 |
+
|
363 |
+
|
364 |
+
# Copied from transformers.models.clip.modeling_clip.CLIPEncoderLayer with CLIP->Kosmos2Vision
|
365 |
+
class Kosmos2VisionEncoderLayer(nn.Module):
|
366 |
+
def __init__(self, config: Kosmos2VisionConfig):
|
367 |
+
super().__init__()
|
368 |
+
self.embed_dim = config.hidden_size
|
369 |
+
self.self_attn = Kosmos2VisionAttention(config)
|
370 |
+
self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
|
371 |
+
self.mlp = Kosmos2VisionMLP(config)
|
372 |
+
self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
|
373 |
+
|
374 |
+
def forward(
|
375 |
+
self,
|
376 |
+
hidden_states: torch.Tensor,
|
377 |
+
attention_mask: torch.Tensor,
|
378 |
+
causal_attention_mask: torch.Tensor,
|
379 |
+
output_attentions: Optional[bool] = False,
|
380 |
+
) -> Tuple[torch.FloatTensor]:
|
381 |
+
"""
|
382 |
+
Args:
|
383 |
+
hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
|
384 |
+
attention_mask (`torch.FloatTensor`): attention mask of size
|
385 |
+
`(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
|
386 |
+
`(config.encoder_attention_heads,)`.
|
387 |
+
output_attentions (`bool`, *optional*):
|
388 |
+
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
|
389 |
+
returned tensors for more detail.
|
390 |
+
"""
|
391 |
+
residual = hidden_states
|
392 |
+
|
393 |
+
hidden_states = self.layer_norm1(hidden_states)
|
394 |
+
hidden_states, attn_weights = self.self_attn(
|
395 |
+
hidden_states=hidden_states,
|
396 |
+
attention_mask=attention_mask,
|
397 |
+
causal_attention_mask=causal_attention_mask,
|
398 |
+
output_attentions=output_attentions,
|
399 |
+
)
|
400 |
+
hidden_states = residual + hidden_states
|
401 |
+
|
402 |
+
residual = hidden_states
|
403 |
+
hidden_states = self.layer_norm2(hidden_states)
|
404 |
+
hidden_states = self.mlp(hidden_states)
|
405 |
+
hidden_states = residual + hidden_states
|
406 |
+
|
407 |
+
outputs = (hidden_states,)
|
408 |
+
|
409 |
+
if output_attentions:
|
410 |
+
outputs += (attn_weights,)
|
411 |
+
|
412 |
+
return outputs
|
413 |
+
|
414 |
+
|
415 |
+
# Copied from transformers.models.clip.modeling_clip.CLIPEncoder with CLIP->Kosmos2Vision
|
416 |
+
class Kosmos2VisionEncoder(nn.Module):
|
417 |
+
"""
|
418 |
+
Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
|
419 |
+
[`Kosmos2VisionEncoderLayer`].
|
420 |
+
|
421 |
+
Args:
|
422 |
+
config: Kosmos2VisionConfig
|
423 |
+
"""
|
424 |
+
|
425 |
+
def __init__(self, config: Kosmos2VisionConfig):
|
426 |
+
super().__init__()
|
427 |
+
self.config = config
|
428 |
+
self.layers = nn.ModuleList([Kosmos2VisionEncoderLayer(config) for _ in range(config.num_hidden_layers)])
|
429 |
+
self.gradient_checkpointing = False
|
430 |
+
|
431 |
+
def forward(
|
432 |
+
self,
|
433 |
+
inputs_embeds,
|
434 |
+
attention_mask: Optional[torch.Tensor] = None,
|
435 |
+
causal_attention_mask: Optional[torch.Tensor] = None,
|
436 |
+
output_attentions: Optional[bool] = None,
|
437 |
+
output_hidden_states: Optional[bool] = None,
|
438 |
+
return_dict: Optional[bool] = None,
|
439 |
+
) -> Union[Tuple, BaseModelOutput]:
|
440 |
+
r"""
|
441 |
+
Args:
|
442 |
+
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
|
443 |
+
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
|
444 |
+
This is useful if you want more control over how to convert `input_ids` indices into associated vectors
|
445 |
+
than the model's internal embedding lookup matrix.
|
446 |
+
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
|
447 |
+
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
|
448 |
+
|
449 |
+
- 1 for tokens that are **not masked**,
|
450 |
+
- 0 for tokens that are **masked**.
|
451 |
+
|
452 |
+
[What are attention masks?](../glossary#attention-mask)
|
453 |
+
causal_attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
|
454 |
+
Causal mask for the text model. Mask values selected in `[0, 1]`:
|
455 |
+
|
456 |
+
- 1 for tokens that are **not masked**,
|
457 |
+
- 0 for tokens that are **masked**.
|
458 |
+
|
459 |
+
[What are attention masks?](../glossary#attention-mask)
|
460 |
+
output_attentions (`bool`, *optional*):
|
461 |
+
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
|
462 |
+
returned tensors for more detail.
|
463 |
+
output_hidden_states (`bool`, *optional*):
|
464 |
+
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
|
465 |
+
for more detail.
|
466 |
+
return_dict (`bool`, *optional*):
|
467 |
+
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
|
468 |
+
"""
|
469 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
470 |
+
output_hidden_states = (
|
471 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
472 |
+
)
|
473 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
474 |
+
|
475 |
+
encoder_states = () if output_hidden_states else None
|
476 |
+
all_attentions = () if output_attentions else None
|
477 |
+
|
478 |
+
hidden_states = inputs_embeds
|
479 |
+
for idx, encoder_layer in enumerate(self.layers):
|
480 |
+
if output_hidden_states:
|
481 |
+
encoder_states = encoder_states + (hidden_states,)
|
482 |
+
if self.gradient_checkpointing and self.training:
|
483 |
+
|
484 |
+
def create_custom_forward(module):
|
485 |
+
def custom_forward(*inputs):
|
486 |
+
return module(*inputs, output_attentions)
|
487 |
+
|
488 |
+
return custom_forward
|
489 |
+
|
490 |
+
layer_outputs = torch.utils.checkpoint.checkpoint(
|
491 |
+
create_custom_forward(encoder_layer),
|
492 |
+
hidden_states,
|
493 |
+
attention_mask,
|
494 |
+
causal_attention_mask,
|
495 |
+
)
|
496 |
+
else:
|
497 |
+
layer_outputs = encoder_layer(
|
498 |
+
hidden_states,
|
499 |
+
attention_mask,
|
500 |
+
causal_attention_mask,
|
501 |
+
output_attentions=output_attentions,
|
502 |
+
)
|
503 |
+
|
504 |
+
hidden_states = layer_outputs[0]
|
505 |
+
|
506 |
+
if output_attentions:
|
507 |
+
all_attentions = all_attentions + (layer_outputs[1],)
|
508 |
+
|
509 |
+
if output_hidden_states:
|
510 |
+
encoder_states = encoder_states + (hidden_states,)
|
511 |
+
|
512 |
+
if not return_dict:
|
513 |
+
return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
|
514 |
+
return BaseModelOutput(
|
515 |
+
last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
|
516 |
+
)
|
517 |
+
|
518 |
+
|
519 |
+
# Copied from transformers.models.clip.modeling_clip.CLIPVisionTransformer with CLIPVision->Kosmos2Vision,CLIP_VISION->KOSMOS2_VISION,CLIP->Kosmos2Vision
|
520 |
+
class Kosmos2VisionTransformer(nn.Module):
|
521 |
+
def __init__(self, config: Kosmos2VisionConfig):
|
522 |
+
super().__init__()
|
523 |
+
self.config = config
|
524 |
+
embed_dim = config.hidden_size
|
525 |
+
|
526 |
+
self.embeddings = Kosmos2VisionEmbeddings(config)
|
527 |
+
self.pre_layrnorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
|
528 |
+
self.encoder = Kosmos2VisionEncoder(config)
|
529 |
+
self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
|
530 |
+
|
531 |
+
@add_start_docstrings_to_model_forward(KOSMOS2_VISION_INPUTS_DOCSTRING)
|
532 |
+
def forward(
|
533 |
+
self,
|
534 |
+
pixel_values: Optional[torch.FloatTensor] = None,
|
535 |
+
output_attentions: Optional[bool] = None,
|
536 |
+
output_hidden_states: Optional[bool] = None,
|
537 |
+
return_dict: Optional[bool] = None,
|
538 |
+
) -> Union[Tuple, BaseModelOutputWithPooling]:
|
539 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
540 |
+
output_hidden_states = (
|
541 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
542 |
+
)
|
543 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
544 |
+
|
545 |
+
if pixel_values is None:
|
546 |
+
raise ValueError("You have to specify pixel_values")
|
547 |
+
|
548 |
+
hidden_states = self.embeddings(pixel_values)
|
549 |
+
hidden_states = self.pre_layrnorm(hidden_states)
|
550 |
+
|
551 |
+
encoder_outputs = self.encoder(
|
552 |
+
inputs_embeds=hidden_states,
|
553 |
+
output_attentions=output_attentions,
|
554 |
+
output_hidden_states=output_hidden_states,
|
555 |
+
return_dict=return_dict,
|
556 |
+
)
|
557 |
+
|
558 |
+
last_hidden_state = encoder_outputs[0]
|
559 |
+
pooled_output = last_hidden_state[:, 0, :]
|
560 |
+
pooled_output = self.post_layernorm(pooled_output)
|
561 |
+
|
562 |
+
if not return_dict:
|
563 |
+
return (last_hidden_state, pooled_output) + encoder_outputs[1:]
|
564 |
+
|
565 |
+
return BaseModelOutputWithPooling(
|
566 |
+
last_hidden_state=last_hidden_state,
|
567 |
+
pooler_output=pooled_output,
|
568 |
+
hidden_states=encoder_outputs.hidden_states,
|
569 |
+
attentions=encoder_outputs.attentions,
|
570 |
+
)
|
571 |
+
|
572 |
+
|
573 |
+
# Copied from transformers.models.m2m_100.modeling_m2m_100.M2M100SinusoidalPositionalEmbedding with M2M100->Kosmos2
|
574 |
+
class Kosmos2TextSinusoidalPositionalEmbedding(nn.Module):
|
575 |
+
"""This module produces sinusoidal positional embeddings of any length."""
|
576 |
+
|
577 |
+
def __init__(self, num_positions: int, embedding_dim: int, padding_idx: Optional[int] = None):
|
578 |
+
super().__init__()
|
579 |
+
self.offset = 2
|
580 |
+
self.embedding_dim = embedding_dim
|
581 |
+
self.padding_idx = padding_idx
|
582 |
+
self.make_weights(num_positions + self.offset, embedding_dim, padding_idx)
|
583 |
+
|
584 |
+
def make_weights(self, num_embeddings: int, embedding_dim: int, padding_idx: Optional[int] = None):
|
585 |
+
emb_weights = self.get_embedding(num_embeddings, embedding_dim, padding_idx)
|
586 |
+
if hasattr(self, "weights"):
|
587 |
+
# in forward put the weights on the correct dtype and device of the param
|
588 |
+
emb_weights = emb_weights.to(dtype=self.weights.dtype, device=self.weights.device)
|
589 |
+
|
590 |
+
self.register_buffer("weights", emb_weights, persistent=False)
|
591 |
+
|
592 |
+
@staticmethod
|
593 |
+
def get_embedding(num_embeddings: int, embedding_dim: int, padding_idx: Optional[int] = None):
|
594 |
+
"""
|
595 |
+
Build sinusoidal embeddings.
|
596 |
+
|
597 |
+
This matches the implementation in tensor2tensor, but differs slightly from the description in Section 3.5 of
|
598 |
+
"Attention Is All You Need".
|
599 |
+
"""
|
600 |
+
half_dim = embedding_dim // 2
|
601 |
+
emb = math.log(10000) / (half_dim - 1)
|
602 |
+
emb = torch.exp(torch.arange(half_dim, dtype=torch.float) * -emb)
|
603 |
+
emb = torch.arange(num_embeddings, dtype=torch.float).unsqueeze(1) * emb.unsqueeze(0)
|
604 |
+
emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1).view(num_embeddings, -1)
|
605 |
+
if embedding_dim % 2 == 1:
|
606 |
+
# zero pad
|
607 |
+
emb = torch.cat([emb, torch.zeros(num_embeddings, 1)], dim=1)
|
608 |
+
if padding_idx is not None:
|
609 |
+
emb[padding_idx, :] = 0
|
610 |
+
|
611 |
+
return emb.to(torch.get_default_dtype())
|
612 |
+
|
613 |
+
@torch.no_grad()
|
614 |
+
def forward(
|
615 |
+
self, input_ids: torch.Tensor = None, inputs_embeds: torch.Tensor = None, past_key_values_length: int = 0
|
616 |
+
):
|
617 |
+
if input_ids is not None:
|
618 |
+
bsz, seq_len = input_ids.size()
|
619 |
+
# Create the position ids from the input token ids. Any padded tokens remain padded.
|
620 |
+
position_ids = create_position_ids_from_input_ids(input_ids, self.padding_idx, past_key_values_length).to(
|
621 |
+
input_ids.device
|
622 |
+
)
|
623 |
+
else:
|
624 |
+
bsz, seq_len = inputs_embeds.size()[:-1]
|
625 |
+
position_ids = self.create_position_ids_from_inputs_embeds(inputs_embeds, past_key_values_length)
|
626 |
+
|
627 |
+
# expand embeddings if needed
|
628 |
+
max_pos = self.padding_idx + 1 + seq_len + past_key_values_length
|
629 |
+
if max_pos > self.weights.size(0):
|
630 |
+
self.make_weights(max_pos + self.offset, self.embedding_dim, self.padding_idx)
|
631 |
+
|
632 |
+
return self.weights.index_select(0, position_ids.view(-1)).view(bsz, seq_len, self.weights.shape[-1]).detach()
|
633 |
+
|
634 |
+
def create_position_ids_from_inputs_embeds(self, inputs_embeds, past_key_values_length):
|
635 |
+
"""
|
636 |
+
We are provided embeddings directly. We cannot infer which are padded so just generate sequential position ids.
|
637 |
+
|
638 |
+
Args:
|
639 |
+
inputs_embeds: torch.Tensor
|
640 |
+
|
641 |
+
Returns: torch.Tensor
|
642 |
+
"""
|
643 |
+
input_shape = inputs_embeds.size()[:-1]
|
644 |
+
sequence_length = input_shape[1]
|
645 |
+
|
646 |
+
position_ids = torch.arange(
|
647 |
+
self.padding_idx + 1, sequence_length + self.padding_idx + 1, dtype=torch.long, device=inputs_embeds.device
|
648 |
+
)
|
649 |
+
return position_ids.unsqueeze(0).expand(input_shape).contiguous() + past_key_values_length
|
650 |
+
|
651 |
+
|
652 |
+
# Similar to transformers.models.bart.modeling_bart.BartAttention with an additional `inner_attn_ln`.
|
653 |
+
class KosmosTextAttention(nn.Module):
|
654 |
+
"""Multi-headed attention from 'Attention Is All You Need' paper"""
|
655 |
+
|
656 |
+
def __init__(
|
657 |
+
self,
|
658 |
+
config,
|
659 |
+
embed_dim: int,
|
660 |
+
num_heads: int,
|
661 |
+
dropout: float = 0.0,
|
662 |
+
is_decoder: bool = False,
|
663 |
+
add_inner_attn_layernorm: bool = False,
|
664 |
+
bias: bool = True,
|
665 |
+
):
|
666 |
+
super().__init__()
|
667 |
+
self.embed_dim = embed_dim
|
668 |
+
self.num_heads = num_heads
|
669 |
+
self.dropout = dropout
|
670 |
+
self.head_dim = embed_dim // num_heads
|
671 |
+
|
672 |
+
if (self.head_dim * num_heads) != self.embed_dim:
|
673 |
+
raise ValueError(
|
674 |
+
f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
|
675 |
+
f" and `num_heads`: {num_heads})."
|
676 |
+
)
|
677 |
+
self.scaling = self.head_dim**-0.5
|
678 |
+
self.is_decoder = is_decoder
|
679 |
+
|
680 |
+
self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
|
681 |
+
self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
|
682 |
+
self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
|
683 |
+
self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
|
684 |
+
|
685 |
+
self.inner_attn_ln = None
|
686 |
+
if add_inner_attn_layernorm:
|
687 |
+
self.inner_attn_ln = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
|
688 |
+
|
689 |
+
def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
|
690 |
+
return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
|
691 |
+
|
692 |
+
def forward(
|
693 |
+
self,
|
694 |
+
hidden_states: torch.Tensor,
|
695 |
+
key_value_states: Optional[torch.Tensor] = None,
|
696 |
+
past_key_value: Optional[Tuple[torch.Tensor]] = None,
|
697 |
+
attention_mask: Optional[torch.Tensor] = None,
|
698 |
+
layer_head_mask: Optional[torch.Tensor] = None,
|
699 |
+
output_attentions: bool = False,
|
700 |
+
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
701 |
+
"""Input shape: Batch x Time x Channel"""
|
702 |
+
|
703 |
+
# if key_value_states are provided this layer is used as a cross-attention layer
|
704 |
+
# for the decoder
|
705 |
+
is_cross_attention = key_value_states is not None
|
706 |
+
|
707 |
+
bsz, tgt_len, _ = hidden_states.size()
|
708 |
+
|
709 |
+
# get query proj
|
710 |
+
query_states = self.q_proj(hidden_states) * self.scaling
|
711 |
+
# get key, value proj
|
712 |
+
# `past_key_value[0].shape[2] == key_value_states.shape[1]`
|
713 |
+
# is checking that the `sequence_length` of the `past_key_value` is the same as
|
714 |
+
# the provided `key_value_states` to support prefix tuning
|
715 |
+
if (
|
716 |
+
is_cross_attention
|
717 |
+
and past_key_value is not None
|
718 |
+
and past_key_value[0].shape[2] == key_value_states.shape[1]
|
719 |
+
):
|
720 |
+
# reuse k,v, cross_attentions
|
721 |
+
key_states = past_key_value[0]
|
722 |
+
value_states = past_key_value[1]
|
723 |
+
elif is_cross_attention:
|
724 |
+
# cross_attentions
|
725 |
+
key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
|
726 |
+
value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
|
727 |
+
elif past_key_value is not None:
|
728 |
+
# reuse k, v, self_attention
|
729 |
+
key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
|
730 |
+
value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
|
731 |
+
key_states = torch.cat([past_key_value[0], key_states], dim=2)
|
732 |
+
value_states = torch.cat([past_key_value[1], value_states], dim=2)
|
733 |
+
else:
|
734 |
+
# self_attention
|
735 |
+
key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
|
736 |
+
value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
|
737 |
+
|
738 |
+
if self.is_decoder:
|
739 |
+
# if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
|
740 |
+
# Further calls to cross_attention layer can then reuse all cross-attention
|
741 |
+
# key/value_states (first "if" case)
|
742 |
+
# if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
|
743 |
+
# all previous decoder key/value_states. Further calls to uni-directional self-attention
|
744 |
+
# can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
|
745 |
+
# if encoder bi-directional self-attention `past_key_value` is always `None`
|
746 |
+
past_key_value = (key_states, value_states)
|
747 |
+
|
748 |
+
proj_shape = (bsz * self.num_heads, -1, self.head_dim)
|
749 |
+
query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
|
750 |
+
key_states = key_states.reshape(*proj_shape)
|
751 |
+
value_states = value_states.reshape(*proj_shape)
|
752 |
+
|
753 |
+
src_len = key_states.size(1)
|
754 |
+
attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
|
755 |
+
|
756 |
+
if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
|
757 |
+
raise ValueError(
|
758 |
+
f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
|
759 |
+
f" {attn_weights.size()}"
|
760 |
+
)
|
761 |
+
|
762 |
+
if attention_mask is not None:
|
763 |
+
if attention_mask.size() != (bsz, 1, tgt_len, src_len):
|
764 |
+
raise ValueError(
|
765 |
+
f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
|
766 |
+
)
|
767 |
+
attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
|
768 |
+
attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
|
769 |
+
|
770 |
+
attn_weights = nn.functional.softmax(attn_weights, dim=-1)
|
771 |
+
|
772 |
+
if layer_head_mask is not None:
|
773 |
+
if layer_head_mask.size() != (self.num_heads,):
|
774 |
+
raise ValueError(
|
775 |
+
f"Head mask for a single layer should be of size {(self.num_heads,)}, but is"
|
776 |
+
f" {layer_head_mask.size()}"
|
777 |
+
)
|
778 |
+
attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
|
779 |
+
attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
|
780 |
+
|
781 |
+
if output_attentions:
|
782 |
+
# this operation is a bit awkward, but it's required to
|
783 |
+
# make sure that attn_weights keeps its gradient.
|
784 |
+
# In order to do so, attn_weights have to be reshaped
|
785 |
+
# twice and have to be reused in the following
|
786 |
+
attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
|
787 |
+
attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
|
788 |
+
else:
|
789 |
+
attn_weights_reshaped = None
|
790 |
+
|
791 |
+
attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
|
792 |
+
|
793 |
+
attn_output = torch.bmm(attn_probs, value_states)
|
794 |
+
|
795 |
+
if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
|
796 |
+
raise ValueError(
|
797 |
+
f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
|
798 |
+
f" {attn_output.size()}"
|
799 |
+
)
|
800 |
+
|
801 |
+
attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
|
802 |
+
attn_output = attn_output.transpose(1, 2)
|
803 |
+
|
804 |
+
# Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
|
805 |
+
# partitioned across GPUs when using tensor-parallelism.
|
806 |
+
attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
|
807 |
+
|
808 |
+
if self.inner_attn_ln is not None:
|
809 |
+
attn_output = self.inner_attn_ln(attn_output)
|
810 |
+
|
811 |
+
attn_output = self.out_proj(attn_output)
|
812 |
+
|
813 |
+
return attn_output, attn_weights_reshaped, past_key_value
|
814 |
+
|
815 |
+
|
816 |
+
class Kosmos2TextFFN(nn.Module):
|
817 |
+
def __init__(self, config: Kosmos2TextConfig):
|
818 |
+
super().__init__()
|
819 |
+
|
820 |
+
self.dropout = config.dropout
|
821 |
+
self.activation_fn = ACT2FN[config.activation_function]
|
822 |
+
self.activation_dropout = config.activation_dropout
|
823 |
+
|
824 |
+
self.fc1 = nn.Linear(config.embed_dim, config.ffn_dim)
|
825 |
+
self.fc2 = nn.Linear(config.ffn_dim, config.embed_dim)
|
826 |
+
|
827 |
+
self.ffn_layernorm = nn.LayerNorm(config.ffn_dim, eps=config.layer_norm_eps)
|
828 |
+
|
829 |
+
def forward(self, hidden_states):
|
830 |
+
hidden_states = self.activation_fn(self.fc1(hidden_states))
|
831 |
+
hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
|
832 |
+
hidden_states = self.ffn_layernorm(hidden_states)
|
833 |
+
hidden_states = self.fc2(hidden_states)
|
834 |
+
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
|
835 |
+
|
836 |
+
return hidden_states
|
837 |
+
|
838 |
+
|
839 |
+
class Kosmos2TextBlock(nn.Module):
|
840 |
+
def __init__(self, config: Kosmos2TextConfig):
|
841 |
+
super().__init__()
|
842 |
+
self.embed_dim = config.embed_dim
|
843 |
+
|
844 |
+
self.self_attn = KosmosTextAttention(
|
845 |
+
config,
|
846 |
+
embed_dim=self.embed_dim,
|
847 |
+
num_heads=config.attention_heads,
|
848 |
+
dropout=config.attention_dropout,
|
849 |
+
is_decoder=True,
|
850 |
+
add_inner_attn_layernorm=True,
|
851 |
+
)
|
852 |
+
self.dropout = config.dropout
|
853 |
+
self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
|
854 |
+
|
855 |
+
if config.add_cross_attention:
|
856 |
+
self.encoder_attn = KosmosTextAttention(
|
857 |
+
config,
|
858 |
+
embed_dim=self.embed_dim,
|
859 |
+
num_heads=config.attention_heads,
|
860 |
+
dropout=config.attention_dropout,
|
861 |
+
is_decoder=True,
|
862 |
+
add_inner_attn_layernorm=False,
|
863 |
+
)
|
864 |
+
self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
|
865 |
+
|
866 |
+
self.ffn = Kosmos2TextFFN(config)
|
867 |
+
self.final_layer_norm = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
|
868 |
+
|
869 |
+
def forward(
|
870 |
+
self,
|
871 |
+
hidden_states: torch.Tensor,
|
872 |
+
attention_mask: Optional[torch.Tensor] = None,
|
873 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
874 |
+
encoder_attention_mask: Optional[torch.Tensor] = None,
|
875 |
+
layer_head_mask: Optional[torch.Tensor] = None,
|
876 |
+
cross_attn_layer_head_mask: Optional[torch.Tensor] = None,
|
877 |
+
past_key_value: Optional[Tuple[torch.Tensor]] = None,
|
878 |
+
output_attentions: Optional[bool] = False,
|
879 |
+
use_cache: Optional[bool] = True,
|
880 |
+
) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
|
881 |
+
residual = hidden_states
|
882 |
+
|
883 |
+
# Self Attention
|
884 |
+
# decoder uni-directional self-attention cached key/values tuple is at positions 1,2
|
885 |
+
self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
|
886 |
+
|
887 |
+
hidden_states = self.self_attn_layer_norm(hidden_states)
|
888 |
+
|
889 |
+
# add present self-attn cache to positions 1,2 of present_key_value tuple
|
890 |
+
hidden_states, self_attn_weights, present_key_value = self.self_attn(
|
891 |
+
hidden_states=hidden_states,
|
892 |
+
past_key_value=self_attn_past_key_value,
|
893 |
+
attention_mask=attention_mask,
|
894 |
+
layer_head_mask=layer_head_mask,
|
895 |
+
output_attentions=output_attentions,
|
896 |
+
)
|
897 |
+
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
|
898 |
+
hidden_states = residual + hidden_states
|
899 |
+
|
900 |
+
# Cross-Attention Block
|
901 |
+
cross_attn_present_key_value = None
|
902 |
+
cross_attn_weights = None
|
903 |
+
if encoder_hidden_states is not None:
|
904 |
+
if not hasattr(self, "encoder_attn"):
|
905 |
+
raise ValueError(
|
906 |
+
f"If `encoder_hidden_states` are passed, {self} has to be instantiated with cross-attention layers"
|
907 |
+
" by setting `config.add_cross_attention=True`"
|
908 |
+
)
|
909 |
+
|
910 |
+
residual = hidden_states
|
911 |
+
|
912 |
+
hidden_states = self.encoder_attn_layer_norm(hidden_states)
|
913 |
+
|
914 |
+
# cross_attn cached key/values tuple is at positions 3,4 of present_key_value tuple
|
915 |
+
cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
|
916 |
+
hidden_states, cross_attn_weights, cross_attn_present_key_value = self.encoder_attn(
|
917 |
+
hidden_states=hidden_states,
|
918 |
+
key_value_states=encoder_hidden_states,
|
919 |
+
attention_mask=encoder_attention_mask,
|
920 |
+
layer_head_mask=cross_attn_layer_head_mask,
|
921 |
+
past_key_value=cross_attn_past_key_value,
|
922 |
+
output_attentions=output_attentions,
|
923 |
+
)
|
924 |
+
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
|
925 |
+
hidden_states = residual + hidden_states
|
926 |
+
|
927 |
+
# add cross-attn to positions 3,4 of present_key_value tuple
|
928 |
+
present_key_value = present_key_value + cross_attn_present_key_value
|
929 |
+
|
930 |
+
# Fully Connected
|
931 |
+
residual = hidden_states
|
932 |
+
|
933 |
+
hidden_states = self.final_layer_norm(hidden_states)
|
934 |
+
|
935 |
+
# FFN
|
936 |
+
hidden_states = self.ffn(hidden_states)
|
937 |
+
hidden_states = residual + hidden_states
|
938 |
+
|
939 |
+
outputs = (hidden_states,)
|
940 |
+
|
941 |
+
if output_attentions:
|
942 |
+
outputs += (self_attn_weights, cross_attn_weights)
|
943 |
+
|
944 |
+
if use_cache:
|
945 |
+
outputs += (present_key_value,)
|
946 |
+
|
947 |
+
return outputs
|
948 |
+
|
949 |
+
|
950 |
+
class Kosmos2TextTransformer(nn.Module):
|
951 |
+
"""
|
952 |
+
Transformer decoder consisting of `config.layers` layers. Each layer is a [`Kosmos2TextBlock`].
|
953 |
+
|
954 |
+
Args:
|
955 |
+
config: Kosmos2TextConfig
|
956 |
+
"""
|
957 |
+
|
958 |
+
def __init__(self, config: Kosmos2TextConfig):
|
959 |
+
super().__init__()
|
960 |
+
self.config = config
|
961 |
+
self.dropout = config.dropout
|
962 |
+
self.layerdrop = config.layerdrop
|
963 |
+
|
964 |
+
self.embed_scale = math.sqrt(config.embed_dim) if config.scale_embedding else 1.0
|
965 |
+
self.embed_tokens = nn.Embedding(config.vocab_size, config.embed_dim, padding_idx=config.pad_token_id)
|
966 |
+
|
967 |
+
self.embed_positions = Kosmos2TextSinusoidalPositionalEmbedding(
|
968 |
+
num_positions=config.max_position_embeddings,
|
969 |
+
embedding_dim=config.embed_dim,
|
970 |
+
padding_idx=config.pad_token_id,
|
971 |
+
)
|
972 |
+
|
973 |
+
self.layers = nn.ModuleList([Kosmos2TextBlock(config) for _ in range(config.layers)])
|
974 |
+
self.layer_norm = nn.LayerNorm(config.embed_dim, config.layer_norm_eps)
|
975 |
+
|
976 |
+
self.gradient_checkpointing = False
|
977 |
+
|
978 |
+
# Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
|
979 |
+
def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
|
980 |
+
# create causal mask
|
981 |
+
# [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
|
982 |
+
combined_attention_mask = None
|
983 |
+
if input_shape[-1] > 1:
|
984 |
+
combined_attention_mask = _make_causal_mask(
|
985 |
+
input_shape,
|
986 |
+
inputs_embeds.dtype,
|
987 |
+
device=inputs_embeds.device,
|
988 |
+
past_key_values_length=past_key_values_length,
|
989 |
+
)
|
990 |
+
|
991 |
+
if attention_mask is not None:
|
992 |
+
# [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
|
993 |
+
expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
|
994 |
+
inputs_embeds.device
|
995 |
+
)
|
996 |
+
combined_attention_mask = (
|
997 |
+
expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
|
998 |
+
)
|
999 |
+
|
1000 |
+
return combined_attention_mask
|
1001 |
+
|
1002 |
+
def forward_embedding(
|
1003 |
+
self, input_ids, inputs_embeds=None, img_features=None, img_input_mask=None, past_key_values_length: int = 0
|
1004 |
+
):
|
1005 |
+
# The argument `inputs_embeds` should be the one without being multiplied by `self.embed_scale`.
|
1006 |
+
if inputs_embeds is None:
|
1007 |
+
inputs_embeds = self.embed_tokens(input_ids)
|
1008 |
+
|
1009 |
+
if img_features is not None:
|
1010 |
+
inputs_embeds[img_input_mask.to(dtype=torch.bool)] = img_features
|
1011 |
+
|
1012 |
+
inputs_embeds = inputs_embeds * self.embed_scale
|
1013 |
+
|
1014 |
+
# embed positions
|
1015 |
+
positions = self.embed_positions(
|
1016 |
+
input_ids=input_ids, inputs_embeds=inputs_embeds, past_key_values_length=past_key_values_length
|
1017 |
+
)
|
1018 |
+
positions = positions.to(inputs_embeds.device)
|
1019 |
+
|
1020 |
+
hidden_states = inputs_embeds + positions
|
1021 |
+
|
1022 |
+
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
|
1023 |
+
|
1024 |
+
return hidden_states
|
1025 |
+
|
1026 |
+
def forward(
|
1027 |
+
self,
|
1028 |
+
input_ids: Optional[torch.Tensor] = None,
|
1029 |
+
attention_mask: Optional[torch.Tensor] = None,
|
1030 |
+
img_features: Optional[torch.Tensor] = None,
|
1031 |
+
img_attn_mask: Optional[torch.Tensor] = None,
|
1032 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
1033 |
+
encoder_attention_mask: Optional[torch.Tensor] = None,
|
1034 |
+
head_mask: Optional[torch.Tensor] = None,
|
1035 |
+
cross_attn_head_mask: Optional[torch.Tensor] = None,
|
1036 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
1037 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
1038 |
+
use_cache: Optional[bool] = None,
|
1039 |
+
output_attentions: Optional[bool] = None,
|
1040 |
+
output_hidden_states: Optional[bool] = None,
|
1041 |
+
return_dict: Optional[bool] = None,
|
1042 |
+
) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]:
|
1043 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
1044 |
+
output_hidden_states = (
|
1045 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
1046 |
+
)
|
1047 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
1048 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
1049 |
+
|
1050 |
+
if input_ids is not None and inputs_embeds is not None:
|
1051 |
+
raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
|
1052 |
+
elif input_ids is not None:
|
1053 |
+
input_shape = input_ids.shape
|
1054 |
+
input_ids = input_ids.view(-1, input_shape[-1])
|
1055 |
+
elif inputs_embeds is not None:
|
1056 |
+
input_shape = inputs_embeds.size()[:-1]
|
1057 |
+
else:
|
1058 |
+
raise ValueError("You have to specify either input_ids or inputs_embeds")
|
1059 |
+
|
1060 |
+
# past_key_values_length
|
1061 |
+
past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
|
1062 |
+
|
1063 |
+
# We don't need img info. when `past_key_values_length` > 0
|
1064 |
+
if past_key_values_length > 0:
|
1065 |
+
img_features = None
|
1066 |
+
img_attn_mask = None
|
1067 |
+
|
1068 |
+
hidden_states = self.forward_embedding(
|
1069 |
+
input_ids=input_ids,
|
1070 |
+
inputs_embeds=inputs_embeds,
|
1071 |
+
img_features=img_features,
|
1072 |
+
img_input_mask=img_attn_mask,
|
1073 |
+
past_key_values_length=past_key_values_length,
|
1074 |
+
)
|
1075 |
+
|
1076 |
+
attention_mask = self._prepare_decoder_attention_mask(
|
1077 |
+
attention_mask, input_shape, hidden_states, past_key_values_length
|
1078 |
+
)
|
1079 |
+
|
1080 |
+
# expand encoder attention mask
|
1081 |
+
if encoder_hidden_states is not None and encoder_attention_mask is not None:
|
1082 |
+
# [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
|
1083 |
+
encoder_attention_mask = _expand_mask(encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1])
|
1084 |
+
|
1085 |
+
hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
|
1086 |
+
|
1087 |
+
if self.gradient_checkpointing and self.training:
|
1088 |
+
if use_cache:
|
1089 |
+
logger.warning_once(
|
1090 |
+
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
|
1091 |
+
)
|
1092 |
+
use_cache = False
|
1093 |
+
|
1094 |
+
# decoder layers
|
1095 |
+
all_hidden_states = () if output_hidden_states else None
|
1096 |
+
all_self_attns = () if output_attentions else None
|
1097 |
+
all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
|
1098 |
+
next_decoder_cache = () if use_cache else None
|
1099 |
+
|
1100 |
+
# check if head_mask/cross_attn_head_mask has a correct number of layers specified if desired
|
1101 |
+
for attn_mask, mask_name in zip([head_mask, cross_attn_head_mask], ["head_mask", "cross_attn_head_mask"]):
|
1102 |
+
if attn_mask is not None:
|
1103 |
+
if attn_mask.size()[0] != (len(self.layers)):
|
1104 |
+
raise ValueError(
|
1105 |
+
f"The `{mask_name}` should be specified for {len(self.layers)} layers, but it is for"
|
1106 |
+
f" {head_mask.size()[0]}."
|
1107 |
+
)
|
1108 |
+
|
1109 |
+
for idx, decoder_layer in enumerate(self.layers):
|
1110 |
+
# add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
|
1111 |
+
if output_hidden_states:
|
1112 |
+
all_hidden_states += (hidden_states,)
|
1113 |
+
if self.training:
|
1114 |
+
dropout_probability = torch.rand([])
|
1115 |
+
if dropout_probability < self.layerdrop:
|
1116 |
+
continue
|
1117 |
+
|
1118 |
+
past_key_value = past_key_values[idx] if past_key_values is not None else None
|
1119 |
+
|
1120 |
+
if self.gradient_checkpointing and self.training:
|
1121 |
+
|
1122 |
+
def create_custom_forward(module):
|
1123 |
+
def custom_forward(*inputs):
|
1124 |
+
# None for past_key_value
|
1125 |
+
return module(*inputs, output_attentions, use_cache)
|
1126 |
+
|
1127 |
+
return custom_forward
|
1128 |
+
|
1129 |
+
layer_outputs = torch.utils.checkpoint.checkpoint(
|
1130 |
+
create_custom_forward(decoder_layer),
|
1131 |
+
hidden_states,
|
1132 |
+
attention_mask,
|
1133 |
+
encoder_hidden_states,
|
1134 |
+
encoder_attention_mask,
|
1135 |
+
head_mask[idx] if head_mask is not None else None,
|
1136 |
+
cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None,
|
1137 |
+
None,
|
1138 |
+
)
|
1139 |
+
else:
|
1140 |
+
layer_outputs = decoder_layer(
|
1141 |
+
hidden_states,
|
1142 |
+
attention_mask=attention_mask,
|
1143 |
+
encoder_hidden_states=encoder_hidden_states,
|
1144 |
+
encoder_attention_mask=encoder_attention_mask,
|
1145 |
+
layer_head_mask=(head_mask[idx] if head_mask is not None else None),
|
1146 |
+
cross_attn_layer_head_mask=(
|
1147 |
+
cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None
|
1148 |
+
),
|
1149 |
+
past_key_value=past_key_value,
|
1150 |
+
output_attentions=output_attentions,
|
1151 |
+
use_cache=use_cache,
|
1152 |
+
)
|
1153 |
+
hidden_states = layer_outputs[0]
|
1154 |
+
|
1155 |
+
if use_cache:
|
1156 |
+
next_decoder_cache += (layer_outputs[3 if output_attentions else 1],)
|
1157 |
+
|
1158 |
+
if output_attentions:
|
1159 |
+
all_self_attns += (layer_outputs[1],)
|
1160 |
+
|
1161 |
+
if encoder_hidden_states is not None:
|
1162 |
+
all_cross_attentions += (layer_outputs[2],)
|
1163 |
+
|
1164 |
+
# add final layer norm
|
1165 |
+
hidden_states = self.layer_norm(hidden_states)
|
1166 |
+
|
1167 |
+
# add hidden states from the last decoder layer
|
1168 |
+
if output_hidden_states:
|
1169 |
+
all_hidden_states += (hidden_states,)
|
1170 |
+
|
1171 |
+
next_cache = next_decoder_cache if use_cache else None
|
1172 |
+
if not return_dict:
|
1173 |
+
return tuple(
|
1174 |
+
v
|
1175 |
+
for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_cross_attentions]
|
1176 |
+
if v is not None
|
1177 |
+
)
|
1178 |
+
return BaseModelOutputWithPastAndCrossAttentions(
|
1179 |
+
last_hidden_state=hidden_states,
|
1180 |
+
past_key_values=next_cache,
|
1181 |
+
hidden_states=all_hidden_states,
|
1182 |
+
attentions=all_self_attns,
|
1183 |
+
cross_attentions=all_cross_attentions,
|
1184 |
+
)
|
1185 |
+
|
1186 |
+
|
1187 |
+
class Kosmos2PreTrainedModel(PreTrainedModel):
|
1188 |
+
"""
|
1189 |
+
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
|
1190 |
+
models.
|
1191 |
+
"""
|
1192 |
+
|
1193 |
+
config_class = Kosmos2Config
|
1194 |
+
supports_gradient_checkpointing = True
|
1195 |
+
|
1196 |
+
|
1197 |
+
@add_start_docstrings(
|
1198 |
+
"""The vision model from KOSMOS-2 without any head or projection on top.""",
|
1199 |
+
KOSMOS2_START_DOCSTRING,
|
1200 |
+
)
|
1201 |
+
class Kosmos2VisionModel(Kosmos2PreTrainedModel):
|
1202 |
+
config_class = Kosmos2VisionConfig
|
1203 |
+
main_input_name = "pixel_values"
|
1204 |
+
|
1205 |
+
# Copied from transformers.models.clip.modeling_clip.CLIPVisionModel.__init__ with CLIP_VISION->KOSMOS2_VISION,CLIP->Kosmos2
|
1206 |
+
def __init__(self, config: Kosmos2VisionConfig):
|
1207 |
+
super().__init__(config)
|
1208 |
+
self.model = Kosmos2VisionTransformer(config)
|
1209 |
+
# Initialize weights and apply final processing
|
1210 |
+
self.post_init()
|
1211 |
+
|
1212 |
+
# Copied from transformers.models.clip.modeling_clip.CLIPVisionModel.get_input_embeddings with CLIP_VISION->KOSMOS2_VISION,CLIP->Kosmos2
|
1213 |
+
def get_input_embeddings(self) -> nn.Module:
|
1214 |
+
return self.model.embeddings.patch_embedding
|
1215 |
+
|
1216 |
+
@add_start_docstrings_to_model_forward(KOSMOS2_VISION_INPUTS_DOCSTRING)
|
1217 |
+
@replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=Kosmos2VisionConfig)
|
1218 |
+
def forward(
|
1219 |
+
self,
|
1220 |
+
pixel_values: Optional[torch.FloatTensor] = None,
|
1221 |
+
output_attentions: Optional[bool] = None,
|
1222 |
+
output_hidden_states: Optional[bool] = None,
|
1223 |
+
return_dict: Optional[bool] = None,
|
1224 |
+
) -> Union[Tuple, BaseModelOutputWithPooling]:
|
1225 |
+
r"""
|
1226 |
+
Returns:
|
1227 |
+
|
1228 |
+
"""
|
1229 |
+
return self.model(
|
1230 |
+
pixel_values=pixel_values,
|
1231 |
+
output_attentions=output_attentions,
|
1232 |
+
output_hidden_states=output_hidden_states,
|
1233 |
+
return_dict=return_dict,
|
1234 |
+
)
|
1235 |
+
|
1236 |
+
|
1237 |
+
@add_start_docstrings(
|
1238 |
+
"""The text model from KOSMOS-2 without any head or projection on top.""",
|
1239 |
+
KOSMOS2_START_DOCSTRING,
|
1240 |
+
)
|
1241 |
+
class Kosmos2TextModel(Kosmos2PreTrainedModel):
|
1242 |
+
config_class = Kosmos2TextConfig
|
1243 |
+
|
1244 |
+
_no_split_modules = ["Kosmos2TextBlock"]
|
1245 |
+
|
1246 |
+
def __init__(self, config: Kosmos2TextConfig):
|
1247 |
+
super().__init__(config)
|
1248 |
+
self.model = Kosmos2TextTransformer(config)
|
1249 |
+
# Initialize weights and apply final processing
|
1250 |
+
self.post_init()
|
1251 |
+
|
1252 |
+
def get_input_embeddings(self) -> nn.Module:
|
1253 |
+
return self.model.embed_tokens
|
1254 |
+
|
1255 |
+
def set_input_embeddings(self, value):
|
1256 |
+
self.model.embed_tokens = value
|
1257 |
+
|
1258 |
+
@add_start_docstrings_to_model_forward(KOSMOS2_TEXT_INPUTS_DOCSTRING)
|
1259 |
+
@replace_return_docstrings(output_type=BaseModelOutputWithPastAndCrossAttentions, config_class=Kosmos2TextConfig)
|
1260 |
+
def forward(
|
1261 |
+
self,
|
1262 |
+
input_ids: Optional[torch.Tensor] = None,
|
1263 |
+
attention_mask: Optional[torch.Tensor] = None,
|
1264 |
+
img_features: Optional[torch.Tensor] = None,
|
1265 |
+
img_attn_mask: Optional[torch.Tensor] = None,
|
1266 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
1267 |
+
encoder_attention_mask: Optional[torch.Tensor] = None,
|
1268 |
+
head_mask: Optional[torch.Tensor] = None,
|
1269 |
+
cross_attn_head_mask: Optional[torch.Tensor] = None,
|
1270 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
1271 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
1272 |
+
use_cache: Optional[bool] = None,
|
1273 |
+
output_attentions: Optional[bool] = None,
|
1274 |
+
output_hidden_states: Optional[bool] = None,
|
1275 |
+
return_dict: Optional[bool] = None,
|
1276 |
+
) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]:
|
1277 |
+
r"""
|
1278 |
+
Returns:
|
1279 |
+
|
1280 |
+
"""
|
1281 |
+
return self.model(
|
1282 |
+
input_ids=input_ids,
|
1283 |
+
attention_mask=attention_mask,
|
1284 |
+
img_features=img_features,
|
1285 |
+
img_attn_mask=img_attn_mask,
|
1286 |
+
encoder_hidden_states=encoder_hidden_states,
|
1287 |
+
encoder_attention_mask=encoder_attention_mask,
|
1288 |
+
head_mask=head_mask,
|
1289 |
+
cross_attn_head_mask=cross_attn_head_mask,
|
1290 |
+
past_key_values=past_key_values,
|
1291 |
+
inputs_embeds=inputs_embeds,
|
1292 |
+
use_cache=use_cache,
|
1293 |
+
output_attentions=output_attentions,
|
1294 |
+
output_hidden_states=output_hidden_states,
|
1295 |
+
return_dict=return_dict,
|
1296 |
+
)
|
1297 |
+
|
1298 |
+
|
1299 |
+
@add_start_docstrings(
|
1300 |
+
"""
|
1301 |
+
The text model from KOSMOS-2 with a language modeling head on top (linear layer with weights tied to the input
|
1302 |
+
embeddings).
|
1303 |
+
""",
|
1304 |
+
KOSMOS2_START_DOCSTRING,
|
1305 |
+
)
|
1306 |
+
class Kosmos2TextForCausalLM(Kosmos2PreTrainedModel):
|
1307 |
+
config_class = Kosmos2TextConfig
|
1308 |
+
_tied_weights_keys = ["lm_head.weight"]
|
1309 |
+
|
1310 |
+
def __init__(self, config: Kosmos2TextConfig):
|
1311 |
+
super().__init__(config)
|
1312 |
+
|
1313 |
+
self.model = Kosmos2TextTransformer(config)
|
1314 |
+
self.lm_head = nn.Linear(in_features=config.embed_dim, out_features=config.vocab_size, bias=False)
|
1315 |
+
|
1316 |
+
# Initialize weights and apply final processing
|
1317 |
+
self.post_init()
|
1318 |
+
|
1319 |
+
def get_input_embeddings(self) -> nn.Module:
|
1320 |
+
return self.model.embed_tokens
|
1321 |
+
|
1322 |
+
def set_input_embeddings(self, value):
|
1323 |
+
self.model.embed_tokens = value
|
1324 |
+
|
1325 |
+
def get_output_embeddings(self) -> nn.Module:
|
1326 |
+
return self.lm_head
|
1327 |
+
|
1328 |
+
def set_output_embeddings(self, new_embeddings):
|
1329 |
+
self.lm_head = new_embeddings
|
1330 |
+
|
1331 |
+
@add_start_docstrings_to_model_forward(KOSMOS2_TEXT_INPUTS_DOCSTRING)
|
1332 |
+
@replace_return_docstrings(output_type=CausalLMOutputWithCrossAttentions, config_class=Kosmos2TextConfig)
|
1333 |
+
def forward(
|
1334 |
+
self,
|
1335 |
+
input_ids: Optional[torch.Tensor] = None,
|
1336 |
+
attention_mask: Optional[torch.Tensor] = None,
|
1337 |
+
img_features: Optional[torch.Tensor] = None,
|
1338 |
+
img_attn_mask: Optional[torch.Tensor] = None,
|
1339 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
1340 |
+
encoder_attention_mask: Optional[torch.Tensor] = None,
|
1341 |
+
head_mask: Optional[torch.Tensor] = None,
|
1342 |
+
cross_attn_head_mask: Optional[torch.Tensor] = None,
|
1343 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
1344 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
1345 |
+
labels: Optional[torch.LongTensor] = None,
|
1346 |
+
use_cache: Optional[bool] = None,
|
1347 |
+
output_attentions: Optional[bool] = None,
|
1348 |
+
output_hidden_states: Optional[bool] = None,
|
1349 |
+
return_dict: Optional[bool] = None,
|
1350 |
+
) -> Union[Tuple, CausalLMOutputWithCrossAttentions]:
|
1351 |
+
r"""
|
1352 |
+
Returns:
|
1353 |
+
|
1354 |
+
"""
|
1355 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
1356 |
+
|
1357 |
+
if labels is not None:
|
1358 |
+
if use_cache:
|
1359 |
+
logger.warning("The `use_cache` argument is changed to `False` since `labels` is provided.")
|
1360 |
+
use_cache = False
|
1361 |
+
|
1362 |
+
outputs = self.model(
|
1363 |
+
input_ids=input_ids,
|
1364 |
+
attention_mask=attention_mask,
|
1365 |
+
img_features=img_features,
|
1366 |
+
img_attn_mask=img_attn_mask,
|
1367 |
+
encoder_hidden_states=encoder_hidden_states,
|
1368 |
+
encoder_attention_mask=encoder_attention_mask,
|
1369 |
+
head_mask=head_mask,
|
1370 |
+
cross_attn_head_mask=cross_attn_head_mask,
|
1371 |
+
past_key_values=past_key_values,
|
1372 |
+
inputs_embeds=inputs_embeds,
|
1373 |
+
use_cache=use_cache,
|
1374 |
+
output_attentions=output_attentions,
|
1375 |
+
output_hidden_states=output_hidden_states,
|
1376 |
+
return_dict=return_dict,
|
1377 |
+
)
|
1378 |
+
logits = self.lm_head(outputs[0])
|
1379 |
+
|
1380 |
+
loss = None
|
1381 |
+
if labels is not None:
|
1382 |
+
# Shift so that tokens < n predict n
|
1383 |
+
shift_logits = logits[..., :-1, :].contiguous()
|
1384 |
+
shift_labels = labels[..., 1:].contiguous()
|
1385 |
+
# Flatten the tokens
|
1386 |
+
loss_fct = CrossEntropyLoss()
|
1387 |
+
shift_logits = shift_logits.view(-1, self.config.vocab_size)
|
1388 |
+
shift_labels = shift_labels.view(-1)
|
1389 |
+
# Enable model parallelism
|
1390 |
+
shift_labels = shift_labels.to(shift_logits.device)
|
1391 |
+
loss = loss_fct(shift_logits, shift_labels)
|
1392 |
+
|
1393 |
+
if not return_dict:
|
1394 |
+
output = (logits,) + outputs[1:]
|
1395 |
+
return (loss,) + output if loss is not None else output
|
1396 |
+
|
1397 |
+
return CausalLMOutputWithCrossAttentions(
|
1398 |
+
loss=loss,
|
1399 |
+
logits=logits,
|
1400 |
+
past_key_values=outputs.past_key_values,
|
1401 |
+
hidden_states=outputs.hidden_states,
|
1402 |
+
attentions=outputs.attentions,
|
1403 |
+
cross_attentions=outputs.cross_attentions,
|
1404 |
+
)
|
1405 |
+
|
1406 |
+
def prepare_inputs_for_generation(
|
1407 |
+
self,
|
1408 |
+
input_ids,
|
1409 |
+
img_features,
|
1410 |
+
img_attn_mask,
|
1411 |
+
past_key_values=None,
|
1412 |
+
attention_mask=None,
|
1413 |
+
use_cache=None,
|
1414 |
+
**model_kwargs,
|
1415 |
+
):
|
1416 |
+
input_shape = input_ids.shape
|
1417 |
+
# if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly
|
1418 |
+
if attention_mask is None:
|
1419 |
+
attention_mask = input_ids.new_ones(input_shape)
|
1420 |
+
|
1421 |
+
# cut input_ids if past_key_values is used
|
1422 |
+
if past_key_values is not None:
|
1423 |
+
input_ids = input_ids[:, -1:]
|
1424 |
+
# the image info. is already encoded into the past keys/values
|
1425 |
+
img_features = None
|
1426 |
+
img_attn_mask = None
|
1427 |
+
elif img_attn_mask is not None:
|
1428 |
+
# appending `False` to `img_attn_mask` (because `input_ids` grows during generation)
|
1429 |
+
batch_size, seq_len = input_ids.size()
|
1430 |
+
mask_len = img_attn_mask.size()[-1]
|
1431 |
+
img_attn_mask = torch.cat(
|
1432 |
+
(img_attn_mask, torch.zeros(size=(batch_size, seq_len - mask_len), dtype=torch.bool)), dim=1
|
1433 |
+
)
|
1434 |
+
|
1435 |
+
return {
|
1436 |
+
"input_ids": input_ids,
|
1437 |
+
"img_features": img_features,
|
1438 |
+
"img_attn_mask": img_attn_mask,
|
1439 |
+
"past_key_values": past_key_values,
|
1440 |
+
"attention_mask": attention_mask,
|
1441 |
+
"use_cache": use_cache,
|
1442 |
+
}
|
1443 |
+
|
1444 |
+
|
1445 |
+
class Kosmos2ImageToTextConnector(nn.Module):
|
1446 |
+
"""The layer that transforms the image model's output to part of the text model's input (namely, image features)"""
|
1447 |
+
|
1448 |
+
def __init__(self, config: Kosmos2Config):
|
1449 |
+
super().__init__()
|
1450 |
+
self.dense = nn.Linear(config.vision_config.hidden_size, config.text_config.embed_dim)
|
1451 |
+
self.latent_query = nn.Parameter(torch.randn(config.latent_query_num, config.text_config.embed_dim))
|
1452 |
+
|
1453 |
+
self.x_attn = KosmosTextAttention(
|
1454 |
+
config.text_config,
|
1455 |
+
config.text_config.embed_dim,
|
1456 |
+
config.text_config.attention_heads,
|
1457 |
+
dropout=config.text_config.attention_dropout,
|
1458 |
+
is_decoder=False,
|
1459 |
+
add_inner_attn_layernorm=False,
|
1460 |
+
)
|
1461 |
+
|
1462 |
+
def forward(self, features):
|
1463 |
+
hidden_states = self.dense(features)
|
1464 |
+
|
1465 |
+
# shape = [batch, latent_query_num, h_dim]
|
1466 |
+
latent_query = self.latent_query.unsqueeze(0).expand(hidden_states.size(0), -1, -1)
|
1467 |
+
key_value_states = torch.cat([hidden_states, latent_query], dim=1)
|
1468 |
+
|
1469 |
+
hidden_states, attn_weights, _ = self.x_attn(
|
1470 |
+
hidden_states=latent_query,
|
1471 |
+
key_value_states=key_value_states,
|
1472 |
+
past_key_value=None,
|
1473 |
+
attention_mask=None,
|
1474 |
+
output_attentions=None,
|
1475 |
+
)
|
1476 |
+
|
1477 |
+
return hidden_states, attn_weights
|
1478 |
+
|
1479 |
+
|
1480 |
+
@add_start_docstrings(
|
1481 |
+
"""
|
1482 |
+
KOSMOS-2 Model for generating text and image features. The model consists of a vision encoder (CLIP) and a language
|
1483 |
+
model.
|
1484 |
+
""",
|
1485 |
+
KOSMOS2_START_DOCSTRING,
|
1486 |
+
)
|
1487 |
+
class Kosmos2Model(Kosmos2PreTrainedModel):
|
1488 |
+
config_class = Kosmos2Config
|
1489 |
+
|
1490 |
+
def __init__(self, config: Kosmos2Config):
|
1491 |
+
super().__init__(config)
|
1492 |
+
|
1493 |
+
self.text_model = Kosmos2TextModel(config.text_config)
|
1494 |
+
self.vision_model = Kosmos2VisionModel(config.vision_config)
|
1495 |
+
self.image_to_text_connector = Kosmos2ImageToTextConnector(config)
|
1496 |
+
|
1497 |
+
# Initialize weights and apply final processing
|
1498 |
+
self.post_init()
|
1499 |
+
|
1500 |
+
def get_input_embeddings(self) -> nn.Module:
|
1501 |
+
return self.text_model.model.embed_tokens
|
1502 |
+
|
1503 |
+
def set_input_embeddings(self, value):
|
1504 |
+
self.text_model.model.embed_tokens = value
|
1505 |
+
|
1506 |
+
@add_start_docstrings_to_model_forward(KOSMOS2_INPUTS_DOCSTRING)
|
1507 |
+
@replace_return_docstrings(output_type=Kosmos2ModelOutput, config_class=Kosmos2Config)
|
1508 |
+
def forward(
|
1509 |
+
self,
|
1510 |
+
pixel_values: Optional[torch.Tensor] = None,
|
1511 |
+
input_ids: Optional[torch.Tensor] = None,
|
1512 |
+
attention_mask: Optional[torch.Tensor] = None,
|
1513 |
+
img_attn_mask: Optional[torch.Tensor] = None,
|
1514 |
+
head_mask: Optional[torch.Tensor] = None,
|
1515 |
+
img_features: Optional[torch.Tensor] = None,
|
1516 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
1517 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
1518 |
+
use_cache: Optional[bool] = None,
|
1519 |
+
output_attentions: Optional[bool] = None,
|
1520 |
+
output_hidden_states: Optional[bool] = None,
|
1521 |
+
return_dict: Optional[bool] = None,
|
1522 |
+
) -> Union[Tuple, Kosmos2ModelOutput]:
|
1523 |
+
# TODO: Add this
|
1524 |
+
r"""
|
1525 |
+
Returns:
|
1526 |
+
|
1527 |
+
```"""
|
1528 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
1529 |
+
output_hidden_states = (
|
1530 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
1531 |
+
)
|
1532 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
1533 |
+
|
1534 |
+
vision_model_output = None
|
1535 |
+
image_connector_attention = None
|
1536 |
+
if img_features is None:
|
1537 |
+
if pixel_values is None:
|
1538 |
+
raise ValueError("You have to specify either `pixel_values` or `img_features`.")
|
1539 |
+
|
1540 |
+
vision_model_output = self.vision_model(pixel_values)
|
1541 |
+
# HF's CLIP has `last_hidden_state` without going through `post_layernorm`.
|
1542 |
+
# Here we need the whole `last_hidden_state` through `post_layernorm` instead of just `pooled_output`.
|
1543 |
+
img_features = self.vision_model.model.post_layernorm(vision_model_output.last_hidden_state)
|
1544 |
+
# normalized features
|
1545 |
+
img_features = nn.functional.normalize(img_features, dim=-1)
|
1546 |
+
img_features, image_connector_attention = self.image_to_text_connector(img_features)
|
1547 |
+
|
1548 |
+
outputs = self.text_model(
|
1549 |
+
input_ids=input_ids,
|
1550 |
+
attention_mask=attention_mask,
|
1551 |
+
img_features=img_features,
|
1552 |
+
img_attn_mask=img_attn_mask,
|
1553 |
+
head_mask=head_mask,
|
1554 |
+
past_key_values=past_key_values,
|
1555 |
+
inputs_embeds=inputs_embeds,
|
1556 |
+
use_cache=use_cache,
|
1557 |
+
output_attentions=output_attentions,
|
1558 |
+
output_hidden_states=output_hidden_states,
|
1559 |
+
return_dict=return_dict,
|
1560 |
+
)
|
1561 |
+
|
1562 |
+
if not return_dict:
|
1563 |
+
outputs = outputs + (img_features, image_connector_attention, vision_model_output)
|
1564 |
+
return tuple(output for output in outputs if output is not None)
|
1565 |
+
|
1566 |
+
return Kosmos2ModelOutput(
|
1567 |
+
last_hidden_states=outputs.last_hidden_state,
|
1568 |
+
past_key_values=outputs.past_key_values,
|
1569 |
+
hidden_states=outputs.hidden_states,
|
1570 |
+
attentions=outputs.attentions,
|
1571 |
+
image_features=img_features,
|
1572 |
+
image_connector_attention=image_connector_attention,
|
1573 |
+
vision_model_output=vision_model_output,
|
1574 |
+
)
|
1575 |
+
|
1576 |
+
|
1577 |
+
@add_start_docstrings(
|
1578 |
+
"""
|
1579 |
+
KOSMOS-2 Model for generating text and bounding boxes given an image. The model consists of a vision encoder (CLIP)
|
1580 |
+
and a language model.
|
1581 |
+
""",
|
1582 |
+
KOSMOS2_START_DOCSTRING,
|
1583 |
+
)
|
1584 |
+
class Kosmos2ForConditionalGeneration(Kosmos2PreTrainedModel):
|
1585 |
+
config_class = Kosmos2Config
|
1586 |
+
_tied_weights_keys = ["text_model.lm_head.weight"]
|
1587 |
+
|
1588 |
+
def __init__(self, config: Kosmos2Config):
|
1589 |
+
super().__init__(config)
|
1590 |
+
|
1591 |
+
self.text_model = Kosmos2TextForCausalLM(config.text_config)
|
1592 |
+
self.vision_model = Kosmos2VisionModel(config.vision_config)
|
1593 |
+
|
1594 |
+
self.image_to_text_connector = Kosmos2ImageToTextConnector(config)
|
1595 |
+
|
1596 |
+
# Initialize weights and apply final processing
|
1597 |
+
self.post_init()
|
1598 |
+
|
1599 |
+
def get_input_embeddings(self) -> nn.Module:
|
1600 |
+
return self.text_model.model.embed_tokens
|
1601 |
+
|
1602 |
+
def set_input_embeddings(self, value):
|
1603 |
+
self.text_model.model.embed_tokens = value
|
1604 |
+
|
1605 |
+
def get_output_embeddings(self) -> nn.Module:
|
1606 |
+
return self.text_model.get_output_embeddings()
|
1607 |
+
|
1608 |
+
def set_output_embeddings(self, new_embeddings):
|
1609 |
+
self.text_model.set_output_embeddings(new_embeddings)
|
1610 |
+
|
1611 |
+
@add_start_docstrings_to_model_forward(KOSMOS2_INPUTS_DOCSTRING)
|
1612 |
+
@replace_return_docstrings(output_type=Kosmos2ForConditionalGenerationModelOutput, config_class=Kosmos2Config)
|
1613 |
+
def forward(
|
1614 |
+
self,
|
1615 |
+
pixel_values: Optional[torch.Tensor] = None,
|
1616 |
+
img_attn_mask=None,
|
1617 |
+
input_ids: Optional[torch.Tensor] = None,
|
1618 |
+
attention_mask=None,
|
1619 |
+
head_mask: Optional[torch.Tensor] = None,
|
1620 |
+
img_features: Optional[List[torch.FloatTensor]] = None,
|
1621 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
1622 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
1623 |
+
labels: Optional[torch.LongTensor] = None,
|
1624 |
+
use_cache: Optional[bool] = None,
|
1625 |
+
output_attentions: Optional[bool] = None,
|
1626 |
+
output_hidden_states: Optional[bool] = None,
|
1627 |
+
return_dict: Optional[bool] = None,
|
1628 |
+
) -> Union[Tuple, Kosmos2ForConditionalGenerationModelOutput]:
|
1629 |
+
r"""
|
1630 |
+
Returns:
|
1631 |
+
|
1632 |
+
Examples:
|
1633 |
+
|
1634 |
+
```python
|
1635 |
+
>>> from PIL import Image
|
1636 |
+
>>> from transformers import AutoProcessor, Kosmos2ForConditionalGeneration
|
1637 |
+
|
1638 |
+
>>> model = Kosmos2ForConditionalGeneration.from_pretrained("ydshieh/kosmos-2-patch14-224")
|
1639 |
+
>>> processor = AutoProcessor.from_pretrained("ydshieh/kosmos-2-patch14-224")
|
1640 |
+
|
1641 |
+
>>> prompt = "<grounding> An image of"
|
1642 |
+
>>> image = Image.open("snowman.jpg")
|
1643 |
+
|
1644 |
+
>>> inputs = processor(text=prompt, images=image, return_tensors="pt")
|
1645 |
+
|
1646 |
+
>>> generated_ids = model.generate(
|
1647 |
+
... pixel_values=inputs["pixel_values"],
|
1648 |
+
... input_ids=inputs["input_ids"][:, :-1],
|
1649 |
+
... attention_mask=inputs["attention_mask"][:, :-1],
|
1650 |
+
... img_features=None,
|
1651 |
+
... img_attn_mask=inputs["img_attn_mask"][:, :-1],
|
1652 |
+
... use_cache=True,
|
1653 |
+
... max_new_tokens=64,
|
1654 |
+
... )
|
1655 |
+
|
1656 |
+
>>> generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
|
1657 |
+
>>> result = processor.post_processor_generation(generated_text)
|
1658 |
+
>>> result
|
1659 |
+
<grounding> An image of<phrase> a snowman</phrase><object><patch_index_0044><patch_index_0863></object> warming himself by<phrase> a fire</phrase><object><patch_index_0005><patch_index_0911></object>.
|
1660 |
+
```"""
|
1661 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
1662 |
+
output_hidden_states = (
|
1663 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
1664 |
+
)
|
1665 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
1666 |
+
|
1667 |
+
vision_model_output = None
|
1668 |
+
image_connector_attention = None
|
1669 |
+
if img_features is None:
|
1670 |
+
if pixel_values is None:
|
1671 |
+
raise ValueError("You have to specify either `pixel_values` or `img_features`.")
|
1672 |
+
|
1673 |
+
vision_model_output = self.vision_model(pixel_values)
|
1674 |
+
# HF's CLIP has `last_hidden_state` without going through `post_layernorm`.
|
1675 |
+
# Here we need the whole `last_hidden_state` through `post_layernorm` instead of just `pooled_output`.
|
1676 |
+
img_features = self.vision_model.model.post_layernorm(vision_model_output.last_hidden_state)
|
1677 |
+
# normalized features
|
1678 |
+
img_features = nn.functional.normalize(img_features, dim=-1)
|
1679 |
+
img_features, image_connector_attention = self.image_to_text_connector(img_features)
|
1680 |
+
|
1681 |
+
lm_outputs = self.text_model(
|
1682 |
+
input_ids=input_ids,
|
1683 |
+
attention_mask=attention_mask,
|
1684 |
+
img_features=img_features,
|
1685 |
+
img_attn_mask=img_attn_mask,
|
1686 |
+
head_mask=head_mask,
|
1687 |
+
past_key_values=past_key_values,
|
1688 |
+
inputs_embeds=inputs_embeds,
|
1689 |
+
labels=labels,
|
1690 |
+
use_cache=use_cache,
|
1691 |
+
output_attentions=output_attentions,
|
1692 |
+
output_hidden_states=output_hidden_states,
|
1693 |
+
return_dict=return_dict,
|
1694 |
+
)
|
1695 |
+
|
1696 |
+
if not return_dict:
|
1697 |
+
outputs = lm_outputs + (img_features, image_connector_attention, vision_model_output)
|
1698 |
+
return tuple(output for output in outputs if output is not None)
|
1699 |
+
|
1700 |
+
return Kosmos2ForConditionalGenerationModelOutput(
|
1701 |
+
loss=lm_outputs.loss,
|
1702 |
+
logits=lm_outputs.logits,
|
1703 |
+
past_key_values=lm_outputs.past_key_values,
|
1704 |
+
hidden_states=lm_outputs.hidden_states,
|
1705 |
+
attentions=lm_outputs.attentions,
|
1706 |
+
image_features=img_features,
|
1707 |
+
image_connector_attention=image_connector_attention,
|
1708 |
+
vision_model_output=vision_model_output,
|
1709 |
+
)
|
1710 |
+
|
1711 |
+
def generate(
|
1712 |
+
self,
|
1713 |
+
input_ids=None,
|
1714 |
+
attention_mask=None,
|
1715 |
+
img_features=None,
|
1716 |
+
inputs_embeds=None,
|
1717 |
+
pixel_values=None,
|
1718 |
+
**kwargs,
|
1719 |
+
):
|
1720 |
+
# in order to allow `inputs` argument (as in `GenerationMixin`)
|
1721 |
+
inputs = kwargs.pop("inputs", None)
|
1722 |
+
if pixel_values is not None and inputs is not None:
|
1723 |
+
raise ValueError(
|
1724 |
+
f"`inputs`: {inputs} were passed alongside `pixel_values` which is not allowed."
|
1725 |
+
f"Make sure to either pass `inputs` or pixel_values=..."
|
1726 |
+
)
|
1727 |
+
if pixel_values is None and inputs is not None:
|
1728 |
+
pixel_values = inputs
|
1729 |
+
|
1730 |
+
if img_features is None:
|
1731 |
+
vision_model_output = self.vision_model(pixel_values)
|
1732 |
+
# HF's CLIP has `last_hidden_state` without going through `post_layernorm`.
|
1733 |
+
# Here we need the whole `last_hidden_state` through `post_layernorm` instead of just `pooled_output`.
|
1734 |
+
img_features = self.vision_model.model.post_layernorm(vision_model_output.last_hidden_state)
|
1735 |
+
# normalized features
|
1736 |
+
img_features = nn.functional.normalize(img_features, dim=-1)
|
1737 |
+
img_features, image_connector_attention = self.image_to_text_connector(img_features)
|
1738 |
+
|
1739 |
+
output = self.text_model.generate(
|
1740 |
+
input_ids=input_ids,
|
1741 |
+
attention_mask=attention_mask,
|
1742 |
+
img_features=img_features,
|
1743 |
+
input_embeds=inputs_embeds,
|
1744 |
+
**kwargs,
|
1745 |
+
)
|
1746 |
+
|
1747 |
+
return output
|
processing_kosmos2.py
ADDED
@@ -0,0 +1,498 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# coding=utf-8
|
2 |
+
# Copyright 2023 The HuggingFace Inc. team.
|
3 |
+
#
|
4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
5 |
+
# you may not use this file except in compliance with the License.
|
6 |
+
# You may obtain a copy of the License at
|
7 |
+
#
|
8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
9 |
+
#
|
10 |
+
# Unless required by applicable law or agreed to in writing, software
|
11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
+
# See the License for the specific language governing permissions and
|
14 |
+
# limitations under the License.
|
15 |
+
"""Processor class for KOSMOS-2."""
|
16 |
+
|
17 |
+
import copy
|
18 |
+
import math
|
19 |
+
import re
|
20 |
+
from typing import List, Optional, Tuple, Union
|
21 |
+
|
22 |
+
import numpy as np
|
23 |
+
|
24 |
+
from ...image_processing_utils import BatchFeature
|
25 |
+
from ...image_utils import ImageInput, is_batched
|
26 |
+
from ...processing_utils import ProcessorMixin
|
27 |
+
from ...tokenization_utils_base import PaddingStrategy, TextInput, TruncationStrategy
|
28 |
+
from ...utils import TensorType, is_tf_available, is_torch_available
|
29 |
+
|
30 |
+
|
31 |
+
if is_torch_available():
|
32 |
+
import torch
|
33 |
+
|
34 |
+
if is_tf_available():
|
35 |
+
import tensorflow as tf
|
36 |
+
|
37 |
+
|
38 |
+
BboxInput = Union[
|
39 |
+
List[Tuple[int, int]],
|
40 |
+
List[Tuple[float, float, float, float]],
|
41 |
+
List[List[Tuple[int, int]]],
|
42 |
+
List[List[Tuple[float, float, float]]],
|
43 |
+
]
|
44 |
+
|
45 |
+
|
46 |
+
class Kosmos2Processor(ProcessorMixin):
|
47 |
+
r"""
|
48 |
+
Constructs an KOSMOS-2 processor which wraps a CLIP image processor and a KOSMOS-2 tokenizer into a single
|
49 |
+
processor.
|
50 |
+
|
51 |
+
[`Kosmos2Processor`] offers all the functionalities of [`CLIPImageProcessor`] and [`Kosmos2TokenizerFast`]. See the
|
52 |
+
docstring of [`~Kosmos2Processor.__call__`] and [`~Kosmos2Processor.decode`] for more information.
|
53 |
+
|
54 |
+
Args:
|
55 |
+
image_processor (`CLIPImageProcessor`):
|
56 |
+
An instance of [`CLIPImageProcessor`]. The image processor is a required input.
|
57 |
+
tokenizer (`Kosmos2TokenizerFast`):
|
58 |
+
An instance of ['Kosmos2TokenizerFast`]. The tokenizer is a required input.
|
59 |
+
"""
|
60 |
+
attributes = ["image_processor", "tokenizer"]
|
61 |
+
image_processor_class = "CLIPImageProcessor"
|
62 |
+
tokenizer_class = ("Kosmos2Tokenizer", "Kosmos2TokenizerFast")
|
63 |
+
|
64 |
+
def __init__(self, image_processor, tokenizer):
|
65 |
+
tokenizer.return_token_type_ids = False
|
66 |
+
super().__init__(image_processor, tokenizer)
|
67 |
+
self.current_processor = self.image_processor
|
68 |
+
|
69 |
+
def __call__(
|
70 |
+
self,
|
71 |
+
images: ImageInput = None,
|
72 |
+
text: Union[TextInput, List[TextInput]] = None,
|
73 |
+
bboxes: BboxInput = None,
|
74 |
+
num_image_tokens: Optional[int] = 64,
|
75 |
+
first_image_token_id: Optional[int] = None,
|
76 |
+
add_special_tokens: bool = True,
|
77 |
+
padding: Union[bool, str, PaddingStrategy] = False,
|
78 |
+
truncation: Union[bool, str, TruncationStrategy] = None,
|
79 |
+
max_length: Optional[int] = None,
|
80 |
+
stride: int = 0,
|
81 |
+
pad_to_multiple_of: Optional[int] = None,
|
82 |
+
return_attention_mask: Optional[bool] = None,
|
83 |
+
return_overflowing_tokens: bool = False,
|
84 |
+
return_special_tokens_mask: bool = False,
|
85 |
+
return_offsets_mapping: bool = False,
|
86 |
+
return_token_type_ids: bool = False,
|
87 |
+
return_length: bool = False,
|
88 |
+
verbose: bool = True,
|
89 |
+
return_tensors: Optional[Union[str, TensorType]] = None,
|
90 |
+
**kwargs,
|
91 |
+
) -> BatchFeature:
|
92 |
+
"""
|
93 |
+
This method uses [`CLIPImageProcessor.__call__`] method to prepare image(s) for the model, and
|
94 |
+
[`Kosmos2TokenizerFast.__call__`] to prepare text for the model.
|
95 |
+
|
96 |
+
Please refer to the docstring of the above two methods for more information.
|
97 |
+
"""
|
98 |
+
if text is None:
|
99 |
+
raise ValueError("You have to specify at least `text`.")
|
100 |
+
|
101 |
+
text = self.preprocess_text(text, images, bboxes, num_image_tokens=num_image_tokens)
|
102 |
+
|
103 |
+
encoding = BatchFeature()
|
104 |
+
|
105 |
+
text_encoding = self.tokenizer(
|
106 |
+
text=text,
|
107 |
+
add_special_tokens=add_special_tokens,
|
108 |
+
padding=padding,
|
109 |
+
truncation=truncation,
|
110 |
+
max_length=max_length,
|
111 |
+
stride=stride,
|
112 |
+
pad_to_multiple_of=pad_to_multiple_of,
|
113 |
+
return_attention_mask=return_attention_mask,
|
114 |
+
return_overflowing_tokens=return_overflowing_tokens,
|
115 |
+
return_special_tokens_mask=return_special_tokens_mask,
|
116 |
+
return_offsets_mapping=return_offsets_mapping,
|
117 |
+
return_token_type_ids=return_token_type_ids,
|
118 |
+
return_length=return_length,
|
119 |
+
verbose=verbose,
|
120 |
+
return_tensors=return_tensors,
|
121 |
+
**kwargs,
|
122 |
+
)
|
123 |
+
encoding.update(text_encoding)
|
124 |
+
|
125 |
+
if images is not None:
|
126 |
+
image_encoding = self.image_processor(images, return_tensors=return_tensors)
|
127 |
+
encoding.update(image_encoding)
|
128 |
+
|
129 |
+
# Use the id of the first token after <unk>
|
130 |
+
if first_image_token_id is None:
|
131 |
+
first_image_token_id = self.tokenizer.unk_token_id + 1
|
132 |
+
|
133 |
+
# To see if we need one more `0` (for `<s>`) at the beginning of `img_attn_mask`.
|
134 |
+
with_bos = add_special_tokens
|
135 |
+
|
136 |
+
# The first (actual) `<image>` token is always at the 1st or 2nd place (after `<s>` if any). Here we look
|
137 |
+
# for the second `<image>` token (which indicate the first image token).
|
138 |
+
start_index = int(with_bos) + 1
|
139 |
+
|
140 |
+
if return_tensors:
|
141 |
+
# change the ids for the fake `<image>` tokens in `input_ids`
|
142 |
+
input_ids = np.array(encoding["input_ids"])
|
143 |
+
input_ids[:, start_index : (start_index + num_image_tokens)] = np.arange(
|
144 |
+
first_image_token_id, first_image_token_id + num_image_tokens
|
145 |
+
)
|
146 |
+
|
147 |
+
batch_size, seq_len = input_ids.shape[:2]
|
148 |
+
img_attn_mask = []
|
149 |
+
if with_bos:
|
150 |
+
# for `<s>`
|
151 |
+
img_attn_mask.append(np.zeros(shape=(batch_size, 1), dtype=np.int64))
|
152 |
+
# for `<image>` (the real one)
|
153 |
+
img_attn_mask.append(np.zeros(shape=(batch_size, 1), dtype=np.int64))
|
154 |
+
# for image tokens
|
155 |
+
img_attn_mask.append(np.ones(shape=(batch_size, 64), dtype=np.int64))
|
156 |
+
# for `</image>`
|
157 |
+
img_attn_mask.append(np.zeros(shape=(batch_size, 1), dtype=np.int64))
|
158 |
+
# trailing part (which are not related to the image)
|
159 |
+
seq_len -= int(with_bos) + 1 + num_image_tokens + 1
|
160 |
+
img_attn_mask.append(np.zeros(shape=(batch_size, seq_len), dtype=np.int64))
|
161 |
+
|
162 |
+
# concatenate along the sequence dimension
|
163 |
+
img_attn_mask = np.concatenate(img_attn_mask, axis=1)
|
164 |
+
|
165 |
+
# to the target tensor type
|
166 |
+
if return_tensors == "pt":
|
167 |
+
input_ids = torch.from_numpy(input_ids)
|
168 |
+
img_attn_mask = torch.from_numpy(img_attn_mask)
|
169 |
+
elif return_tensors == "tf":
|
170 |
+
input_ids = tf.convert_to_tensor(input_ids)
|
171 |
+
img_attn_mask = tf.convert_to_tensor(img_attn_mask)
|
172 |
+
|
173 |
+
encoding["input_ids"] = input_ids
|
174 |
+
encoding["img_attn_mask"] = img_attn_mask
|
175 |
+
|
176 |
+
else:
|
177 |
+
# Add `img_attn_mask`: the leading and trailing `0` are for `boi` and `eoi` tokens. The `1` indicates
|
178 |
+
# the places of image tokens.
|
179 |
+
image_token_ids = list(range(first_image_token_id, first_image_token_id + num_image_tokens))
|
180 |
+
base_img_attn_mask = [0] + [1] * num_image_tokens + [0]
|
181 |
+
|
182 |
+
# loop over `encoding["input_ids"]`
|
183 |
+
input_ids = []
|
184 |
+
img_attn_mask = []
|
185 |
+
all_input_ids = encoding["input_ids"]
|
186 |
+
# not batched -> (changed to) batch of size 1
|
187 |
+
if isinstance(text, str):
|
188 |
+
all_input_ids = [all_input_ids]
|
189 |
+
for text_ids in all_input_ids:
|
190 |
+
# change the ids for the fake `<image>` tokens in `input_ids`
|
191 |
+
text_ids = text_ids[:start_index] + image_token_ids + text_ids[start_index + num_image_tokens :]
|
192 |
+
input_ids.append(text_ids)
|
193 |
+
|
194 |
+
mask = copy.copy(base_img_attn_mask)
|
195 |
+
if with_bos:
|
196 |
+
# for `<s>`
|
197 |
+
mask = [0] + mask
|
198 |
+
# trailing part (which are not related to the image)
|
199 |
+
mask += [0] * (len(text_ids) - len(mask))
|
200 |
+
img_attn_mask.append(mask)
|
201 |
+
|
202 |
+
# un-batch if necessary
|
203 |
+
if isinstance(text, str):
|
204 |
+
input_ids = input_ids[0]
|
205 |
+
img_attn_mask = img_attn_mask[0]
|
206 |
+
|
207 |
+
encoding["input_ids"] = input_ids
|
208 |
+
encoding["img_attn_mask"] = img_attn_mask
|
209 |
+
|
210 |
+
return encoding
|
211 |
+
|
212 |
+
def preprocess_text(
|
213 |
+
self,
|
214 |
+
texts: Union[TextInput, List[TextInput]],
|
215 |
+
images: ImageInput = None,
|
216 |
+
bboxes: BboxInput = None,
|
217 |
+
num_image_tokens: Optional[int] = 64,
|
218 |
+
) -> Union[str, List[str]]:
|
219 |
+
"""Add image and bounding box information to `texts` as image and patch index tokens.
|
220 |
+
|
221 |
+
Args:
|
222 |
+
texts (`Union[TextInput, List[TextInput]]`): The texts to be processed.
|
223 |
+
images (`ImageInput`, *optional*): The images associated to `texts`.
|
224 |
+
bboxes (`Union[List[Tuple[int]], List[Tuple[float]], List[List[Tuple[int]]], List[List[Tuple[float]]]]`, *optional*): The bounding bboxes associated to `texts`.
|
225 |
+
num_image_tokens (`int`, *optional*, defaults to 64): The number of image tokens (used as latent queries). This should corresponds to the `latent_query_num` attribute in `Kosmos2Config`.
|
226 |
+
|
227 |
+
Returns:
|
228 |
+
`Union[TextInput, List[TextInput]]`: The processed texts with image and patch index tokens.
|
229 |
+
"""
|
230 |
+
# These are fake `<image>` tokens enclosed between (the actual) `<image>` token and `</image>`.
|
231 |
+
img_tokens = ["<image>"] * num_image_tokens
|
232 |
+
img_info = " ".join(["<image>"] + img_tokens + ["</image>"])
|
233 |
+
|
234 |
+
def check_bboxes_for_single_text(bboxes):
|
235 |
+
"""
|
236 |
+
Check `bboxes` for a single text example. It could be
|
237 |
+
- `None`: no bounding box associated to a text.
|
238 |
+
- A list with each element being the bounding boxes associated to one `<phrase> ... </phrase>` pair
|
239 |
+
found in a text. This could be:
|
240 |
+
- `None`: no bounding box associated to a `<phrase> ... </phrase>` pair.
|
241 |
+
- A tuple of 2 integers: A single bounding box specified by patch indices.
|
242 |
+
- A tuple of 4 float point number: A single bounding box specified by (normalized) coordinates.
|
243 |
+
- A list containing the above 2 tuple types: Multiple bounding boxes for a
|
244 |
+
`<phrase> ... </phrase>` pair.
|
245 |
+
"""
|
246 |
+
if bboxes is None:
|
247 |
+
return
|
248 |
+
elif not isinstance(bboxes, list):
|
249 |
+
raise ValueError("`bboxes` (for a single text example) should be `None` or a list.")
|
250 |
+
|
251 |
+
# `bbox` is the bounding boxes for a single <phrase> </phrase> pair
|
252 |
+
for bbox in bboxes:
|
253 |
+
if bbox is None:
|
254 |
+
continue
|
255 |
+
elif not isinstance(bbox, list):
|
256 |
+
bbox = [bbox]
|
257 |
+
for elt in bbox:
|
258 |
+
if not isinstance(elt, tuple) or not (
|
259 |
+
(len(elt) == 2 and all(isinstance(x, int) for x in elt))
|
260 |
+
or (len(elt) == 4 and all(isinstance(x, float) for x in elt))
|
261 |
+
):
|
262 |
+
raise ValueError(
|
263 |
+
"Each element in `bboxes` (for a single text example) should be `None`, a tuple containing "
|
264 |
+
"2 integers or 4 float point numbers, or a list containing such tuples. Also "
|
265 |
+
"make sure the arguments `texts` and `bboxes` passed to `preprocess_text` are both in "
|
266 |
+
"batches or both for a single example."
|
267 |
+
)
|
268 |
+
|
269 |
+
def preprocess_single(text, image, bboxes):
|
270 |
+
if image is not None:
|
271 |
+
# Add `<image> ... (fake) image tokens ... </image>`
|
272 |
+
text = f"{img_info} {text}"
|
273 |
+
|
274 |
+
# Add `<object> <patch_idx_xxxx> <patch_idx_yyy> </object>` after `<phrase> phrase text </phrase>`
|
275 |
+
text = self._insert_patch_index_tokens(text, bboxes)
|
276 |
+
text = self._add_remove_spaces_around_tag_tokens(text)
|
277 |
+
|
278 |
+
return text
|
279 |
+
|
280 |
+
# make batch to simplify processing logic
|
281 |
+
batched = True
|
282 |
+
if isinstance(texts, str):
|
283 |
+
batched = False
|
284 |
+
texts = [texts]
|
285 |
+
|
286 |
+
if images is None:
|
287 |
+
images = [None] * len(texts)
|
288 |
+
elif not is_batched(images):
|
289 |
+
images = [images]
|
290 |
+
if len(texts) != len(images):
|
291 |
+
raise ValueError(
|
292 |
+
f"The number of examples in `texts` and `images` should be the same. Got {len(texts)} v.s. {len(images)} instead."
|
293 |
+
)
|
294 |
+
|
295 |
+
if not batched:
|
296 |
+
check_bboxes_for_single_text(bboxes)
|
297 |
+
bboxes = [bboxes]
|
298 |
+
elif bboxes is not None:
|
299 |
+
if not isinstance(bboxes, list):
|
300 |
+
raise ValueError("`bboxes` should be `None` or a list (as a batch) when `texts` is passed as a batch.")
|
301 |
+
for x in bboxes:
|
302 |
+
check_bboxes_for_single_text(x)
|
303 |
+
else:
|
304 |
+
bboxes = [None] * len(texts)
|
305 |
+
|
306 |
+
if len(bboxes) != len(texts):
|
307 |
+
raise ValueError(
|
308 |
+
f"The number of examples in `texts` and `bboxes` should be the same. Got {len(texts)} v.s. {len(bboxes)} instead."
|
309 |
+
)
|
310 |
+
|
311 |
+
result = [preprocess_single(text, image, bbox) for text, image, bbox in zip(texts, images, bboxes)]
|
312 |
+
# un-batch if necessary
|
313 |
+
if not batched:
|
314 |
+
result = result[0]
|
315 |
+
|
316 |
+
return result
|
317 |
+
|
318 |
+
# Copied from transformers.models.blip.processing_blip.BlipProcessor.batch_decode with BertTokenizerFast->PreTrainedTokenizer
|
319 |
+
def batch_decode(self, *args, **kwargs):
|
320 |
+
"""
|
321 |
+
This method forwards all its arguments to PreTrainedTokenizer's [`~PreTrainedTokenizer.batch_decode`]. Please
|
322 |
+
refer to the docstring of this method for more information.
|
323 |
+
"""
|
324 |
+
return self.tokenizer.batch_decode(*args, **kwargs)
|
325 |
+
|
326 |
+
# Copied from transformers.models.blip.processing_blip.BlipProcessor.decode with BertTokenizerFast->PreTrainedTokenizer
|
327 |
+
def decode(self, *args, **kwargs):
|
328 |
+
"""
|
329 |
+
This method forwards all its arguments to PreTrainedTokenizer's [`~PreTrainedTokenizer.decode`]. Please refer
|
330 |
+
to the docstring of this method for more information.
|
331 |
+
"""
|
332 |
+
return self.tokenizer.decode(*args, **kwargs)
|
333 |
+
|
334 |
+
def post_processor_generation(self, text):
|
335 |
+
return text.split("</image>")[-1]
|
336 |
+
|
337 |
+
@property
|
338 |
+
# Copied from transformers.models.blip.processing_blip.BlipProcessor.model_input_names
|
339 |
+
def model_input_names(self):
|
340 |
+
tokenizer_input_names = self.tokenizer.model_input_names
|
341 |
+
image_processor_input_names = self.image_processor.model_input_names
|
342 |
+
return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
|
343 |
+
|
344 |
+
def _insert_patch_index_tokens(self, text: str, bboxes: Union[List[Tuple[int]], List[Tuple[float]]]) -> str:
|
345 |
+
if bboxes is None or len(bboxes) == 0:
|
346 |
+
return text
|
347 |
+
|
348 |
+
matched_phrases = list(re.finditer(r"<phrase>.+?</phrase>", string=text))
|
349 |
+
if len(matched_phrases) != len(bboxes):
|
350 |
+
raise ValueError(
|
351 |
+
f"The number of elements in `bboxes` should be the same as the number of `<phrase> ... </phrase>` pairs in `text`. Got {len(matched_phrases)} v.s. {len(bboxes)} instead."
|
352 |
+
)
|
353 |
+
|
354 |
+
# insert object's patch index tokens
|
355 |
+
# the found `<phrase> ... </phrase>` pairs.
|
356 |
+
curr_pos = 0
|
357 |
+
buffer = []
|
358 |
+
for matched, bbox in zip(matched_phrases, bboxes):
|
359 |
+
_, end = matched.span()
|
360 |
+
buffer.append(text[curr_pos:end])
|
361 |
+
curr_pos = end
|
362 |
+
# A phrase without bbox
|
363 |
+
if bbox is None:
|
364 |
+
continue
|
365 |
+
# A phrase with a single bbox
|
366 |
+
if isinstance(bbox, tuple):
|
367 |
+
bbox = [bbox]
|
368 |
+
patch_index_strings = []
|
369 |
+
# A phrase could have multiple bboxes
|
370 |
+
for box in bbox:
|
371 |
+
patch_index_1, patch_index_2 = self._convert_bbox_to_patch_index_tokens(box)
|
372 |
+
patch_index_strings.append(f"{patch_index_1} {patch_index_2}")
|
373 |
+
position_str = " </delimiter_of_multi_objects/> ".join(patch_index_strings)
|
374 |
+
buffer.append(f"<object> {position_str} </object>")
|
375 |
+
# remaining
|
376 |
+
if curr_pos < len(text):
|
377 |
+
buffer.append(text[curr_pos:])
|
378 |
+
|
379 |
+
text = "".join(buffer)
|
380 |
+
return text
|
381 |
+
|
382 |
+
def _convert_bbox_to_patch_index_tokens(
|
383 |
+
self, bbox: Union[Tuple[int, int], Tuple[float, float, float, float]]
|
384 |
+
) -> Tuple[str, str]:
|
385 |
+
# already computed patch indices
|
386 |
+
if len(bbox) == 2:
|
387 |
+
idx_1, idx_2 = bbox
|
388 |
+
# bbox specified with (normalized) coordinates
|
389 |
+
else:
|
390 |
+
# use `self.tokenizer` to get `num_patches_per_side`
|
391 |
+
num_patches_per_side = int(math.sqrt(self.tokenizer.num_patch_index_tokens))
|
392 |
+
idx_1, idx_2 = coordinate_to_patch_index(bbox, num_patches_per_side)
|
393 |
+
|
394 |
+
token_1 = f"<patch_index_{str(idx_1).zfill(4)}>"
|
395 |
+
token_2 = f"<patch_index_{str(idx_2).zfill(4)}>"
|
396 |
+
|
397 |
+
return token_1, token_2
|
398 |
+
|
399 |
+
def _add_remove_spaces_around_tag_tokens(self, text):
|
400 |
+
"""
|
401 |
+
Remove spaces before tag tokens (e.g. `<x>`). Also ensure a space after a tag token, if it is not followed by
|
402 |
+
another tag token (this is not technically necessary, but good for a standard/consistent format). This avoids
|
403 |
+
the inconsistency of tokenization results between kosmos-2 slow and fast tokenizers.
|
404 |
+
"""
|
405 |
+
|
406 |
+
tag_tokens = set(
|
407 |
+
self.tokenizer.tag_tokens
|
408 |
+
+ [f"<patch_index_{str(x).zfill(4)}>" for x in range(self.tokenizer.num_patch_index_tokens)]
|
409 |
+
)
|
410 |
+
pattern = "|".join(tag_tokens)
|
411 |
+
splits = re.split(rf"({pattern})", text)
|
412 |
+
|
413 |
+
output = ""
|
414 |
+
prev_str_in_targets = False
|
415 |
+
for split in splits:
|
416 |
+
if split in tag_tokens:
|
417 |
+
prev_str_in_targets = True
|
418 |
+
output = output.rstrip() + split
|
419 |
+
else:
|
420 |
+
# we don't need to ensure a space before a normal token that is after a tag token. But having it and
|
421 |
+
# keeps a standard format is good anyway.
|
422 |
+
if prev_str_in_targets and not split.startswith(" "):
|
423 |
+
output += " " + split
|
424 |
+
else:
|
425 |
+
output += split
|
426 |
+
prev_str_in_targets = False
|
427 |
+
|
428 |
+
return output
|
429 |
+
|
430 |
+
|
431 |
+
def coordinate_to_patch_index(bbox: Tuple[float, float, float, float], num_patches_per_side: int) -> Tuple[int, int]:
|
432 |
+
"""Convert a bounding box to a pair of patch indices.
|
433 |
+
|
434 |
+
Args:
|
435 |
+
bbox (`Tuple[float, float, float, float]`):
|
436 |
+
The 4 coordinates of the bounding box, with the format being (x1, y1, x2, y2) specifying the upper-left
|
437 |
+
and lower-right corners of the box. It should have x2 > x1 and y1 > y2.
|
438 |
+
num_patches_per_side (`int`): the number of patches along each side.
|
439 |
+
|
440 |
+
Returns:
|
441 |
+
`Tuple[int, int]`: A pair of patch indices.
|
442 |
+
"""
|
443 |
+
(x1, y1, x2, y2) = bbox
|
444 |
+
|
445 |
+
ul_x = math.floor(x1 * num_patches_per_side)
|
446 |
+
ul_y = math.floor(y1 * num_patches_per_side)
|
447 |
+
|
448 |
+
lr_x = math.ceil(x2 * num_patches_per_side - 1)
|
449 |
+
lr_y = math.ceil(y2 * num_patches_per_side - 1)
|
450 |
+
|
451 |
+
ul_idx = ul_y * num_patches_per_side + ul_x
|
452 |
+
lr_idx = lr_y * num_patches_per_side + lr_x
|
453 |
+
|
454 |
+
return ul_idx, lr_idx
|
455 |
+
|
456 |
+
|
457 |
+
# copied from https://github.com/microsoft/unilm/blob/97e4923e97d3ee10b57e97013556e3fd0d207a9b/kosmos-2/demo/decode_string.py#L35C1-L75C38
|
458 |
+
def patch_index_to_coordinate(ul_idx: int, lr_idx: int, num_patches_per_side: int):
|
459 |
+
"""
|
460 |
+
Given a grid of length `num_patches_per_side` and the indices of the upper-left and lower-right corners of a
|
461 |
+
bounding box, returns the normalized coordinates of the bounding box, in the form (x1, y1, x2, y2).
|
462 |
+
|
463 |
+
Args:
|
464 |
+
ul_idx (`int`): the index of the grid cell that corresponds to the upper-left corner of the bounding box.
|
465 |
+
lr_idx (`int`): the index of the grid cell that corresponds to the lower-right corner of the bounding box.
|
466 |
+
num_patches_per_side (`int`): the number of patches along each side.
|
467 |
+
|
468 |
+
Returns:
|
469 |
+
`Tuple[float]`: the normalized coordinates of the bounding box, in the form (x1, y1, x2, y2).
|
470 |
+
"""
|
471 |
+
# Compute the size of each cell in the grid
|
472 |
+
cell_size = 1.0 / num_patches_per_side
|
473 |
+
|
474 |
+
# Compute the x and y indices of the upper-left and lower-right corners of the bounding box
|
475 |
+
ul_x = ul_idx % num_patches_per_side
|
476 |
+
ul_y = ul_idx // num_patches_per_side
|
477 |
+
|
478 |
+
lr_x = lr_idx % num_patches_per_side
|
479 |
+
lr_y = lr_idx // num_patches_per_side
|
480 |
+
|
481 |
+
# Compute the normalized coordinates of the bounding box
|
482 |
+
if ul_idx == lr_idx:
|
483 |
+
x1 = ul_x * cell_size
|
484 |
+
y1 = ul_y * cell_size
|
485 |
+
x2 = lr_x * cell_size + cell_size
|
486 |
+
y2 = lr_y * cell_size + cell_size
|
487 |
+
elif ul_x == lr_x or ul_y == lr_y:
|
488 |
+
x1 = ul_x * cell_size
|
489 |
+
y1 = ul_y * cell_size
|
490 |
+
x2 = lr_x * cell_size + cell_size
|
491 |
+
y2 = lr_y * cell_size + cell_size
|
492 |
+
else:
|
493 |
+
x1 = ul_x * cell_size + cell_size / 2
|
494 |
+
y1 = ul_y * cell_size + cell_size / 2
|
495 |
+
x2 = lr_x * cell_size + cell_size / 2
|
496 |
+
y2 = lr_y * cell_size + cell_size / 2
|
497 |
+
|
498 |
+
return x1, y1, x2, y2
|
tokenization_kosmos2.py
ADDED
@@ -0,0 +1,413 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# coding=utf-8
|
2 |
+
# Copyright 2023 Microsoft Research and The HuggingFace Inc. team. All rights reserved.
|
3 |
+
#
|
4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
5 |
+
# you may not use this file except in compliance with the License.
|
6 |
+
# You may obtain a copy of the License at
|
7 |
+
#
|
8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
9 |
+
#
|
10 |
+
# Unless required by applicable law or agreed to in writing, software
|
11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
+
# See the License for the specific language governing permissions and
|
14 |
+
# limitations under the License.
|
15 |
+
""" Tokenization classes for KOSMOS-2 model."""
|
16 |
+
|
17 |
+
|
18 |
+
import os
|
19 |
+
from shutil import copyfile
|
20 |
+
from typing import Any, Dict, List, Optional, Tuple
|
21 |
+
|
22 |
+
import sentencepiece as spm
|
23 |
+
|
24 |
+
from ...tokenization_utils import AddedToken, PreTrainedTokenizer
|
25 |
+
from ...utils import logging
|
26 |
+
|
27 |
+
|
28 |
+
logger = logging.get_logger(__name__)
|
29 |
+
|
30 |
+
SPIECE_UNDERLINE = "▁"
|
31 |
+
|
32 |
+
VOCAB_FILES_NAMES = {"vocab_file": "sentencepiece.bpe.model"}
|
33 |
+
|
34 |
+
PRETRAINED_VOCAB_FILES_MAP = {
|
35 |
+
"vocab_file": {
|
36 |
+
"microsoft/kosmos-2-patch14-224": "https://huggingface.co/microsoft/kosmos-2-patch14-224/resolve/main/sentencepiece.bpe.model",
|
37 |
+
}
|
38 |
+
}
|
39 |
+
|
40 |
+
PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
|
41 |
+
"microsoft/kosmos-2-patch14-224": 2048,
|
42 |
+
}
|
43 |
+
|
44 |
+
|
45 |
+
class Kosmos2Tokenizer(PreTrainedTokenizer):
|
46 |
+
"""
|
47 |
+
Adapted from [`RobertaTokenizer`] and [`XLNetTokenizer`]. Based on
|
48 |
+
[SentencePiece](https://github.com/google/sentencepiece).
|
49 |
+
|
50 |
+
This tokenizer inherits from [`PreTrainedTokenizer`] which contains most of the main methods. Users should refer to
|
51 |
+
this superclass for more information regarding those methods.
|
52 |
+
|
53 |
+
Args:
|
54 |
+
vocab_file (`str`):
|
55 |
+
Path to the vocabulary file.
|
56 |
+
bos_token (`str`, *optional*, defaults to `"<s>"`):
|
57 |
+
The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.
|
58 |
+
|
59 |
+
<Tip>
|
60 |
+
|
61 |
+
When building a sequence using special tokens, this is not the token that is used for the beginning of
|
62 |
+
sequence. The token used is the `cls_token`.
|
63 |
+
|
64 |
+
</Tip>
|
65 |
+
|
66 |
+
eos_token (`str`, *optional*, defaults to `"</s>"`):
|
67 |
+
The end of sequence token.
|
68 |
+
|
69 |
+
<Tip>
|
70 |
+
|
71 |
+
When building a sequence using special tokens, this is not the token that is used for the end of sequence.
|
72 |
+
The token used is the `sep_token`.
|
73 |
+
|
74 |
+
</Tip>
|
75 |
+
|
76 |
+
sep_token (`str`, *optional*, defaults to `"</s>"`):
|
77 |
+
The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for
|
78 |
+
sequence classification or for a text and a question for question answering. It is also used as the last
|
79 |
+
token of a sequence built with special tokens.
|
80 |
+
cls_token (`str`, *optional*, defaults to `"<s>"`):
|
81 |
+
The classifier token which is used when doing sequence classification (classification of the whole sequence
|
82 |
+
instead of per-token classification). It is the first token of the sequence when built with special tokens.
|
83 |
+
unk_token (`str`, *optional*, defaults to `"<unk>"`):
|
84 |
+
The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
|
85 |
+
token instead.
|
86 |
+
pad_token (`str`, *optional*, defaults to `"<pad>"`):
|
87 |
+
The token used for padding, for example when batching sequences of different lengths.
|
88 |
+
mask_token (`str`, *optional*, defaults to `"<mask>"`):
|
89 |
+
The token used for masking values. This is the token used when training this model with masked language
|
90 |
+
modeling. This is the token which the model will try to predict.
|
91 |
+
additional_special_tokens (`List[str]`, *optional*, defaults to `["<s>NOTUSED", "</s>NOTUSED"]`):
|
92 |
+
Additional special tokens used by the tokenizer.
|
93 |
+
num_patch_index_tokens (`int`, *optional*, defaults to `1024`):
|
94 |
+
The number of tokens used to specify the patch indices of bounding boxes in an image. These tokens have the
|
95 |
+
format `<patch_index_xxxx>` where `xxxx` is an integer.
|
96 |
+
sp_model_kwargs (`dict`, *optional*):
|
97 |
+
Will be passed to the `SentencePieceProcessor.__init__()` method. The [Python wrapper for
|
98 |
+
SentencePiece](https://github.com/google/sentencepiece/tree/master/python) can be used, among other things,
|
99 |
+
to set:
|
100 |
+
|
101 |
+
- `enable_sampling`: Enable subword regularization.
|
102 |
+
- `nbest_size`: Sampling parameters for unigram. Invalid for BPE-Dropout.
|
103 |
+
|
104 |
+
- `nbest_size = {0,1}`: No sampling is performed.
|
105 |
+
- `nbest_size > 1`: samples from the nbest_size results.
|
106 |
+
- `nbest_size < 0`: assuming that nbest_size is infinite and samples from the all hypothesis (lattice)
|
107 |
+
using forward-filtering-and-backward-sampling algorithm.
|
108 |
+
|
109 |
+
- `alpha`: Smoothing parameter for unigram sampling, and dropout probability of merge operations for
|
110 |
+
BPE-dropout.
|
111 |
+
|
112 |
+
Attributes:
|
113 |
+
sp_model (`SentencePieceProcessor`):
|
114 |
+
The *SentencePiece* processor that is used for every conversion (string, tokens and IDs).
|
115 |
+
"""
|
116 |
+
|
117 |
+
vocab_files_names = VOCAB_FILES_NAMES
|
118 |
+
pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
|
119 |
+
max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
|
120 |
+
model_input_names = ["input_ids", "attention_mask"]
|
121 |
+
|
122 |
+
def __init__(
|
123 |
+
self,
|
124 |
+
vocab_file,
|
125 |
+
bos_token="<s>",
|
126 |
+
eos_token="</s>",
|
127 |
+
sep_token="</s>",
|
128 |
+
cls_token="<s>",
|
129 |
+
unk_token="<unk>",
|
130 |
+
pad_token="<pad>",
|
131 |
+
mask_token="<mask>",
|
132 |
+
num_patch_index_tokens=1024,
|
133 |
+
add_tag_and_patch_index_tokens=False,
|
134 |
+
sp_model_kwargs: Optional[Dict[str, Any]] = None,
|
135 |
+
**kwargs,
|
136 |
+
) -> None:
|
137 |
+
# Mask token behave like a normal word, i.e. include the space before it
|
138 |
+
mask_token = AddedToken(mask_token, lstrip=True, rstrip=False) if isinstance(mask_token, str) else mask_token
|
139 |
+
|
140 |
+
self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
|
141 |
+
|
142 |
+
super().__init__(
|
143 |
+
bos_token=bos_token,
|
144 |
+
eos_token=eos_token,
|
145 |
+
unk_token=unk_token,
|
146 |
+
sep_token=sep_token,
|
147 |
+
cls_token=cls_token,
|
148 |
+
pad_token=pad_token,
|
149 |
+
mask_token=mask_token,
|
150 |
+
sp_model_kwargs=self.sp_model_kwargs,
|
151 |
+
**kwargs,
|
152 |
+
)
|
153 |
+
|
154 |
+
self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
|
155 |
+
self.sp_model.Load(str(vocab_file))
|
156 |
+
self.vocab_file = vocab_file
|
157 |
+
|
158 |
+
# Original fairseq vocab and spm vocab must be "aligned":
|
159 |
+
# Vocab | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9
|
160 |
+
# -------- | ------- | ------- | ------ | ------- | ------ | ------ | ------ | ------ | ------- | ------
|
161 |
+
# fairseq | '<s>' | '<pad>' | '</s>' | '<unk>' | '.' | '_the' | ',' | '▁to' | '▁and' | '▁of'
|
162 |
+
# spm | '<unk>' | '<s>' | '</s>' | '.' | '_the' | ',' | '▁to' | '▁and' | '▁of' | '▁a'
|
163 |
+
|
164 |
+
# Mimic fairseq token-to-id alignment for the first 4 token
|
165 |
+
self.fairseq_tokens_to_ids = {"<s>": 0, "<pad>": 1, "</s>": 2, "<unk>": 3}
|
166 |
+
|
167 |
+
# The first "real" token "," has position 4 in the original fairseq vocab and position 3 in the spm vocab
|
168 |
+
self.fairseq_offset = 1
|
169 |
+
|
170 |
+
self.fairseq_tokens_to_ids["<mask>"] = len(self.sp_model) + self.fairseq_offset
|
171 |
+
self.fairseq_ids_to_tokens = {v: k for k, v in self.fairseq_tokens_to_ids.items()}
|
172 |
+
|
173 |
+
self.eod_token = "</doc>"
|
174 |
+
|
175 |
+
self.boi_token = "<image>"
|
176 |
+
self.eoi_token = "</image>"
|
177 |
+
|
178 |
+
self.eoc_token = "</chunk>"
|
179 |
+
self.eol_token = "</line>"
|
180 |
+
|
181 |
+
self.bop_token = "<phrase>"
|
182 |
+
self.eop_token = "</phrase>"
|
183 |
+
|
184 |
+
self.boo_token = "<object>"
|
185 |
+
self.eoo_token = "</object>"
|
186 |
+
|
187 |
+
self.dom_token = "</delimiter_of_multi_objects/>"
|
188 |
+
|
189 |
+
self.grd_token = "<grounding>"
|
190 |
+
|
191 |
+
self.tag_tokens = [
|
192 |
+
self.eod_token,
|
193 |
+
self.boi_token,
|
194 |
+
self.eoi_token,
|
195 |
+
self.eoc_token,
|
196 |
+
self.eol_token,
|
197 |
+
self.bop_token,
|
198 |
+
self.eop_token,
|
199 |
+
self.boo_token,
|
200 |
+
self.eoo_token,
|
201 |
+
self.dom_token,
|
202 |
+
self.grd_token,
|
203 |
+
]
|
204 |
+
|
205 |
+
self.num_patch_index_tokens = num_patch_index_tokens
|
206 |
+
patch_index_tokens = [f"<patch_index_{str(x).zfill(4)}>" for x in range(self.num_patch_index_tokens)]
|
207 |
+
|
208 |
+
if add_tag_and_patch_index_tokens:
|
209 |
+
for idx, token in enumerate(self.tag_tokens + patch_index_tokens):
|
210 |
+
# we can't add them as special tokens, as the slow tokenizer doesn't save the information of a token
|
211 |
+
# being special when it is added through `add_tokens`, but the fast tokenizer is able to do so.
|
212 |
+
self.add_tokens(AddedToken(token, lstrip=True, rstrip=False), special_tokens=True)
|
213 |
+
|
214 |
+
def _decode(
|
215 |
+
self,
|
216 |
+
token_ids: List[int],
|
217 |
+
skip_special_tokens: bool = False,
|
218 |
+
clean_up_tokenization_spaces: bool = None,
|
219 |
+
spaces_between_special_tokens: bool = True,
|
220 |
+
**kwargs,
|
221 |
+
) -> str:
|
222 |
+
self._decode_use_source_tokenizer = kwargs.pop("use_source_tokenizer", False)
|
223 |
+
|
224 |
+
filtered_tokens = self.convert_ids_to_tokens(token_ids, skip_special_tokens=skip_special_tokens)
|
225 |
+
|
226 |
+
# To avoid mixing byte-level and unicode for byte-level BPT
|
227 |
+
# we need to build string separately for added tokens and byte-level tokens
|
228 |
+
# cf. https://github.com/huggingface/transformers/issues/1133
|
229 |
+
sub_texts = []
|
230 |
+
current_sub_text = []
|
231 |
+
is_first_current_sub_text = True
|
232 |
+
for token in filtered_tokens:
|
233 |
+
if skip_special_tokens and token in self.all_special_ids:
|
234 |
+
continue
|
235 |
+
if token in self.added_tokens_encoder:
|
236 |
+
if current_sub_text:
|
237 |
+
sub_text = self.convert_tokens_to_string(current_sub_text)
|
238 |
+
# `convert_tokens_to_string` removes the leading space, which is undesired if we are not at the
|
239 |
+
# beginning part of the text. We can't use `spaces_between_special_tokens` to add this space back
|
240 |
+
# neither, as it will also add a space before a tag/patch_index token (which is not the case with
|
241 |
+
# the fast tokenizer - it doesn't even support `spaces_between_special_tokens`), which is not the
|
242 |
+
# ideal output format.
|
243 |
+
# The condition `not spaces_between_special_tokens` is to avoid double spaces.
|
244 |
+
if not is_first_current_sub_text and not spaces_between_special_tokens:
|
245 |
+
sub_text = " " + sub_text
|
246 |
+
sub_texts.append(sub_text)
|
247 |
+
current_sub_text = []
|
248 |
+
is_first_current_sub_text = False
|
249 |
+
sub_texts.append(token)
|
250 |
+
else:
|
251 |
+
current_sub_text.append(token)
|
252 |
+
if current_sub_text:
|
253 |
+
sub_texts.append(self.convert_tokens_to_string(current_sub_text))
|
254 |
+
|
255 |
+
if spaces_between_special_tokens:
|
256 |
+
text = " ".join(sub_texts)
|
257 |
+
else:
|
258 |
+
text = "".join(sub_texts)
|
259 |
+
|
260 |
+
clean_up_tokenization_spaces = (
|
261 |
+
clean_up_tokenization_spaces
|
262 |
+
if clean_up_tokenization_spaces is not None
|
263 |
+
else self.clean_up_tokenization_spaces
|
264 |
+
)
|
265 |
+
if clean_up_tokenization_spaces:
|
266 |
+
clean_text = self.clean_up_tokenization(text)
|
267 |
+
return clean_text
|
268 |
+
else:
|
269 |
+
return text
|
270 |
+
|
271 |
+
def __getstate__(self):
|
272 |
+
state = self.__dict__.copy()
|
273 |
+
state["sp_model"] = None
|
274 |
+
state["sp_model_proto"] = self.sp_model.serialized_model_proto()
|
275 |
+
return state
|
276 |
+
|
277 |
+
def __setstate__(self, d):
|
278 |
+
self.__dict__ = d
|
279 |
+
|
280 |
+
# for backward compatibility
|
281 |
+
if not hasattr(self, "sp_model_kwargs"):
|
282 |
+
self.sp_model_kwargs = {}
|
283 |
+
|
284 |
+
self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
|
285 |
+
self.sp_model.LoadFromSerializedProto(self.sp_model_proto)
|
286 |
+
|
287 |
+
def build_inputs_with_special_tokens(
|
288 |
+
self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
|
289 |
+
) -> List[int]:
|
290 |
+
"""
|
291 |
+
Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
|
292 |
+
adding special tokens. An XLM-RoBERTa sequence has the following format:
|
293 |
+
|
294 |
+
- single sequence: `<s> X </s>`
|
295 |
+
- pair of sequences: `<s> A </s></s> B </s>`
|
296 |
+
|
297 |
+
Args:
|
298 |
+
token_ids_0 (`List[int]`):
|
299 |
+
List of IDs to which the special tokens will be added.
|
300 |
+
token_ids_1 (`List[int]`, *optional*):
|
301 |
+
Optional second list of IDs for sequence pairs.
|
302 |
+
|
303 |
+
Returns:
|
304 |
+
`List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens.
|
305 |
+
"""
|
306 |
+
|
307 |
+
if token_ids_1 is None:
|
308 |
+
return [self.cls_token_id] + token_ids_0 + [self.sep_token_id]
|
309 |
+
cls = [self.cls_token_id]
|
310 |
+
sep = [self.sep_token_id]
|
311 |
+
return cls + token_ids_0 + sep + sep + token_ids_1 + sep
|
312 |
+
|
313 |
+
def get_special_tokens_mask(
|
314 |
+
self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
|
315 |
+
) -> List[int]:
|
316 |
+
"""
|
317 |
+
Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
|
318 |
+
special tokens using the tokenizer `prepare_for_model` method.
|
319 |
+
|
320 |
+
Args:
|
321 |
+
token_ids_0 (`List[int]`):
|
322 |
+
List of IDs.
|
323 |
+
token_ids_1 (`List[int]`, *optional*):
|
324 |
+
Optional second list of IDs for sequence pairs.
|
325 |
+
already_has_special_tokens (`bool`, *optional*, defaults to `False`):
|
326 |
+
Whether or not the token list is already formatted with special tokens for the model.
|
327 |
+
|
328 |
+
Returns:
|
329 |
+
`List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
|
330 |
+
"""
|
331 |
+
|
332 |
+
if already_has_special_tokens:
|
333 |
+
return super().get_special_tokens_mask(
|
334 |
+
token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
|
335 |
+
)
|
336 |
+
|
337 |
+
if token_ids_1 is None:
|
338 |
+
return [1] + ([0] * len(token_ids_0)) + [1]
|
339 |
+
return [1] + ([0] * len(token_ids_0)) + [1, 1] + ([0] * len(token_ids_1)) + [1]
|
340 |
+
|
341 |
+
def create_token_type_ids_from_sequences(
|
342 |
+
self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
|
343 |
+
) -> List[int]:
|
344 |
+
"""
|
345 |
+
Create a mask from the two sequences passed to be used in a sequence-pair classification task. XLM-RoBERTa does
|
346 |
+
not make use of token type ids, therefore a list of zeros is returned.
|
347 |
+
|
348 |
+
Args:
|
349 |
+
token_ids_0 (`List[int]`):
|
350 |
+
List of IDs.
|
351 |
+
token_ids_1 (`List[int]`, *optional*):
|
352 |
+
Optional second list of IDs for sequence pairs.
|
353 |
+
|
354 |
+
Returns:
|
355 |
+
`List[int]`: List of zeros.
|
356 |
+
|
357 |
+
"""
|
358 |
+
|
359 |
+
sep = [self.sep_token_id]
|
360 |
+
cls = [self.cls_token_id]
|
361 |
+
|
362 |
+
if token_ids_1 is None:
|
363 |
+
return len(cls + token_ids_0 + sep) * [0]
|
364 |
+
return len(cls + token_ids_0 + sep + sep + token_ids_1 + sep) * [0]
|
365 |
+
|
366 |
+
@property
|
367 |
+
def vocab_size(self):
|
368 |
+
return len(self.sp_model) + self.fairseq_offset + 1 # Add the <mask> token
|
369 |
+
|
370 |
+
def get_vocab(self):
|
371 |
+
vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}
|
372 |
+
vocab.update(self.added_tokens_encoder)
|
373 |
+
return vocab
|
374 |
+
|
375 |
+
def _tokenize(self, text: str) -> List[str]:
|
376 |
+
return self.sp_model.encode(text, out_type=str)
|
377 |
+
|
378 |
+
def _convert_token_to_id(self, token):
|
379 |
+
"""Converts a token (str) in an id using the vocab."""
|
380 |
+
if token in self.fairseq_tokens_to_ids:
|
381 |
+
return self.fairseq_tokens_to_ids[token]
|
382 |
+
spm_id = self.sp_model.PieceToId(token)
|
383 |
+
|
384 |
+
# Need to return unknown token if the SP model returned 0
|
385 |
+
return spm_id + self.fairseq_offset if spm_id else self.unk_token_id
|
386 |
+
|
387 |
+
def _convert_id_to_token(self, index):
|
388 |
+
"""Converts an index (integer) in a token (str) using the vocab."""
|
389 |
+
if index in self.fairseq_ids_to_tokens:
|
390 |
+
return self.fairseq_ids_to_tokens[index]
|
391 |
+
return self.sp_model.IdToPiece(index - self.fairseq_offset)
|
392 |
+
|
393 |
+
def convert_tokens_to_string(self, tokens):
|
394 |
+
"""Converts a sequence of tokens (strings for sub-words) in a single string."""
|
395 |
+
out_string = "".join(tokens).replace(SPIECE_UNDERLINE, " ").strip()
|
396 |
+
return out_string
|
397 |
+
|
398 |
+
def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
|
399 |
+
if not os.path.isdir(save_directory):
|
400 |
+
logger.error(f"Vocabulary path ({save_directory}) should be a directory")
|
401 |
+
return
|
402 |
+
out_vocab_file = os.path.join(
|
403 |
+
save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
|
404 |
+
)
|
405 |
+
|
406 |
+
if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file) and os.path.isfile(self.vocab_file):
|
407 |
+
copyfile(self.vocab_file, out_vocab_file)
|
408 |
+
elif not os.path.isfile(self.vocab_file):
|
409 |
+
with open(out_vocab_file, "wb") as fi:
|
410 |
+
content_spiece_model = self.sp_model.serialized_model_proto()
|
411 |
+
fi.write(content_spiece_model)
|
412 |
+
|
413 |
+
return (out_vocab_file,)
|
tokenization_kosmos2_fast.py
ADDED
@@ -0,0 +1,250 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# coding=utf-8
|
2 |
+
# Copyright 2023 Microsoft Research and The HuggingFace Inc. team. All rights reserved.
|
3 |
+
#
|
4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
5 |
+
# you may not use this file except in compliance with the License.
|
6 |
+
# You may obtain a copy of the License at
|
7 |
+
#
|
8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
9 |
+
#
|
10 |
+
# Unless required by applicable law or agreed to in writing, software
|
11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
+
# See the License for the specific language governing permissions and
|
14 |
+
# limitations under the License.
|
15 |
+
""" Tokenization classes for KOSMOS-2 model."""
|
16 |
+
|
17 |
+
|
18 |
+
import os
|
19 |
+
from shutil import copyfile
|
20 |
+
from typing import List, Optional, Tuple
|
21 |
+
|
22 |
+
from ...tokenization_utils import AddedToken
|
23 |
+
from ...tokenization_utils_fast import PreTrainedTokenizerFast
|
24 |
+
from ...utils import is_sentencepiece_available, logging
|
25 |
+
|
26 |
+
|
27 |
+
if is_sentencepiece_available():
|
28 |
+
from .tokenization_kosmos2 import Kosmos2Tokenizer
|
29 |
+
else:
|
30 |
+
Kosmos2TokenizerFast = None
|
31 |
+
|
32 |
+
|
33 |
+
logger = logging.get_logger(__name__)
|
34 |
+
|
35 |
+
VOCAB_FILES_NAMES = {"vocab_file": "sentencepiece.bpe.model", "tokenizer_file": "tokenizer.json"}
|
36 |
+
|
37 |
+
PRETRAINED_VOCAB_FILES_MAP = {
|
38 |
+
"vocab_file": {
|
39 |
+
"microsoft/kosmos-2-patch14-224": "https://huggingface.co/microsoft/kosmos-2-patch14-224/resolve/main/sentencepiece.bpe.model",
|
40 |
+
}
|
41 |
+
}
|
42 |
+
|
43 |
+
PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
|
44 |
+
"microsoft/kosmos-2-patch14-224": 2048,
|
45 |
+
}
|
46 |
+
|
47 |
+
|
48 |
+
class Kosmos2TokenizerFast(PreTrainedTokenizerFast):
|
49 |
+
"""
|
50 |
+
Construct a "fast" KOSMOS-2 tokenizer (backed by HuggingFace's *tokenizers* library). Adapted from
|
51 |
+
[`RobertaTokenizer`] and [`XLNetTokenizer`]. Based on
|
52 |
+
[BPE](https://huggingface.co/docs/tokenizers/python/latest/components.html?highlight=BPE#models).
|
53 |
+
|
54 |
+
This tokenizer inherits from [`PreTrainedTokenizerFast`] which contains most of the main methods. Users should
|
55 |
+
refer to this superclass for more information regarding those methods.
|
56 |
+
|
57 |
+
Args:
|
58 |
+
vocab_file (`str`):
|
59 |
+
Path to the vocabulary file.
|
60 |
+
bos_token (`str`, *optional*, defaults to `"<s>"`):
|
61 |
+
The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.
|
62 |
+
|
63 |
+
<Tip>
|
64 |
+
|
65 |
+
When building a sequence using special tokens, this is not the token that is used for the beginning of
|
66 |
+
sequence. The token used is the `cls_token`.
|
67 |
+
|
68 |
+
</Tip>
|
69 |
+
|
70 |
+
eos_token (`str`, *optional*, defaults to `"</s>"`):
|
71 |
+
The end of sequence token.
|
72 |
+
|
73 |
+
<Tip>
|
74 |
+
|
75 |
+
When building a sequence using special tokens, this is not the token that is used for the end of sequence.
|
76 |
+
The token used is the `sep_token`.
|
77 |
+
|
78 |
+
</Tip>
|
79 |
+
|
80 |
+
sep_token (`str`, *optional*, defaults to `"</s>"`):
|
81 |
+
The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for
|
82 |
+
sequence classification or for a text and a question for question answering. It is also used as the last
|
83 |
+
token of a sequence built with special tokens.
|
84 |
+
cls_token (`str`, *optional*, defaults to `"<s>"`):
|
85 |
+
The classifier token which is used when doing sequence classification (classification of the whole sequence
|
86 |
+
instead of per-token classification). It is the first token of the sequence when built with special tokens.
|
87 |
+
unk_token (`str`, *optional*, defaults to `"<unk>"`):
|
88 |
+
The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
|
89 |
+
token instead.
|
90 |
+
pad_token (`str`, *optional*, defaults to `"<pad>"`):
|
91 |
+
The token used for padding, for example when batching sequences of different lengths.
|
92 |
+
mask_token (`str`, *optional*, defaults to `"<mask>"`):
|
93 |
+
The token used for masking values. This is the token used when training this model with masked language
|
94 |
+
modeling. This is the token which the model will try to predict.
|
95 |
+
additional_special_tokens (`List[str]`, *optional*, defaults to `["<s>NOTUSED", "</s>NOTUSED"]`):
|
96 |
+
Additional special tokens used by the tokenizer.
|
97 |
+
num_patch_index_tokens (`int`, *optional*, defaults to `1024`):
|
98 |
+
The number of tokens used to specify the patch indices of bounding boxes in an image. These tokens have the
|
99 |
+
format `<patch_index_xxxx>` where `xxxx` is an integer.
|
100 |
+
"""
|
101 |
+
|
102 |
+
vocab_files_names = VOCAB_FILES_NAMES
|
103 |
+
pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
|
104 |
+
max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
|
105 |
+
model_input_names = ["input_ids", "attention_mask"]
|
106 |
+
slow_tokenizer_class = Kosmos2Tokenizer
|
107 |
+
|
108 |
+
def __init__(
|
109 |
+
self,
|
110 |
+
vocab_file=None,
|
111 |
+
tokenizer_file=None,
|
112 |
+
bos_token="<s>",
|
113 |
+
eos_token="</s>",
|
114 |
+
sep_token="</s>",
|
115 |
+
cls_token="<s>",
|
116 |
+
unk_token="<unk>",
|
117 |
+
pad_token="<pad>",
|
118 |
+
mask_token="<mask>",
|
119 |
+
num_patch_index_tokens=1024,
|
120 |
+
add_tag_and_patch_index_tokens=False,
|
121 |
+
**kwargs,
|
122 |
+
):
|
123 |
+
# Mask token behave like a normal word, i.e. include the space before it
|
124 |
+
mask_token = AddedToken(mask_token, lstrip=True, rstrip=False) if isinstance(mask_token, str) else mask_token
|
125 |
+
|
126 |
+
super().__init__(
|
127 |
+
vocab_file,
|
128 |
+
tokenizer_file=tokenizer_file,
|
129 |
+
bos_token=bos_token,
|
130 |
+
eos_token=eos_token,
|
131 |
+
sep_token=sep_token,
|
132 |
+
cls_token=cls_token,
|
133 |
+
unk_token=unk_token,
|
134 |
+
pad_token=pad_token,
|
135 |
+
mask_token=mask_token,
|
136 |
+
**kwargs,
|
137 |
+
)
|
138 |
+
|
139 |
+
self.vocab_file = vocab_file
|
140 |
+
self.can_save_slow_tokenizer = False if not self.vocab_file else True
|
141 |
+
|
142 |
+
self.eod_token = "</doc>"
|
143 |
+
|
144 |
+
self.boi_token = "<image>"
|
145 |
+
self.eoi_token = "</image>"
|
146 |
+
|
147 |
+
self.eoc_token = "</chunk>"
|
148 |
+
self.eol_token = "</line>"
|
149 |
+
|
150 |
+
self.bop_token = "<phrase>"
|
151 |
+
self.eop_token = "</phrase>"
|
152 |
+
|
153 |
+
self.boo_token = "<object>"
|
154 |
+
self.eoo_token = "</object>"
|
155 |
+
|
156 |
+
self.dom_token = "</delimiter_of_multi_objects/>"
|
157 |
+
|
158 |
+
self.grd_token = "<grounding>"
|
159 |
+
|
160 |
+
self.tag_tokens = [
|
161 |
+
self.eod_token,
|
162 |
+
self.boi_token,
|
163 |
+
self.eoi_token,
|
164 |
+
self.eoc_token,
|
165 |
+
self.eol_token,
|
166 |
+
self.bop_token,
|
167 |
+
self.eop_token,
|
168 |
+
self.boo_token,
|
169 |
+
self.eoo_token,
|
170 |
+
self.dom_token,
|
171 |
+
self.grd_token,
|
172 |
+
]
|
173 |
+
|
174 |
+
self.num_patch_index_tokens = num_patch_index_tokens
|
175 |
+
patch_index_tokens = [f"<patch_index_{str(x).zfill(4)}>" for x in range(self.num_patch_index_tokens)]
|
176 |
+
|
177 |
+
if add_tag_and_patch_index_tokens:
|
178 |
+
for idx, token in enumerate(self.tag_tokens + patch_index_tokens):
|
179 |
+
# we need to set `special_tokens=False` to be the same as in the slow tokenizer.
|
180 |
+
self.add_tokens(AddedToken(token, lstrip=True, rstrip=False), special_tokens=False)
|
181 |
+
|
182 |
+
def build_inputs_with_special_tokens(
|
183 |
+
self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
|
184 |
+
) -> List[int]:
|
185 |
+
"""
|
186 |
+
Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
|
187 |
+
adding special tokens. An XLM-RoBERTa sequence has the following format:
|
188 |
+
|
189 |
+
- single sequence: `<s> X </s>`
|
190 |
+
- pair of sequences: `<s> A </s></s> B </s>`
|
191 |
+
|
192 |
+
Args:
|
193 |
+
token_ids_0 (`List[int]`):
|
194 |
+
List of IDs to which the special tokens will be added.
|
195 |
+
token_ids_1 (`List[int]`, *optional*):
|
196 |
+
Optional second list of IDs for sequence pairs.
|
197 |
+
|
198 |
+
Returns:
|
199 |
+
`List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens.
|
200 |
+
"""
|
201 |
+
|
202 |
+
if token_ids_1 is None:
|
203 |
+
return [self.cls_token_id] + token_ids_0 + [self.sep_token_id]
|
204 |
+
cls = [self.cls_token_id]
|
205 |
+
sep = [self.sep_token_id]
|
206 |
+
return cls + token_ids_0 + sep + sep + token_ids_1 + sep
|
207 |
+
|
208 |
+
def create_token_type_ids_from_sequences(
|
209 |
+
self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
|
210 |
+
) -> List[int]:
|
211 |
+
"""
|
212 |
+
Create a mask from the two sequences passed to be used in a sequence-pair classification task. XLM-RoBERTa does
|
213 |
+
not make use of token type ids, therefore a list of zeros is returned.
|
214 |
+
|
215 |
+
Args:
|
216 |
+
token_ids_0 (`List[int]`):
|
217 |
+
List of IDs.
|
218 |
+
token_ids_1 (`List[int]`, *optional*):
|
219 |
+
Optional second list of IDs for sequence pairs.
|
220 |
+
|
221 |
+
Returns:
|
222 |
+
`List[int]`: List of zeros.
|
223 |
+
|
224 |
+
"""
|
225 |
+
|
226 |
+
sep = [self.sep_token_id]
|
227 |
+
cls = [self.cls_token_id]
|
228 |
+
|
229 |
+
if token_ids_1 is None:
|
230 |
+
return len(cls + token_ids_0 + sep) * [0]
|
231 |
+
return len(cls + token_ids_0 + sep + sep + token_ids_1 + sep) * [0]
|
232 |
+
|
233 |
+
def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
|
234 |
+
if not self.can_save_slow_tokenizer:
|
235 |
+
raise ValueError(
|
236 |
+
"Your fast tokenizer does not have the necessary information to save the vocabulary for a slow "
|
237 |
+
"tokenizer."
|
238 |
+
)
|
239 |
+
|
240 |
+
if not os.path.isdir(save_directory):
|
241 |
+
logger.error(f"Vocabulary path ({save_directory}) should be a directory.")
|
242 |
+
return
|
243 |
+
out_vocab_file = os.path.join(
|
244 |
+
save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
|
245 |
+
)
|
246 |
+
|
247 |
+
if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file):
|
248 |
+
copyfile(self.vocab_file, out_vocab_file)
|
249 |
+
|
250 |
+
return (out_vocab_file,)
|