Support returning hidden state from CLIPTextEncoder

tests/models/clip/test_text_encoder.py

@@ -17,7 +17,7 @@ class TestCLIPTextEncoder:
  def start(self):
  set_rng_seed(1234)
  context_length = 77
- batch_size, embedding_dim, heads = 2, 4, 2
+ batch_size, embedding_dim, heads, width = 2, 4, 2, 512
 
  def build_text(text_length):
  return torch.randint(1, 10, (batch_size, text_length), dtype=torch.long)
@@ -27,12 +27,16 @@ def build_encoder(
  use_clip_init=True,
  context_length=context_length,
  heads=heads,
+ width=width,
+ return_hidden_state=False,
  ):
  return CLIPTextEncoder(
  embedding_dim=embedding_dim,
  use_clip_init=use_clip_init,
  context_length=context_length,
+ width=width,
  heads=heads,
+ return_hidden_state=return_hidden_state,
  )
 
  return build_encoder, build_text
@@ -117,6 +121,47 @@ def test_forward(self, start):
  actual=actual_clip_init, expected=expected_clip_init, rtol=0, atol=1e-4
  )
 
+ def test_forward_return_hidden_state(self, start):
+ build_encoder, build_text = start
+ text = build_text(text_length=3)
+
+ text_encoder = build_encoder(
+ context_length=3, width=4, return_hidden_state=True
+ )
+ assert isinstance(text_encoder, torch.nn.Module)
+
+ actual_clip_init, actual_hidden_state = text_encoder(text)
+ print(actual_hidden_state)
+ expected_clip_init = torch.Tensor(
+ [
+ [-0.366838, -1.596611, -0.330413, -0.593790],
+ [-0.790419, 0.876780, -0.970667, -0.727134],
+ ]
+ )
+ expected_hidden_state = torch.Tensor(
+ [
+ [
+ [6.348165e-01, -4.137459e-02, -1.604239e00, 1.010798e00],
+ [6.204837e-01, -3.028658e-02, -1.606570e00, 1.016373e00],
+ [5.915626e-01, -1.666874e-03, -1.613292e00, 1.023396e00],
+ ],
+ [
+ [5.910631e-01, -1.515219e-02, -1.607913e00, 1.032002e00],
+ [1.467783e-01, -1.675803e00, 7.402021e-01, 7.888227e-01],
+ [6.721084e-01, -2.896671e-01, -1.493379e00, 1.110938e00],
+ ],
+ ]
+ )
+ assert_expected(
+ actual=actual_clip_init, expected=expected_clip_init, rtol=0, atol=1e-4
+ )
+ assert_expected(
+ actual=actual_hidden_state,
+ expected=expected_hidden_state,
+ rtol=0,
+ atol=1e-4,
+ )
+
  def test_forward_over_context_length(self, start):
  build_encoder, build_text = start
 

torchmultimodal/models/clip/text_encoder.py

@@ -5,6 +5,8 @@
 # LICENSE file in the root directory of this source tree.
 
 
+from typing import Tuple, Union
+
 import torch
 from torch import nn, Tensor
 
@@ -28,6 +30,8 @@ class CLIPTextEncoder(nn.Module):
  heads (int): Number of heads in Transformer encoder.
  layers (int): Number of layers in Transformer encoder.
  use_clip_init (bool): Whether to use CLIP-specific initialization.
+ return_hidden_state (bool): Whether to return the last hidden state.
+ If True, forward returns a tuple of final embedding and last hidden state.
 
  Inputs: text (Tensor): Tensor containing text features.
  """
@@ -45,6 +49,7 @@ def __init__(
  heads: int = 8,
  layers: int = 12,
  use_clip_init: bool = True,
+ return_hidden_state: bool = False,
  ):
  super().__init__()
  torch._C._log_api_usage_once(f"torchmultimodal.{self.__class__.__name__}")
@@ -77,6 +82,8 @@ def __init__(
  if use_clip_init:
  self.initialize_parameters()
 
+ self.return_hidden_state = return_hidden_state
+
  def initialize_parameters(self) -> None:
  # Initialize token and positional embeddings
  nn.init.normal_(self.token_embedding.weight, std=self.TOKEN_EMBEDDING_INIT_STD)
@@ -108,7 +115,7 @@ def build_attention_mask(self) -> Tensor:
  ).triu(1)
  return mask
 
- def forward(self, text: Tensor) -> Tensor:
+ def forward(self, text: Tensor) -> Union[Tensor, Tuple[Tensor, Tensor]]:
  if text.size(1) != self.context_length:
  raise ValueError(
  f"length of input should be {self.context_length} but found {text.size(1)}"
@@ -120,11 +127,13 @@ def forward(self, text: Tensor) -> Tensor:
 
  # [n_ctx, bs, transformer.width] -> [bs, n_ctx, transformer.width]
  embeddings = torch.permute(embeddings, (1, 0, 2))
- embeddings = self.ln_final(embeddings)
+ hidden_state = self.ln_final(embeddings)
  # take features from the eot embedding (the highest number in each sequence)
  embeddings = self.projection(
- embeddings[torch.arange(embeddings.shape[0]), text.argmax(dim=-1)]
+ hidden_state[torch.arange(hidden_state.shape[0]), text.argmax(dim=-1)]
  )
  # embeddings now has size [bs, embedding_dim]
 
+ if self.return_hidden_state:
+ return embeddings, hidden_state
  return embeddings