Source code for transformers.models.vit.modeling_vit

# coding=utf-8
# Copyright 2021 Google AI, Ross Wightman, The HuggingFace Inc. team. All rights reserved.
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# See the License for the specific language governing permissions and
# limitations under the License.
""" PyTorch ViT model. """

import math

import torch
import torch.utils.checkpoint
from torch import nn
from torch.nn import CrossEntropyLoss, MSELoss

from ...activations import ACT2FN
from ...file_utils import add_start_docstrings, add_start_docstrings_to_model_forward, replace_return_docstrings
from ...modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling, SequenceClassifierOutput
from ...modeling_utils import PreTrainedModel, find_pruneable_heads_and_indices, prune_linear_layer
from ...utils import logging
from .configuration_vit import ViTConfig

logger = logging.get_logger(__name__)

_CHECKPOINT_FOR_DOC = "google/vit-base-patch16-224"

    # See all ViT models at

# Inspired by
# From PyTorch internals
def to_2tuple(x):
    if isinstance(x,
        return x
    return (x, x)

# Based on timm implementation, which can be found here:

class ViTEmbeddings(nn.Module):
    Construct the CLS token, position and patch embeddings.


    def __init__(self, config):

        self.cls_token = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
        self.patch_embeddings = PatchEmbeddings(
        num_patches = self.patch_embeddings.num_patches
        self.position_embeddings = nn.Parameter(torch.zeros(1, num_patches + 1, config.hidden_size))
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.config = config

    def interpolate_pos_encoding(self, embeddings, height, width):
        This method allows to interpolate the pre-trained position encodings, to be able to use the model on higher
        resolution images.


        npatch = embeddings.shape[1] - 1
        N = self.position_embeddings.shape[1] - 1
        if npatch == N and height == width:
            return self.position_embeddings
        class_pos_embed = self.position_embeddings[:, 0]
        patch_pos_embed = self.position_embeddings[:, 1:]
        dim = embeddings.shape[-1]
        h0 = height // self.config.patch_size
        w0 = width // self.config.patch_size
        # we add a small number to avoid floating point error in the interpolation
        # see discussion at
        h0, w0 = h0 + 0.1, w0 + 0.1
        patch_pos_embed = nn.functional.interpolate(
            patch_pos_embed.reshape(1, int(math.sqrt(N)), int(math.sqrt(N)), dim).permute(0, 3, 1, 2),
            scale_factor=(h0 / math.sqrt(N), w0 / math.sqrt(N)),
        assert int(h0) == patch_pos_embed.shape[-2] and int(w0) == patch_pos_embed.shape[-1]
        patch_pos_embed = patch_pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
        return, patch_pos_embed), dim=1)

    def forward(self, pixel_values, interpolate_pos_encoding=False):
        batch_size, num_channels, height, width = pixel_values.shape
        embeddings = self.patch_embeddings(pixel_values, interpolate_pos_encoding=interpolate_pos_encoding)

        # add the [CLS] token to the embedded patch tokens
        cls_tokens = self.cls_token.expand(batch_size, -1, -1)
        embeddings =, embeddings), dim=1)

        # add positional encoding to each token
        if interpolate_pos_encoding:
            embeddings = embeddings + self.interpolate_pos_encoding(embeddings, height, width)
            embeddings = embeddings + self.position_embeddings

        embeddings = self.dropout(embeddings)

        return embeddings

# Based on timm implementation, which can be found here:
class PatchEmbeddings(nn.Module):
    Image to Patch Embedding.


    def __init__(self, image_size=224, patch_size=16, num_channels=3, embed_dim=768):
        image_size = to_2tuple(image_size)
        patch_size = to_2tuple(patch_size)
        num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
        self.image_size = image_size
        self.patch_size = patch_size
        self.num_patches = num_patches

        self.projection = nn.Conv2d(num_channels, embed_dim, kernel_size=patch_size, stride=patch_size)

    def forward(self, pixel_values, interpolate_pos_encoding=False):
        batch_size, num_channels, height, width = pixel_values.shape
        if not interpolate_pos_encoding:
            if height != self.image_size[0] or width != self.image_size[1]:
                raise ValueError(
                    f"Input image size ({height}*{width}) doesn't match model ({self.image_size[0]}*{self.image_size[1]})."
        x = self.projection(pixel_values).flatten(2).transpose(1, 2)
        return x

class ViTSelfAttention(nn.Module):
    def __init__(self, config):
        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
            raise ValueError(
                f"The hidden size {config.hidden_size,} is not a multiple of the number of attention "
                f"heads {config.num_attention_heads}."

        self.num_attention_heads = config.num_attention_heads
        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
        self.all_head_size = self.num_attention_heads * self.attention_head_size

        self.query = nn.Linear(config.hidden_size, self.all_head_size)
        self.key = nn.Linear(config.hidden_size, self.all_head_size)
        self.value = nn.Linear(config.hidden_size, self.all_head_size)

        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)

    def transpose_for_scores(self, x):
        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
        x = x.view(*new_x_shape)
        return x.permute(0, 2, 1, 3)

    def forward(self, hidden_states, head_mask=None, output_attentions=False):
        mixed_query_layer = self.query(hidden_states)

        key_layer = self.transpose_for_scores(self.key(hidden_states))
        value_layer = self.transpose_for_scores(self.value(hidden_states))
        query_layer = self.transpose_for_scores(mixed_query_layer)

        # Take the dot product between "query" and "key" to get the raw attention scores.
        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))

        attention_scores = attention_scores / math.sqrt(self.attention_head_size)

        # Normalize the attention scores to probabilities.
        attention_probs = nn.Softmax(dim=-1)(attention_scores)

        # This is actually dropping out entire tokens to attend to, which might
        # seem a bit unusual, but is taken from the original Transformer paper.
        attention_probs = self.dropout(attention_probs)

        # Mask heads if we want to
        if head_mask is not None:
            attention_probs = attention_probs * head_mask

        context_layer = torch.matmul(attention_probs, value_layer)

        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
        context_layer = context_layer.view(*new_context_layer_shape)

        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)

        return outputs

class ViTSelfOutput(nn.Module):
    The residual connection is defined in ViTLayer instead of here (as is the case with other models), due to the
    layernorm applied before each block.

    def __init__(self, config):
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

    def forward(self, hidden_states, input_tensor):

        hidden_states = self.dense(hidden_states)
        hidden_states = self.dropout(hidden_states)

        return hidden_states

class ViTAttention(nn.Module):
    def __init__(self, config):
        self.attention = ViTSelfAttention(config)
        self.output = ViTSelfOutput(config)
        self.pruned_heads = set()

    def prune_heads(self, heads):
        if len(heads) == 0:
        heads, index = find_pruneable_heads_and_indices(
            heads, self.attention.num_attention_heads, self.attention.attention_head_size, self.pruned_heads

        # Prune linear layers
        self.attention.query = prune_linear_layer(self.attention.query, index)
        self.attention.key = prune_linear_layer(self.attention.key, index)
        self.attention.value = prune_linear_layer(self.attention.value, index)
        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)

        # Update hyper params and store pruned heads
        self.attention.num_attention_heads = self.attention.num_attention_heads - len(heads)
        self.attention.all_head_size = self.attention.attention_head_size * self.attention.num_attention_heads
        self.pruned_heads = self.pruned_heads.union(heads)

    def forward(self, hidden_states, head_mask=None, output_attentions=False):
        self_outputs = self.attention(hidden_states, head_mask, output_attentions)

        attention_output = self.output(self_outputs[0], hidden_states)

        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
        return outputs

class ViTIntermediate(nn.Module):
    def __init__(self, config):
        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
        if isinstance(config.hidden_act, str):
            self.intermediate_act_fn = ACT2FN[config.hidden_act]
            self.intermediate_act_fn = config.hidden_act

    def forward(self, hidden_states):

        hidden_states = self.dense(hidden_states)
        hidden_states = self.intermediate_act_fn(hidden_states)

        return hidden_states

class ViTOutput(nn.Module):
    def __init__(self, config):
        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

    def forward(self, hidden_states, input_tensor):
        hidden_states = self.dense(hidden_states)
        hidden_states = self.dropout(hidden_states)

        hidden_states = hidden_states + input_tensor

        return hidden_states

class ViTLayer(nn.Module):
    """This corresponds to the Block class in the timm implementation."""

    def __init__(self, config):
        self.chunk_size_feed_forward = config.chunk_size_feed_forward
        self.seq_len_dim = 1
        self.attention = ViTAttention(config)
        self.intermediate = ViTIntermediate(config)
        self.output = ViTOutput(config)
        self.layernorm_before = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.layernorm_after = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)

    def forward(self, hidden_states, head_mask=None, output_attentions=False):
        self_attention_outputs = self.attention(
            self.layernorm_before(hidden_states),  # in ViT, layernorm is applied before self-attention
        attention_output = self_attention_outputs[0]
        outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights

        # first residual connection
        hidden_states = attention_output + hidden_states

        # in ViT, layernorm is also applied after self-attention
        layer_output = self.layernorm_after(hidden_states)

        # TODO feedforward chunking not working for now
        # layer_output = apply_chunking_to_forward(
        #     self.feed_forward_chunk, self.chunk_size_feed_forward, self.seq_len_dim, layer_output
        # )

        layer_output = self.intermediate(layer_output)

        # second residual connection is done here
        layer_output = self.output(layer_output, hidden_states)

        outputs = (layer_output,) + outputs

        return outputs

    def feed_forward_chunk(self, attention_output):
        intermediate_output = self.intermediate(attention_output)
        layer_output = self.output(intermediate_output)
        return layer_output

class ViTEncoder(nn.Module):
    def __init__(self, config):
        self.config = config
        self.layer = nn.ModuleList([ViTLayer(config) for _ in range(config.num_hidden_layers)])
        self.gradient_checkpointing = False

    def forward(
        all_hidden_states = () if output_hidden_states else None
        all_self_attentions = () if output_attentions else None

        for i, layer_module in enumerate(self.layer):
            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

            layer_head_mask = head_mask[i] if head_mask is not None else None

            if self.gradient_checkpointing and

                def create_custom_forward(module):
                    def custom_forward(*inputs):
                        return module(*inputs, output_attentions)

                    return custom_forward

                layer_outputs = torch.utils.checkpoint.checkpoint(
                layer_outputs = layer_module(hidden_states, layer_head_mask, output_attentions)

            hidden_states = layer_outputs[0]

            if output_attentions:
                all_self_attentions = all_self_attentions + (layer_outputs[1],)

        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        if not return_dict:
            return tuple(v for v in [hidden_states, all_hidden_states, all_self_attentions] if v is not None)
        return BaseModelOutput(

class ViTPreTrainedModel(PreTrainedModel):
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained

    config_class = ViTConfig
    base_model_prefix = "vit"
    supports_gradient_checkpointing = True

    def _init_weights(self, module):
        """Initialize the weights"""
        if isinstance(module, (nn.Linear, nn.Conv2d)):
            # Slightly different from the TF version which uses truncated_normal for initialization
            # cf
  , std=self.config.initializer_range)
            if module.bias is not None:
        elif isinstance(module, nn.Embedding):
  , std=self.config.initializer_range)
            if module.padding_idx is not None:
        elif isinstance(module, nn.LayerNorm):

    def _set_gradient_checkpointing(self, module, value=False):
        if isinstance(module, ViTEncoder):
            module.gradient_checkpointing = value

    This model is a PyTorch `torch.nn.Module <>`_ subclass. Use
    it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and

        config (:class:`~transformers.ViTConfig`): Model configuration class with all the parameters of the model.
            Initializing with a config file does not load the weights associated with the model, only the
            configuration. Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model

        pixel_values (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, num_channels, height, width)`):
            Pixel values. Pixel values can be obtained using :class:`~transformers.ViTFeatureExtractor`. See
            :meth:`transformers.ViTFeatureExtractor.__call__` for details.

        head_mask (:obj:`torch.FloatTensor` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`):
            Mask to nullify selected heads of the self-attention modules. Mask values selected in ``[0, 1]``:

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.

        output_attentions (:obj:`bool`, `optional`):
            Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
            tensors for more detail.
        output_hidden_states (:obj:`bool`, `optional`):
            Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
            more detail.
        interpolate_pos_encoding (:obj:`bool`, `optional`):
            Whether to interpolate the pre-trained position encodings.
        return_dict (:obj:`bool`, `optional`):
            Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.

[docs]@add_start_docstrings( "The bare ViT Model transformer outputting raw hidden-states without any specific head on top.", VIT_START_DOCSTRING, ) class ViTModel(ViTPreTrainedModel): def __init__(self, config, add_pooling_layer=True): super().__init__(config) self.config = config self.embeddings = ViTEmbeddings(config) self.encoder = ViTEncoder(config) self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) self.pooler = ViTPooler(config) if add_pooling_layer else None self.init_weights() def get_input_embeddings(self): return self.embeddings.patch_embeddings def _prune_heads(self, heads_to_prune): """ Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base class PreTrainedModel """ for layer, heads in heads_to_prune.items(): self.encoder.layer[layer].attention.prune_heads(heads)
[docs] @add_start_docstrings_to_model_forward(VIT_INPUTS_DOCSTRING) @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=_CONFIG_FOR_DOC) def forward( self, pixel_values=None, head_mask=None, output_attentions=None, output_hidden_states=None, interpolate_pos_encoding=None, return_dict=None, ): r""" Returns: Examples:: >>> from transformers import ViTFeatureExtractor, ViTModel >>> from PIL import Image >>> import requests >>> url = '' >>> image =, stream=True).raw) >>> feature_extractor = ViTFeatureExtractor.from_pretrained('google/vit-base-patch16-224-in21k') >>> model = ViTModel.from_pretrained('google/vit-base-patch16-224-in21k') >>> inputs = feature_extractor(images=image, return_tensors="pt") >>> outputs = model(**inputs) >>> last_hidden_states = outputs.last_hidden_state """ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) return_dict = return_dict if return_dict is not None else self.config.use_return_dict if pixel_values is None: raise ValueError("You have to specify pixel_values") # Prepare head mask if needed # 1.0 in head_mask indicate we keep the head # attention_probs has shape bsz x n_heads x N x N # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads] # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length] head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers) embedding_output = self.embeddings(pixel_values, interpolate_pos_encoding=interpolate_pos_encoding) encoder_outputs = self.encoder( embedding_output, head_mask=head_mask, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) sequence_output = encoder_outputs[0] sequence_output = self.layernorm(sequence_output) pooled_output = self.pooler(sequence_output) if self.pooler is not None else None if not return_dict: return (sequence_output, pooled_output) + encoder_outputs[1:] return BaseModelOutputWithPooling( last_hidden_state=sequence_output, pooler_output=pooled_output, hidden_states=encoder_outputs.hidden_states, attentions=encoder_outputs.attentions, )
class ViTPooler(nn.Module): def __init__(self, config): super().__init__() self.dense = nn.Linear(config.hidden_size, config.hidden_size) self.activation = nn.Tanh() def forward(self, hidden_states): # We "pool" the model by simply taking the hidden state corresponding # to the first token. first_token_tensor = hidden_states[:, 0] pooled_output = self.dense(first_token_tensor) pooled_output = self.activation(pooled_output) return pooled_output
[docs]@add_start_docstrings( """ ViT Model transformer with an image classification head on top (a linear layer on top of the final hidden state of the [CLS] token) e.g. for ImageNet. """, VIT_START_DOCSTRING, ) class ViTForImageClassification(ViTPreTrainedModel): def __init__(self, config): super().__init__(config) self.num_labels = config.num_labels self.vit = ViTModel(config, add_pooling_layer=False) # Classifier head self.classifier = nn.Linear(config.hidden_size, config.num_labels) if config.num_labels > 0 else nn.Identity() self.init_weights()
[docs] @add_start_docstrings_to_model_forward(VIT_INPUTS_DOCSTRING) @replace_return_docstrings(output_type=SequenceClassifierOutput, config_class=_CONFIG_FOR_DOC) def forward( self, pixel_values=None, head_mask=None, labels=None, output_attentions=None, output_hidden_states=None, interpolate_pos_encoding=None, return_dict=None, ): r""" labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`): Labels for computing the image classification/regression loss. Indices should be in :obj:`[0, ..., config.num_labels - 1]`. If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss), If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy). Returns: Examples:: >>> from transformers import ViTFeatureExtractor, ViTForImageClassification >>> from PIL import Image >>> import requests >>> url = '' >>> image =, stream=True).raw) >>> feature_extractor = ViTFeatureExtractor.from_pretrained('google/vit-base-patch16-224') >>> model = ViTForImageClassification.from_pretrained('google/vit-base-patch16-224') >>> inputs = feature_extractor(images=image, return_tensors="pt") >>> outputs = model(**inputs) >>> logits = outputs.logits >>> # model predicts one of the 1000 ImageNet classes >>> predicted_class_idx = logits.argmax(-1).item() >>> print("Predicted class:", model.config.id2label[predicted_class_idx]) """ return_dict = return_dict if return_dict is not None else self.config.use_return_dict outputs = self.vit( pixel_values, head_mask=head_mask, output_attentions=output_attentions, output_hidden_states=output_hidden_states, interpolate_pos_encoding=interpolate_pos_encoding, return_dict=return_dict, ) sequence_output = outputs[0] logits = self.classifier(sequence_output[:, 0, :]) loss = None if labels is not None: if self.num_labels == 1: # We are doing regression loss_fct = MSELoss() loss = loss_fct(logits.view(-1), labels.view(-1)) else: loss_fct = CrossEntropyLoss() loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1)) if not return_dict: output = (logits,) + outputs[2:] return ((loss,) + output) if loss is not None else output return SequenceClassifierOutput( loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions, )