Make CLIP model could use new added tokens with meaningful pooling (#24777)

* fix

---------

Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

src/transformers/models/clip/modeling_clip.py

@@ -701,6 +701,9 @@ class CLIPTextTransformer(nn.Module):
         self.encoder = CLIPEncoder(config)
         self.final_layer_norm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
 
+        # For `pooled_output` computation
+        self.eos_token_id = config.eos_token_id
+
     @add_start_docstrings_to_model_forward(CLIP_TEXT_INPUTS_DOCSTRING)
     @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=CLIPTextConfig)
     def forward(
@@ -750,13 +753,26 @@ class CLIPTextTransformer(nn.Module):
         last_hidden_state = encoder_outputs[0]
         last_hidden_state = self.final_layer_norm(last_hidden_state)
 
-        # text_embeds.shape = [batch_size, sequence_length, transformer.width]
-        # take features from the eot embedding (eot_token is the highest number in each sequence)
-        # casting to torch.int for onnx compatibility: argmax doesn't support int64 inputs with opset 14
-        pooled_output = last_hidden_state[
-            torch.arange(last_hidden_state.shape[0], device=last_hidden_state.device),
-            input_ids.to(dtype=torch.int, device=last_hidden_state.device).argmax(dim=-1),
-        ]
+        if self.eos_token_id == 2:
+            # The `eos_token_id` was incorrect before PR #24773: Let's keep what have been done here.
+            # A CLIP model with such `eos_token_id` in the config can't work correctly with extra new tokens added
+            # ------------------------------------------------------------
+            # text_embeds.shape = [batch_size, sequence_length, transformer.width]
+            # take features from the eot embedding (eot_token is the highest number in each sequence)
+            # casting to torch.int for onnx compatibility: argmax doesn't support int64 inputs with opset 14
+            pooled_output = last_hidden_state[
+                torch.arange(last_hidden_state.shape[0], device=last_hidden_state.device),
+                input_ids.to(dtype=torch.int, device=last_hidden_state.device).argmax(dim=-1),
+            ]
+        else:
+            # The config gets updated `eos_token_id` from PR #24773 (so the use of exta new tokens is possible)
+            pooled_output = last_hidden_state[
+                torch.arange(last_hidden_state.shape[0], device=last_hidden_state.device),
+                # We need to get the first position of `eos_token_id` value (`pad_token_ids` might equal to `eos_token_id`)
+                (input_ids.to(dtype=torch.int, device=last_hidden_state.device) == self.eos_token_id)
+                .int()
+                .argmax(dim=-1),
+            ]
 
         if not return_dict:
             return (last_hidden_state, pooled_output) + encoder_outputs[1:]

src/transformers/models/clip/modeling_flax_clip.py

@@ -487,6 +487,9 @@ class FlaxCLIPTextTransformer(nn.Module):
         self.encoder = FlaxCLIPEncoder(self.config, dtype=self.dtype)
         self.final_layer_norm = nn.LayerNorm(epsilon=self.config.layer_norm_eps, dtype=self.dtype)
 
+        # For `pooled_output` computation
+        self.eos_token_id = self.config.eos_token_id
+
     def __call__(
         self,
         input_ids,
@@ -517,9 +520,18 @@ class FlaxCLIPTextTransformer(nn.Module):
         last_hidden_state = encoder_outputs[0]
         last_hidden_state = self.final_layer_norm(last_hidden_state)
 
-        # text_embeds.shape = [batch_size, sequence_length, transformer.width]
-        # take features from the EOS embedding (eos_token_id is the highest number in each sequence)
-        pooled_output = last_hidden_state[jnp.arange(last_hidden_state.shape[0]), input_ids.argmax(axis=-1)]
+        if self.eos_token_id == 2:
+            # The `eos_token_id` was incorrect before PR #24773: Let's keep what have been done here.
+            # A CLIP model with such `eos_token_id` in the config can't work correctly with extra new tokens added
+            # ------------------------------------------------------------
+            # text_embeds.shape = [batch_size, sequence_length, transformer.width]
+            # take features from the EOS embedding (eos_token_id is the highest number in each sequence)
+            pooled_output = last_hidden_state[jnp.arange(last_hidden_state.shape[0]), input_ids.argmax(axis=-1)]
+        else:
+            # (no need to cast from bool to int after comparing to `eos_token_id`)
+            pooled_output = last_hidden_state[
+                jnp.arange(last_hidden_state.shape[0]), (input_ids == self.eos_token_id).argmax(axis=-1)
+            ]
 
         if not return_dict:
             return (last_hidden_state, pooled_output) + encoder_outputs[1:]

src/transformers/models/clip/modeling_tf_clip.py

@@ -494,6 +494,9 @@ class TFCLIPTextTransformer(tf.keras.layers.Layer):
             epsilon=config.layer_norm_eps, name="final_layer_norm"
         )
 
+        # For `pooled_output` computation
+        self.eos_token_id = config.eos_token_id
+
     def call(
         self,
         input_ids: TFModelInputType,
@@ -530,14 +533,30 @@ class TFCLIPTextTransformer(tf.keras.layers.Layer):
         sequence_output = encoder_outputs[0]
         sequence_output = self.final_layer_norm(inputs=sequence_output)
 
-        # text_embeds.shape = [batch_size, n_ctx, transformer.width]
-        # take features from the eot embedding (eot_token is the highest number in each sequence)
-        pooled_output = tf.gather_nd(
-            params=sequence_output,
-            indices=tf.stack(
-                values=(tf.range(input_shape[0], dtype=tf.int64), tf.math.argmax(input_ids, axis=-1)), axis=1
-            ),
-        )
+        if self.eos_token_id == 2:
+            # The `eos_token_id` was incorrect before PR #24773: Let's keep what have been done here.
+            # A CLIP model with such `eos_token_id` in the config can't work correctly with extra new tokens added
+            # ------------------------------------------------------------
+            # text_embeds.shape = [batch_size, n_ctx, transformer.width]
+            # take features from the eot embedding (eot_token is the highest number in each sequence)
+            pooled_output = tf.gather_nd(
+                params=sequence_output,
+                indices=tf.stack(
+                    values=(tf.range(input_shape[0], dtype=tf.int64), tf.math.argmax(input_ids, axis=-1)), axis=1
+                ),
+            )
+        else:
+            # The config gets updated `eos_token_id` from PR #24773 (so the use of exta new tokens is possible)
+            pooled_output = tf.gather_nd(
+                params=sequence_output,
+                indices=tf.stack(
+                    values=(
+                        tf.range(input_shape[0], dtype=tf.int64),
+                        tf.math.argmax(tf.cast(input_ids == self.eos_token_id, dtype=tf.int8), axis=-1),
+                    ),
+                    axis=1,
+                ),
+            )
 
         if not return_dict:
             return (sequence_output, pooled_output) + encoder_outputs[1:]

src/transformers/models/clipseg/configuration_clipseg.py

@@ -97,8 +97,8 @@ class CLIPSegTextConfig(PretrainedConfig):
         initializer_range=0.02,
         initializer_factor=1.0,
         pad_token_id=1,
-        bos_token_id=0,
-        eos_token_id=2,
+        bos_token_id=49406,
+        eos_token_id=49407,
         **kwargs,
     ):
         super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)

src/transformers/models/clipseg/modeling_clipseg.py

@@ -712,6 +712,9 @@ class CLIPSegTextTransformer(nn.Module):
         self.encoder = CLIPSegEncoder(config)
         self.final_layer_norm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
 
+        # For `pooled_output` computation
+        self.eos_token_id = config.eos_token_id
+
     @add_start_docstrings_to_model_forward(CLIPSEG_TEXT_INPUTS_DOCSTRING)
     @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=CLIPSegTextConfig)
     # Copied from transformers.models.clip.modeling_clip.CLIPTextTransformer.forward with clip->clipseg, CLIP->CLIPSeg
@@ -762,13 +765,26 @@ class CLIPSegTextTransformer(nn.Module):
         last_hidden_state = encoder_outputs[0]
         last_hidden_state = self.final_layer_norm(last_hidden_state)
 
-        # text_embeds.shape = [batch_size, sequence_length, transformer.width]
-        # take features from the eot embedding (eot_token is the highest number in each sequence)
-        # casting to torch.int for onnx compatibility: argmax doesn't support int64 inputs with opset 14
-        pooled_output = last_hidden_state[
-            torch.arange(last_hidden_state.shape[0], device=last_hidden_state.device),
-            input_ids.to(dtype=torch.int, device=last_hidden_state.device).argmax(dim=-1),
-        ]
+        if self.eos_token_id == 2:
+            # The `eos_token_id` was incorrect before PR #24773: Let's keep what have been done here.
+            # A CLIPSeg model with such `eos_token_id` in the config can't work correctly with extra new tokens added
+            # ------------------------------------------------------------
+            # text_embeds.shape = [batch_size, sequence_length, transformer.width]
+            # take features from the eot embedding (eot_token is the highest number in each sequence)
+            # casting to torch.int for onnx compatibility: argmax doesn't support int64 inputs with opset 14
+            pooled_output = last_hidden_state[
+                torch.arange(last_hidden_state.shape[0], device=last_hidden_state.device),
+                input_ids.to(dtype=torch.int, device=last_hidden_state.device).argmax(dim=-1),
+            ]
+        else:
+            # The config gets updated `eos_token_id` from PR #24773 (so the use of exta new tokens is possible)
+            pooled_output = last_hidden_state[
+                torch.arange(last_hidden_state.shape[0], device=last_hidden_state.device),
+                # We need to get the first position of `eos_token_id` value (`pad_token_ids` might equal to `eos_token_id`)
+                (input_ids.to(dtype=torch.int, device=last_hidden_state.device) == self.eos_token_id)
+                .int()
+                .argmax(dim=-1),
+            ]
 
         if not return_dict:
             return (last_hidden_state, pooled_output) + encoder_outputs[1:]

src/transformers/models/groupvit/configuration_groupvit.py

@@ -106,8 +106,8 @@ class GroupViTTextConfig(PretrainedConfig):
         initializer_range=0.02,
         initializer_factor=1.0,
         pad_token_id=1,
-        bos_token_id=0,
-        eos_token_id=2,
+        bos_token_id=49406,
+        eos_token_id=49407,
         **kwargs,
     ):
         super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)

src/transformers/models/groupvit/modeling_groupvit.py

@@ -1095,6 +1095,9 @@ class GroupViTTextTransformer(nn.Module):
         self.encoder = GroupViTTextEncoder(config)
         self.final_layer_norm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
 
+        # For `pooled_output` computation
+        self.eos_token_id = config.eos_token_id
+
     @add_start_docstrings_to_model_forward(GROUPVIT_TEXT_INPUTS_DOCSTRING)
     @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=GroupViTTextConfig)
     def forward(
@@ -1144,13 +1147,26 @@ class GroupViTTextTransformer(nn.Module):
         last_hidden_state = encoder_outputs[0]
         last_hidden_state = self.final_layer_norm(last_hidden_state)
 
-        # text_embeds.shape = [batch_size, sequence_length, transformer.width]
-        # take features from the eot embedding (eot_token is the highest number in each sequence)
-        # casting to torch.int for onnx compatibility: argmax doesn't support int64 inputs with opset 14
-        pooled_output = last_hidden_state[
-            torch.arange(last_hidden_state.shape[0], device=last_hidden_state.device),
-            input_ids.to(dtype=torch.int, device=last_hidden_state.device).argmax(dim=-1),
-        ]
+        if self.eos_token_id == 2:
+            # The `eos_token_id` was incorrect before PR #24773: Let's keep what have been done here.
+            # A CLIP model with such `eos_token_id` in the config can't work correctly with extra new tokens added
+            # ------------------------------------------------------------
+            # text_embeds.shape = [batch_size, sequence_length, transformer.width]
+            # take features from the eot embedding (eot_token is the highest number in each sequence)
+            # casting to torch.int for onnx compatibility: argmax doesn't support int64 inputs with opset 14
+            pooled_output = last_hidden_state[
+                torch.arange(last_hidden_state.shape[0], device=last_hidden_state.device),
+                input_ids.to(dtype=torch.int, device=last_hidden_state.device).argmax(dim=-1),
+            ]
+        else:
+            # The config gets updated `eos_token_id` from PR #24773 (so the use of exta new tokens is possible)
+            pooled_output = last_hidden_state[
+                torch.arange(last_hidden_state.shape[0], device=last_hidden_state.device),
+                # We need to get the first position of `eos_token_id` value (`pad_token_ids` might equal to `eos_token_id`)
+                (input_ids.to(dtype=torch.int, device=last_hidden_state.device) == self.eos_token_id)
+                .int()
+                .argmax(dim=-1),
+            ]
 
         if not return_dict:
             return (last_hidden_state, pooled_output) + encoder_outputs[1:]

src/transformers/models/groupvit/modeling_tf_groupvit.py

@@ -1002,6 +1002,9 @@ class TFGroupViTTextTransformer(tf.keras.layers.Layer):
             epsilon=config.layer_norm_eps, name="final_layer_norm"
         )
 
+        # For `pooled_output` computation
+        self.eos_token_id = config.eos_token_id
+
     def call(
         self,
         input_ids: TFModelInputType,
@@ -1038,14 +1041,30 @@ class TFGroupViTTextTransformer(tf.keras.layers.Layer):
         sequence_output = encoder_outputs[0]
         sequence_output = self.final_layer_norm(inputs=sequence_output)
 
-        # text_embeds.shape = [batch_size, n_ctx, transformer.width]
-        # take features from the eot embedding (eot_token is the highest number in each sequence)
-        pooled_output = tf.gather_nd(
-            params=sequence_output,
-            indices=tf.stack(
-                values=(tf.range(input_shape[0], dtype=tf.int64), tf.math.argmax(input_ids, axis=-1)), axis=1
-            ),
-        )
+        if self.eos_token_id == 2:
+            # The `eos_token_id` was incorrect before PR #24773: Let's keep what have been done here.
+            # A CLIP model with such `eos_token_id` in the config can't work correctly with extra new tokens added
+            # ------------------------------------------------------------
+            # text_embeds.shape = [batch_size, n_ctx, transformer.width]
+            # take features from the eot embedding (eot_token is the highest number in each sequence)
+            pooled_output = tf.gather_nd(
+                params=sequence_output,
+                indices=tf.stack(
+                    values=(tf.range(input_shape[0], dtype=tf.int64), tf.math.argmax(input_ids, axis=-1)), axis=1
+                ),
+            )
+        else:
+            # The config gets updated `eos_token_id` from PR #24773 (so the use of exta new tokens is possible)
+            pooled_output = tf.gather_nd(
+                params=sequence_output,
+                indices=tf.stack(
+                    values=(
+                        tf.range(input_shape[0], dtype=tf.int64),
+                        tf.math.argmax(tf.cast(input_ids == self.eos_token_id, dtype=tf.int8), axis=-1),
+                    ),
+                    axis=1,
+                ),
+            )
 
         if not return_dict:
             return (sequence_output, pooled_output) + encoder_outputs[1:]