Upload model

config.json

@@ -0,0 +1,48 @@
+{
+  "activation_function": "gelu_new",
+  "architectures": [
+    "BackpackGPT2NLIModel"
+  ],
+  "attn_pdrop": 0.1,
+  "auto_map": {
+    "AutoConfig": "configuration_backpack_gpt2_nli.BackpackGPT2NLIConfig",
+    "AutoModelForSequenceClassification": "modeling_backpack_gpt2_nli.BackpackGPT2NLIModel"
+  },
+  "bos_token_id": 50256,
+  "embd_pdrop": 0.1,
+  "eos_token_id": 50256,
+  "freeze_backpack": true,
+  "id2label": {
+    "0": "LABEL_0",
+    "1": "LABEL_1",
+    "2": "LABEL_2"
+  },
+  "initializer_range": 0.02,
+  "label2id": {
+    "LABEL_0": 0,
+    "LABEL_1": 1,
+    "LABEL_2": 2
+  },
+  "layer_norm_epsilon": 1e-05,
+  "model_type": "gpt2",
+  "n_embd": 768,
+  "n_head": 12,
+  "n_inner": null,
+  "n_layer": 12,
+  "n_positions": 512,
+  "num_senses": 16,
+  "reorder_and_upcast_attn": false,
+  "resid_pdrop": 0.1,
+  "scale_attn_by_inverse_layer_idx": true,
+  "scale_attn_weights": true,
+  "sense_intermediate_scale": 4,
+  "summary_activation": null,
+  "summary_first_dropout": 0.1,
+  "summary_proj_to_labels": true,
+  "summary_type": "cls_index",
+  "summary_use_proj": true,
+  "torch_dtype": "float32",
+  "transformers_version": "4.31.0",
+  "use_cache": true,
+  "vocab_size": 50264
+}

configuration_backpack_gpt2.py

@@ -0,0 +1,42 @@
+from transformers.models.gpt2.configuration_gpt2 import GPT2Config
+
+class BackpackGPT2Config(GPT2Config):
+  """
+    This is the configuration class to store the configuration of a [`GPT2Model`] or a [`TFGPT2Model`]. It is used to
+    instantiate a Backpack GPT-2 model according to the specified arguments, defining the model architecture.
+
+    Configuration objects inherit from [`GPT2Config`] and can be used to control the model outputs. Read the
+    documentation from [`GPT2Config`] for more information.
+
+    Args:
+        num_senses (`int`, *optional*, defaults to 16):
+            The number of sense vectors to define for each word.
+        sense_intermediate_scale (`int`, *optional*, defaults ot 4):
+            The hidden dimensionality of the sense vector network.
+
+    Example:
+
+    ```python
+    >>> from transformers import BackpackGPT2Config, BackpackGPT2Model
+
+    >>> # Initializing a GPT2 configuration
+    >>> configuration = BackpackGPT2Config()
+
+    >>> # Initializing a model (with random weights) from the configuration
+    >>> model = BackpackGPT2Model(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+  """
+
+  def __init__(self,
+               vocab_size=50264,
+               num_senses=16,
+               sense_intermediate_scale=4,
+               n_positions=512,
+               scale_attn_by_inverse_layer_idx=True,
+               **kwargs,
+  ):
+    self.num_senses = num_senses
+    self.sense_intermediate_scale = sense_intermediate_scale
+    super().__init__(vocab_size=vocab_size, n_positions=n_positions, scale_attn_by_inverse_layer_idx=scale_attn_by_inverse_layer_idx, **kwargs)

configuration_backpack_gpt2_nli.py

@@ -0,0 +1,19 @@
+from transformers.models.gpt2.configuration_gpt2 import GPT2Config
+
+class BackpackGPT2NLIConfig(GPT2Config):
+  def __init__(self,
+               freeze_backpack=True,
+               num_labels=3,
+               vocab_size=50264,
+               num_senses=16,
+               sense_intermediate_scale=4,
+               n_positions=512,
+               scale_attn_by_inverse_layer_idx=True,
+               **kwargs,
+  ):
+    self.freeze_backpack = freeze_backpack
+    #self.num_labels = num_labels
+
+    self.num_senses = num_senses
+    self.sense_intermediate_scale = sense_intermediate_scale
+    super().__init__(vocab_size=vocab_size, n_positions=n_positions, scale_attn_by_inverse_layer_idx=scale_attn_by_inverse_layer_idx, num_labels=num_labels, **kwargs)

modeling_backpack_gpt2.py

@@ -0,0 +1,229 @@
+import math
+from dataclasses import dataclass
+from typing import Optional, Tuple
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+
+from transformers.activations import ACT2FN
+from transformers.pytorch_utils import Conv1D
+from transformers.utils import (
+    ModelOutput,
+    logging,
+)
+from transformers.models.gpt2.modeling_gpt2 import GPT2Model, GPT2PreTrainedModel
+from .configuration_backpack_gpt2 import BackpackGPT2Config # From the same directory
+
+logger = logging.get_logger(__name__)
+
+
+### Backpack-Specific
+class BackpackGPT2PreTrainedModel(GPT2PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+    _keys_to_ignore_on_load_missing = [r"attn.masked_bias", r"attn.bias"]
+
+    config_class = BackpackGPT2Config
+    base_model_prefix = "backpack"
+    is_parallelizable = True
+    supports_gradient_checkpointing = False
+    _no_split_modules = ["GPT2Block", "BackpackNoMixBlock"]
+
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+class BackpackMLP(nn.Module):
+
+  def __init__(self, embed_dim, intermediate_dim, out_dim, config):
+        super().__init__()
+        self.c_fc = Conv1D(intermediate_dim, embed_dim)
+        self.c_proj = Conv1D(out_dim, intermediate_dim)
+        self.act = ACT2FN[config.activation_function]
+        self.dropout = nn.Dropout(config.resid_pdrop)
+
+  def forward(self, hidden_states: Optional[Tuple[torch.FloatTensor]]) -> torch.FloatTensor:
+      hidden_states = self.c_fc(hidden_states) # run order: 5 
+      hidden_states = self.act(hidden_states)
+      hidden_states = self.c_proj(hidden_states)
+      hidden_states = self.dropout(hidden_states)
+      return hidden_states
+
+class BackpackNoMixBlock(nn.Module):
+
+  def __init__(self, config):
+    super().__init__()
+    self.ln_1 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+    self.ln_2 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+    self.mlp = BackpackMLP(config.n_embd, config.n_embd*4, config.n_embd, config)
+    self.resid_dropout1 = nn.Dropout(config.resid_pdrop)
+    self.resid_dropout2 = nn.Dropout(config.resid_pdrop)
+
+  def forward(self, hidden_states, residual):
+    residual = self.resid_dropout1(hidden_states) + residual # run order: 4
+    hidden_states = self.ln_1(residual)
+    mlp_out = self.mlp(hidden_states)
+    residual = self.resid_dropout2(mlp_out) + residual
+    hidden_states = self.ln_2(residual)
+    return hidden_states
+
+class BackpackSenseNetwork(nn.Module):
+    def __init__(self, config, num_senses, device=None, dtype=None):
+        super().__init__()
+        self.num_senses = num_senses # paper: k
+        #self.embeddings = embeddings
+        self.n_embd = config.n_embd # paper: d
+
+        self.dropout = nn.Dropout(config.embd_pdrop)
+        self.block = BackpackNoMixBlock(config)
+        self.ln = nn.LayerNorm(self.n_embd, eps=config.layer_norm_epsilon)
+        self.final_mlp = BackpackMLP(
+            embed_dim=config.n_embd,
+            intermediate_dim=config.sense_intermediate_scale*config.n_embd,
+            out_dim=config.n_embd*config.num_senses,
+            config=config,
+            )
+
+    def forward(self, input_embeds):
+      residual = self.dropout(input_embeds) # run order: 3
+      hidden_states = self.ln(residual)
+      hidden_states = self.block(hidden_states, residual)
+      senses = self.final_mlp(hidden_states)
+      bs, s, nvd = senses.shape
+      return senses.reshape(bs, s, self.num_senses, self.n_embd).transpose(1,2) # (bs, nv, s, d)
+
+class BackpackWeightNetwork(nn.Module):
+
+  def __init__(self, num_senses, embed_dim):
+    super().__init__()
+    self.n_embd = embed_dim
+    self.num_senses = num_senses
+    self.embed_per_sense = embed_dim // num_senses
+    self.c_attn = nn.Linear(embed_dim, 2 * num_senses * self.embed_per_sense)
+    self.softmax_scale = None
+
+  def forward(self, encoded):
+    b, s, d = encoded.shape
+    encoded = self.c_attn(encoded) # (b, s, 2*d)
+    encoded = encoded.reshape(b, s, 2, self.num_senses, self.embed_per_sense) #(b, s, 2, nv, d//nv)
+    batch_size, seqlen = encoded.shape[0], encoded.shape[1]
+
+    # compute scores & mask
+    q, k = encoded.unbind(dim=2)
+    softmax_scale = self.softmax_scale or 1.0 / math.sqrt(q.shape[-1])
+    scores = torch.einsum('bthd,bshd->bhts', q, k * softmax_scale)
+    causal_mask = torch.triu(torch.full((seqlen, seqlen), -10000.0, device=scores.device), 1)
+    scores = scores + causal_mask.to(dtype=scores.dtype)
+
+    return torch.softmax(scores, dim=-1, dtype=q.dtype)
+  
+
+@dataclass
+class BackpackGPT2BaseModelOutput(ModelOutput):
+    hidden_states: torch.FloatTensor = None
+    contextualization: torch.FloatTensor = None
+
+class BackpackGPT2Model(BackpackGPT2PreTrainedModel):
+    _keys_to_ignore_on_load_missing = [r".*attn.masked_bias", r".*attn.bias"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.embed_dim = config.n_embd # paper: d
+        self.num_senses = config.num_senses # paper: k
+
+        self.gpt2_model = GPT2Model(config) # paper: parameterizing contextualization weights: Transformer Model
+        
+        self.sense_network = BackpackSenseNetwork(config, self.num_senses, self.gpt2_model.wte) # paper: parametrizing senses
+        self.word_embeddings = self.gpt2_model.wte
+        self.position_embeddings = self.gpt2_model.wpe
+        self.sense_weight_net = BackpackWeightNetwork(self.num_senses, self.embed_dim) # paper: parameterizing contextualization weights
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+        self.gradient_checkpointing = False
+
+    def get_num_senses(self):
+        return self.num_senses
+
+    def get_word_embeddings(self):
+        return self.word_embeddings
+
+    def get_sense_network(self):
+        return self.sense_network
+
+    def forward(self, input_ids, position_ids):
+        # Compute senses
+        sense_input_embeds = self.word_embeddings(input_ids) # GPT2 word emb without pos # run order: 2
+        senses = self.sense_network(sense_input_embeds) # (bs, nv, s, d) (batch size, num toks, num senses, sense dim)
+
+        # Compute contextualization weights
+        contextl_hidden_states = self.gpt2_model(input_ids, position_ids=position_ids).last_hidden_state # (bs, s, d)
+        contextualization = self.sense_weight_net(contextl_hidden_states) # (bs, nv, s, s)
+
+        # Compute resulting outputs
+        hidden_states = torch.sum(contextualization @ senses, dim=1) # (bs, nv, s, d) -> (bs, s, d)
+        return BackpackGPT2BaseModelOutput(
+            hidden_states=hidden_states, # paper: o
+            contextualization=contextualization,
+        )
+    
+    def run_with_custom_contextualization(self, input_ids, contextualization):
+        # Compute senses
+        sense_input_embeds = self.word_embeddings(input_ids)
+        senses = self.sense_network(sense_input_embeds) # (bs, nv, s, d)
+
+        # Compute resulting outputs
+        hidden_states = torch.sum(contextualization @ senses, dim=1) # (bs, nv, s, d) -> (bs, s, d)
+        return BackpackGPT2BaseModelOutput(
+            hidden_states=hidden_states,
+            contextualization=contextualization,
+        )
+
+@dataclass
+class BackpackGPT2LMHeadModelOutput(ModelOutput):
+    logits: torch.FloatTensor = None
+    contextualization: torch.FloatTensor = None
+
+class BackpackGPT2LMHeadModel(BackpackGPT2PreTrainedModel):
+  _keys_to_ignore_on_load_missing = [r".*attn.masked_bias", r".*attn.bias"]
+
+  def __init__(self, config):
+    super().__init__(config)
+    self.backpack = BackpackGPT2Model(config)
+    self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+
+    # Model parallel
+    self.model_parallel = False
+    self.device_map = None
+
+    self.tie_weights()
+
+  def tie_weights(self):
+      self.lm_head.weight = self.backpack.word_embeddings.weight # also tied with the underlying underlying transf
+
+  def get_lm_head(self):
+      return self.lm_head
+
+  def forward(self, input_ids, position_ids=None): # run order: 1
+      outputs = self.backpack(input_ids, position_ids=position_ids)
+      hidden_states, contextualization = outputs.hidden_states, outputs.contextualization
+      lm_logits = self.lm_head(hidden_states) # (bs, s, V)
+      return BackpackGPT2LMHeadModelOutput(
+            logits=lm_logits,
+            contextualization=contextualization,
+        ) # you should return hidden_states as contextualized emb of input words in the specified sequence
+            # use these contextualized embeddings as input to classification head, actually you should write a new class similar to this one but
+            # use different output size for lm_head param (call it clf_head)
+
+  def run_with_custom_contextualization(self, input_ids, contextualization):
+      outputs = self.backpack.run_with_custom_contextualization(input_ids, contextualization)
+      hidden_states, contextualization = outputs.hidden_states, outputs.contextualization
+      lm_logits = self.lm_head(hidden_states)
+      return BackpackGPT2LMHeadModelOutput(
+        logits=lm_logits,
+        contextualization=contextualization,
+    )
+            

modeling_backpack_gpt2_nli.py

@@ -0,0 +1,55 @@
+import torch
+import torch.utils.checkpoint
+from torch import nn
+
+from transformers.models.gpt2.modeling_gpt2 import GPT2PreTrainedModel
+from .configuration_backpack_gpt2_nli import BackpackGPT2NLIConfig
+from .modeling_backpack_gpt2 import BackpackGPT2Model
+
+
+class BackpackGPT2NLIModel(GPT2PreTrainedModel):
+    config_class = BackpackGPT2NLIConfig
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.backpack = BackpackGPT2Model(config)
+
+        self.n_embd = config.n_embd
+
+        self.num_labels = config.num_labels # 0: Entailment -- 1: Neutral -- 2: Contradiction
+        
+        self.nli_head = nn.Sequential(
+            nn.Linear(self.n_embd, self.n_embd),
+            nn.Dropout(0.1),
+            nn.Linear(self.n_embd, self.num_labels)
+        )
+
+        # Freeze The Encoder if Needed
+        self.backpack.requires_grad_(not config.freeze_backpack)
+
+        self.loss_func = nn.CrossEntropyLoss()
+
+        # Model parallel
+        self.model_parallel = False
+
+
+    def forward(self, input_ids=None, attention_mask=None, labels=None):
+
+        backpack_outputs = self.backpack(input_ids=input_ids, position_ids=None)
+
+        backpack_hidden_states, backpack_contextualization = backpack_outputs.hidden_states, backpack_outputs.contextualization
+        last_toks_indices = attention_mask.shape[1] - 1 - attention_mask.flip((1,)).argmax(dim=1) # index of the last token of the input (according to att mask)
+        last_backpack_hidden_states = backpack_hidden_states[torch.arange(backpack_hidden_states.shape[0]), last_toks_indices, :]
+
+        logits = self.nli_head(last_backpack_hidden_states)
+
+        if labels is not None:
+            # Flatten the logits and labels, considering the attention mask
+            flat_logits = logits
+            flat_labels = labels.view(-1)
+
+            loss = self.loss_func(flat_logits, flat_labels)
+            return {'logits': logits, 'loss': loss}
+        else:
+            return {'logits': logits}

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7bc9bbb9567385a0e5d334d1c2391ac80b77becda3be7e947aa9d50edc607306
+size 682759533