File size: 6,305 Bytes
92740f3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
# Copyright (c) 2024 NVIDIA CORPORATION. 
#   Licensed under the MIT license.

# Adapted from https://github.com/mlfoundations/open_flamingo under the MIT license.
#   LICENSE is in incl_licenses directory.

import torch.nn as nn

try:
    from .helpers import GatedCrossAttentionBlock
    from .utils import getattr_recursive, setattr_recursive
except:
    from helpers import GatedCrossAttentionBlock
    from utils import getattr_recursive, setattr_recursive


class FlamingoLayer(nn.Module):
    """
    FlamingoLayer is a wrapper around the GatedCrossAttentionBlock and DecoderLayer.
    """

    def __init__(
        self, gated_cross_attn_layer, decoder_layer, gradient_checkpointing=False
    ):
        super().__init__()
        self.gated_cross_attn_layer = gated_cross_attn_layer
        self.decoder_layer = decoder_layer
        self.audio_x = None
        self.audio_x_mask = None
        self.few_shot_mask = None
        self.media_locations = None
        if self.gated_cross_attn_layer is not None:
            self.gated_cross_attn_layer._use_gradient_checkpointing = (
                gradient_checkpointing
            )
        self.decoder_layer._use_gradient_checkpointing = gradient_checkpointing

    def is_conditioned(self) -> bool:
        """Check whether the layer is conditioned."""
        return (self.audio_x is not None) and (self.audio_x_mask is not None) and (self.media_locations is not None)

    def condition_audio_x(self, audio_x, audio_x_mask):
        self.audio_x = audio_x
        self.audio_x_mask = audio_x_mask

    def condition_media_locations(self, media_locations):
        self.media_locations = media_locations

    def condition_use_cached_media(self, use_cached_media):
        self.use_cached_media = use_cached_media

    def forward(
        self,
        lang_x,
        attention_mask=None,
        **decoder_layer_kwargs,
    ):
        if self.gated_cross_attn_layer is not None:
            if self.audio_x is None:
                raise ValueError("audio_x must be conditioned before forward pass")

            if self.media_locations is None:
                raise ValueError(
                    "media_locations must be conditioned before forward pass"
                )

            lang_x = self.gated_cross_attn_layer(
                lang_x,
                self.audio_x,
                self.audio_x_mask,
                media_locations=self.media_locations,
                use_cached_media=self.use_cached_media,
            )

        # Normal decoder layer
        lang_x = self.decoder_layer(
            lang_x, attention_mask=attention_mask, **decoder_layer_kwargs
        )
        return lang_x


class FlamingoLMMixin(nn.Module):
    """
    Mixin to add cross-attention layers to a language model.
    """

    def set_decoder_layers_attr_name(self, decoder_layers_attr_name):
        self.decoder_layers_attr_name = decoder_layers_attr_name

    def _get_decoder_layers(self):
        return getattr_recursive(self, self.decoder_layers_attr_name)

    def _set_decoder_layers(self, value):
        setattr_recursive(self, self.decoder_layers_attr_name, value)

    def init_flamingo(
        self,
        media_token_id,
        lang_hidden_size,
        audio_hidden_size,
        max_window_per_audio,
        cross_attn_every_n_layers,
        gradient_checkpointing,
    ):
        """
        Initialize Flamingo by adding a new gated cross attn to the decoder. Store the media token id for computing the media locations.
        """
        self.old_decoder_blocks = self._get_decoder_layers()
        self.gated_cross_attn_layers = nn.ModuleList(
            [
                GatedCrossAttentionBlock(
                    dim=lang_hidden_size, 
                    dim_audio=audio_hidden_size,
                    max_window_per_audio=max_window_per_audio, 
                    only_attend_immediate_media=False,
                )
                if (layer_idx + 1) % cross_attn_every_n_layers == 0
                else None
                for layer_idx, _ in enumerate(self._get_decoder_layers())
            ]
        )
        self.init_flamingo_layers(gradient_checkpointing)
        self.media_token_id = media_token_id
        self.initialized_flamingo = True
        self._use_cached_audio_x = False

    def init_flamingo_layers(self, gradient_checkpointing):
        """
        Re initializes the FlamingoLayers.
        Propagates any changes made to self.gated_corss_attn_layers or self.old_decoder_blocks
        """
        self._set_decoder_layers(
            nn.ModuleList(
                [
                    FlamingoLayer(
                        gated_cross_attn_layer, decoder_layer, gradient_checkpointing
                    )
                    for gated_cross_attn_layer, decoder_layer in zip(
                        self.gated_cross_attn_layers, self.old_decoder_blocks
                    )
                ]
            )
        )

    def forward(self, input_ids, attention_mask, **kwargs):
        """Condition the Flamingo layers on the media locations before forward()"""
        if not self.initialized_flamingo:
            raise ValueError(
                "Flamingo layers are not initialized. Please call `init_flamingo` first."
            )

        media_locations = input_ids == self.media_token_id

        use_cached_media_locations = (
            self._use_cached_audio_x
            and self.is_conditioned()
            and not media_locations.any()
        )

        for layer in self._get_decoder_layers():
            if not use_cached_media_locations:
                layer.condition_media_locations(media_locations)
            layer.condition_use_cached_media(use_cached_media_locations)

        kwargs["input_ids"] = input_ids
        kwargs["attention_mask"] = attention_mask
        return super().forward(**kwargs)

    def is_conditioned(self) -> bool:
        """Check whether all decoder layers are already conditioned."""
        return all(l.is_conditioned() for l in self._get_decoder_layers())

    def clear_conditioned_layers(self):
        for layer in self._get_decoder_layers():
            layer.condition_audio_x(None, None)
            layer.condition_media_locations(None)
            layer.condition_use_cached_media(None)