Aluode/PerceptionLabPortable

1# coding=utf-82# Copyright 2021 NVIDIA The HuggingFace Inc. team. All rights reserved.3#4# Licensed under the Apache License, Version 2.0 (the "License");5# you may not use this file except in compliance with the License.6# You may obtain a copy of the License at7#8#     http://www.apache.org/licenses/LICENSE-2.09#10# Unless required by applicable law or agreed to in writing, software11# distributed under the License is distributed on an "AS IS" BASIS,12# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.13# See the License for the specific language governing permissions and14# limitations under the License.15"""PyTorch SegFormer model."""16 17import math18from typing import Optional, Union19 20import torch21from torch import nn22from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss23 24from ...activations import ACT2FN25from ...modeling_outputs import BaseModelOutput, ImageClassifierOutput, SemanticSegmenterOutput26from ...modeling_utils import PreTrainedModel27from ...pytorch_utils import find_pruneable_heads_and_indices, prune_linear_layer28from ...utils import auto_docstring, logging29from .configuration_segformer import SegformerConfig30 31 32logger = logging.get_logger(__name__)33 34 35class SegFormerImageClassifierOutput(ImageClassifierOutput):36    """37    Base class for outputs of image classification models.38 39    Args:40        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):41            Classification (or regression if config.num_labels==1) loss.42        logits (`torch.FloatTensor` of shape `(batch_size, config.num_labels)`):43            Classification (or regression if config.num_labels==1) scores (before SoftMax).44        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):45            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +46            one for the output of each stage) of shape `(batch_size, num_channels, height, width)`. Hidden-states (also47            called feature maps) of the model at the output of each stage.48        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):49            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, patch_size,50            sequence_length)`.51 52            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention53            heads.54    """55 56    loss: Optional[torch.FloatTensor] = None57    logits: Optional[torch.FloatTensor] = None58    hidden_states: Optional[tuple[torch.FloatTensor]] = None59    attentions: Optional[tuple[torch.FloatTensor]] = None60 61 62# Copied from transformers.models.beit.modeling_beit.drop_path63def drop_path(input: torch.Tensor, drop_prob: float = 0.0, training: bool = False) -> torch.Tensor:64    """65    Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).66 67    Comment by Ross Wightman: This is the same as the DropConnect impl I created for EfficientNet, etc networks,68    however, the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...69    See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for changing the70    layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use 'survival rate' as the71    argument.72    """73    if drop_prob == 0.0 or not training:74        return input75    keep_prob = 1 - drop_prob76    shape = (input.shape[0],) + (1,) * (input.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets77    random_tensor = keep_prob + torch.rand(shape, dtype=input.dtype, device=input.device)78    random_tensor.floor_()  # binarize79    output = input.div(keep_prob) * random_tensor80    return output81 82 83# Copied from transformers.models.convnext.modeling_convnext.ConvNextDropPath with ConvNext->Segformer84class SegformerDropPath(nn.Module):85    """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks)."""86 87    def __init__(self, drop_prob: Optional[float] = None) -> None:88        super().__init__()89        self.drop_prob = drop_prob90 91    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:92        return drop_path(hidden_states, self.drop_prob, self.training)93 94    def extra_repr(self) -> str:95        return f"p={self.drop_prob}"96 97 98class SegformerOverlapPatchEmbeddings(nn.Module):99    """Construct the overlapping patch embeddings."""100 101    def __init__(self, patch_size, stride, num_channels, hidden_size):102        super().__init__()103        self.proj = nn.Conv2d(104            num_channels,105            hidden_size,106            kernel_size=patch_size,107            stride=stride,108            padding=patch_size // 2,109        )110 111        self.layer_norm = nn.LayerNorm(hidden_size)112 113    def forward(self, pixel_values):114        embeddings = self.proj(pixel_values)115        _, _, height, width = embeddings.shape116        # (batch_size, num_channels, height, width) -> (batch_size, num_channels, height*width) -> (batch_size, height*width, num_channels)117        # this can be fed to a Transformer layer118        embeddings = embeddings.flatten(2).transpose(1, 2)119        embeddings = self.layer_norm(embeddings)120        return embeddings, height, width121 122 123class SegformerEfficientSelfAttention(nn.Module):124    """SegFormer's efficient self-attention mechanism. Employs the sequence reduction process introduced in the [PvT125    paper](https://huggingface.co/papers/2102.12122)."""126 127    def __init__(self, config, hidden_size, num_attention_heads, sequence_reduction_ratio):128        super().__init__()129        self.hidden_size = hidden_size130        self.num_attention_heads = num_attention_heads131 132        if self.hidden_size % self.num_attention_heads != 0:133            raise ValueError(134                f"The hidden size ({self.hidden_size}) is not a multiple of the number of attention "135                f"heads ({self.num_attention_heads})"136            )137 138        self.attention_head_size = int(self.hidden_size / self.num_attention_heads)139        self.all_head_size = self.num_attention_heads * self.attention_head_size140 141        self.query = nn.Linear(self.hidden_size, self.all_head_size)142        self.key = nn.Linear(self.hidden_size, self.all_head_size)143        self.value = nn.Linear(self.hidden_size, self.all_head_size)144 145        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)146 147        self.sr_ratio = sequence_reduction_ratio148        if sequence_reduction_ratio > 1:149            self.sr = nn.Conv2d(150                hidden_size, hidden_size, kernel_size=sequence_reduction_ratio, stride=sequence_reduction_ratio151            )152            self.layer_norm = nn.LayerNorm(hidden_size)153 154    def forward(155        self,156        hidden_states,157        height,158        width,159        output_attentions=False,160    ):161        batch_size, seq_length, _ = hidden_states.shape162        query_layer = (163            self.query(hidden_states)164            .view(batch_size, -1, self.num_attention_heads, self.attention_head_size)165            .transpose(1, 2)166        )167 168        if self.sr_ratio > 1:169            batch_size, seq_len, num_channels = hidden_states.shape170            # Reshape to (batch_size, num_channels, height, width)171            hidden_states = hidden_states.permute(0, 2, 1).reshape(batch_size, num_channels, height, width)172            # Apply sequence reduction173            hidden_states = self.sr(hidden_states)174            # Reshape back to (batch_size, seq_len, num_channels)175            hidden_states = hidden_states.reshape(batch_size, num_channels, -1).permute(0, 2, 1)176            hidden_states = self.layer_norm(hidden_states)177 178        key_layer = (179            self.key(hidden_states)180            .view(batch_size, -1, self.num_attention_heads, self.attention_head_size)181            .transpose(1, 2)182        )183        value_layer = (184            self.value(hidden_states)185            .view(batch_size, -1, self.num_attention_heads, self.attention_head_size)186            .transpose(1, 2)187        )188 189        # Take the dot product between "query" and "key" to get the raw attention scores.190        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))191 192        attention_scores = attention_scores / math.sqrt(self.attention_head_size)193 194        # Normalize the attention scores to probabilities.195        attention_probs = nn.functional.softmax(attention_scores, dim=-1)196 197        # This is actually dropping out entire tokens to attend to, which might198        # seem a bit unusual, but is taken from the original Transformer paper.199        attention_probs = self.dropout(attention_probs)200 201        context_layer = torch.matmul(attention_probs, value_layer)202 203        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()204        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)205        context_layer = context_layer.view(new_context_layer_shape)206 207        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)208 209        return outputs210 211 212class SegformerSelfOutput(nn.Module):213    def __init__(self, config, hidden_size):214        super().__init__()215        self.dense = nn.Linear(hidden_size, hidden_size)216        self.dropout = nn.Dropout(config.hidden_dropout_prob)217 218    def forward(self, hidden_states, input_tensor):219        hidden_states = self.dense(hidden_states)220        hidden_states = self.dropout(hidden_states)221        return hidden_states222 223 224class SegformerAttention(nn.Module):225    def __init__(self, config, hidden_size, num_attention_heads, sequence_reduction_ratio):226        super().__init__()227        self.self = SegformerEfficientSelfAttention(228            config=config,229            hidden_size=hidden_size,230            num_attention_heads=num_attention_heads,231            sequence_reduction_ratio=sequence_reduction_ratio,232        )233        self.output = SegformerSelfOutput(config, hidden_size=hidden_size)234        self.pruned_heads = set()235 236    def prune_heads(self, heads):237        if len(heads) == 0:238            return239        heads, index = find_pruneable_heads_and_indices(240            heads, self.self.num_attention_heads, self.self.attention_head_size, self.pruned_heads241        )242 243        # Prune linear layers244        self.self.query = prune_linear_layer(self.self.query, index)245        self.self.key = prune_linear_layer(self.self.key, index)246        self.self.value = prune_linear_layer(self.self.value, index)247        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)248 249        # Update hyper params and store pruned heads250        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)251        self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads252        self.pruned_heads = self.pruned_heads.union(heads)253 254    def forward(self, hidden_states, height, width, output_attentions=False):255        self_outputs = self.self(hidden_states, height, width, output_attentions)256 257        attention_output = self.output(self_outputs[0], hidden_states)258        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them259        return outputs260 261 262class SegformerDWConv(nn.Module):263    def __init__(self, dim=768):264        super().__init__()265        self.dwconv = nn.Conv2d(dim, dim, 3, 1, 1, bias=True, groups=dim)266 267    def forward(self, hidden_states, height, width):268        batch_size, seq_len, num_channels = hidden_states.shape269        hidden_states = hidden_states.transpose(1, 2).view(batch_size, num_channels, height, width)270        hidden_states = self.dwconv(hidden_states)271        hidden_states = hidden_states.flatten(2).transpose(1, 2)272 273        return hidden_states274 275 276class SegformerMixFFN(nn.Module):277    def __init__(self, config, in_features, hidden_features=None, out_features=None):278        super().__init__()279        out_features = out_features or in_features280        self.dense1 = nn.Linear(in_features, hidden_features)281        self.dwconv = SegformerDWConv(hidden_features)282        if isinstance(config.hidden_act, str):283            self.intermediate_act_fn = ACT2FN[config.hidden_act]284        else:285            self.intermediate_act_fn = config.hidden_act286        self.dense2 = nn.Linear(hidden_features, out_features)287        self.dropout = nn.Dropout(config.hidden_dropout_prob)288 289    def forward(self, hidden_states, height, width):290        hidden_states = self.dense1(hidden_states)291        hidden_states = self.dwconv(hidden_states, height, width)292        hidden_states = self.intermediate_act_fn(hidden_states)293        hidden_states = self.dropout(hidden_states)294        hidden_states = self.dense2(hidden_states)295        hidden_states = self.dropout(hidden_states)296        return hidden_states297 298 299class SegformerLayer(nn.Module):300    """This corresponds to the Block class in the original implementation."""301 302    def __init__(self, config, hidden_size, num_attention_heads, drop_path, sequence_reduction_ratio, mlp_ratio):303        super().__init__()304        self.layer_norm_1 = nn.LayerNorm(hidden_size)305        self.attention = SegformerAttention(306            config,307            hidden_size=hidden_size,308            num_attention_heads=num_attention_heads,309            sequence_reduction_ratio=sequence_reduction_ratio,310        )311        self.drop_path = SegformerDropPath(drop_path) if drop_path > 0.0 else nn.Identity()312        self.layer_norm_2 = nn.LayerNorm(hidden_size)313        mlp_hidden_size = int(hidden_size * mlp_ratio)314        self.mlp = SegformerMixFFN(config, in_features=hidden_size, hidden_features=mlp_hidden_size)315 316    def forward(self, hidden_states, height, width, output_attentions=False):317        self_attention_outputs = self.attention(318            self.layer_norm_1(hidden_states),  # in Segformer, layernorm is applied before self-attention319            height,320            width,321            output_attentions=output_attentions,322        )323 324        attention_output = self_attention_outputs[0]325        outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights326 327        # first residual connection (with stochastic depth)328        attention_output = self.drop_path(attention_output)329        hidden_states = attention_output + hidden_states330 331        mlp_output = self.mlp(self.layer_norm_2(hidden_states), height, width)332 333        # second residual connection (with stochastic depth)334        mlp_output = self.drop_path(mlp_output)335        layer_output = mlp_output + hidden_states336 337        outputs = (layer_output,) + outputs338 339        return outputs340 341 342class SegformerEncoder(nn.Module):343    def __init__(self, config):344        super().__init__()345        self.config = config346 347        # stochastic depth decay rule348        drop_path_decays = [349            x.item() for x in torch.linspace(0, config.drop_path_rate, sum(config.depths), device="cpu")350        ]351 352        # patch embeddings353        embeddings = []354        for i in range(config.num_encoder_blocks):355            embeddings.append(356                SegformerOverlapPatchEmbeddings(357                    patch_size=config.patch_sizes[i],358                    stride=config.strides[i],359                    num_channels=config.num_channels if i == 0 else config.hidden_sizes[i - 1],360                    hidden_size=config.hidden_sizes[i],361                )362            )363        self.patch_embeddings = nn.ModuleList(embeddings)364 365        # Transformer blocks366        blocks = []367        cur = 0368        for i in range(config.num_encoder_blocks):369            # each block consists of layers370            layers = []371            if i != 0:372                cur += config.depths[i - 1]373            for j in range(config.depths[i]):374                layers.append(375                    SegformerLayer(376                        config,377                        hidden_size=config.hidden_sizes[i],378                        num_attention_heads=config.num_attention_heads[i],379                        drop_path=drop_path_decays[cur + j],380                        sequence_reduction_ratio=config.sr_ratios[i],381                        mlp_ratio=config.mlp_ratios[i],382                    )383                )384            blocks.append(nn.ModuleList(layers))385 386        self.block = nn.ModuleList(blocks)387 388        # Layer norms389        self.layer_norm = nn.ModuleList(390            [nn.LayerNorm(config.hidden_sizes[i]) for i in range(config.num_encoder_blocks)]391        )392 393    def forward(394        self,395        pixel_values: torch.FloatTensor,396        output_attentions: Optional[bool] = False,397        output_hidden_states: Optional[bool] = False,398        return_dict: Optional[bool] = True,399    ) -> Union[tuple, BaseModelOutput]:400        all_hidden_states = () if output_hidden_states else None401        all_self_attentions = () if output_attentions else None402 403        batch_size = pixel_values.shape[0]404 405        hidden_states = pixel_values406        for idx, x in enumerate(zip(self.patch_embeddings, self.block, self.layer_norm)):407            embedding_layer, block_layer, norm_layer = x408            # first, obtain patch embeddings409            hidden_states, height, width = embedding_layer(hidden_states)410            # second, send embeddings through blocks411            for i, blk in enumerate(block_layer):412                layer_outputs = blk(hidden_states, height, width, output_attentions)413                hidden_states = layer_outputs[0]414                if output_attentions:415                    all_self_attentions = all_self_attentions + (layer_outputs[1],)416            # third, apply layer norm417            hidden_states = norm_layer(hidden_states)418            # fourth, optionally reshape back to (batch_size, num_channels, height, width)419            if idx != len(self.patch_embeddings) - 1 or (420                idx == len(self.patch_embeddings) - 1 and self.config.reshape_last_stage421            ):422                hidden_states = hidden_states.reshape(batch_size, height, width, -1).permute(0, 3, 1, 2).contiguous()423            if output_hidden_states:424                all_hidden_states = all_hidden_states + (hidden_states,)425 426        if not return_dict:427            return tuple(v for v in [hidden_states, all_hidden_states, all_self_attentions] if v is not None)428        return BaseModelOutput(429            last_hidden_state=hidden_states,430            hidden_states=all_hidden_states,431            attentions=all_self_attentions,432        )433 434 435@auto_docstring436class SegformerPreTrainedModel(PreTrainedModel):437    config: SegformerConfig438    base_model_prefix = "segformer"439    main_input_name = "pixel_values"440 441    def _init_weights(self, module):442        """Initialize the weights"""443        if isinstance(module, (nn.Linear, nn.Conv2d)):444            # Slightly different from the TF version which uses truncated_normal for initialization445            # cf https://github.com/pytorch/pytorch/pull/5617446            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)447            if module.bias is not None:448                module.bias.data.zero_()449        elif isinstance(module, nn.Embedding):450            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)451            if module.padding_idx is not None:452                module.weight.data[module.padding_idx].zero_()453        elif isinstance(module, (nn.LayerNorm, nn.BatchNorm2d)):454            module.bias.data.zero_()455            module.weight.data.fill_(1.0)456 457 458@auto_docstring459class SegformerModel(SegformerPreTrainedModel):460    def __init__(self, config):461        super().__init__(config)462        self.config = config463 464        # hierarchical Transformer encoder465        self.encoder = SegformerEncoder(config)466 467        # Initialize weights and apply final processing468        self.post_init()469 470    def _prune_heads(self, heads_to_prune):471        """472        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base473        class PreTrainedModel474        """475        for layer, heads in heads_to_prune.items():476            self.encoder.layer[layer].attention.prune_heads(heads)477 478    @auto_docstring479    def forward(480        self,481        pixel_values: torch.FloatTensor,482        output_attentions: Optional[bool] = None,483        output_hidden_states: Optional[bool] = None,484        return_dict: Optional[bool] = None,485    ) -> Union[tuple, BaseModelOutput]:486        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions487        output_hidden_states = (488            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states489        )490        return_dict = return_dict if return_dict is not None else self.config.use_return_dict491 492        encoder_outputs = self.encoder(493            pixel_values,494            output_attentions=output_attentions,495            output_hidden_states=output_hidden_states,496            return_dict=return_dict,497        )498        sequence_output = encoder_outputs[0]499 500        if not return_dict:501            return (sequence_output,) + encoder_outputs[1:]502 503        return BaseModelOutput(504            last_hidden_state=sequence_output,505            hidden_states=encoder_outputs.hidden_states,506            attentions=encoder_outputs.attentions,507        )508 509 510@auto_docstring(511    custom_intro="""512    SegFormer Model transformer with an image classification head on top (a linear layer on top of the final hidden513    states) e.g. for ImageNet.514    """515)516class SegformerForImageClassification(SegformerPreTrainedModel):517    def __init__(self, config):518        super().__init__(config)519 520        self.num_labels = config.num_labels521        self.segformer = SegformerModel(config)522 523        # Classifier head524        self.classifier = nn.Linear(config.hidden_sizes[-1], config.num_labels)525 526        # Initialize weights and apply final processing527        self.post_init()528 529    @auto_docstring530    def forward(531        self,532        pixel_values: Optional[torch.FloatTensor] = None,533        labels: Optional[torch.LongTensor] = None,534        output_attentions: Optional[bool] = None,535        output_hidden_states: Optional[bool] = None,536        return_dict: Optional[bool] = None,537    ) -> Union[tuple, SegFormerImageClassifierOutput]:538        r"""539        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):540            Labels for computing the image classification/regression loss. Indices should be in `[0, ...,541            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If542            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).543        """544        return_dict = return_dict if return_dict is not None else self.config.use_return_dict545 546        outputs = self.segformer(547            pixel_values,548            output_attentions=output_attentions,549            output_hidden_states=output_hidden_states,550            return_dict=return_dict,551        )552 553        sequence_output = outputs[0]554 555        # convert last hidden states to (batch_size, height*width, hidden_size)556        batch_size = sequence_output.shape[0]557        if self.config.reshape_last_stage:558            # (batch_size, num_channels, height, width) -> (batch_size, height, width, num_channels)559            sequence_output = sequence_output.permute(0, 2, 3, 1)560        sequence_output = sequence_output.reshape(batch_size, -1, self.config.hidden_sizes[-1])561 562        # global average pooling563        sequence_output = sequence_output.mean(dim=1)564 565        logits = self.classifier(sequence_output)566 567        loss = None568        if labels is not None:569            loss = self.loss_function(labels, logits, self.config)570 571        if not return_dict:572            output = (logits,) + outputs[1:]573            return ((loss,) + output) if loss is not None else output574 575        return SegFormerImageClassifierOutput(576            loss=loss,577            logits=logits,578            hidden_states=outputs.hidden_states,579            attentions=outputs.attentions,580        )581 582 583class SegformerMLP(nn.Module):584    """585    Linear Embedding.586    """587 588    def __init__(self, config: SegformerConfig, input_dim):589        super().__init__()590        self.proj = nn.Linear(input_dim, config.decoder_hidden_size)591 592    def forward(self, hidden_states: torch.Tensor):593        hidden_states = hidden_states.flatten(2).transpose(1, 2)594        hidden_states = self.proj(hidden_states)595        return hidden_states596 597 598class SegformerDecodeHead(SegformerPreTrainedModel):599    def __init__(self, config):600        super().__init__(config)601        # linear layers which will unify the channel dimension of each of the encoder blocks to the same config.decoder_hidden_size602        mlps = []603        for i in range(config.num_encoder_blocks):604            mlp = SegformerMLP(config, input_dim=config.hidden_sizes[i])605            mlps.append(mlp)606        self.linear_c = nn.ModuleList(mlps)607 608        # the following 3 layers implement the ConvModule of the original implementation609        self.linear_fuse = nn.Conv2d(610            in_channels=config.decoder_hidden_size * config.num_encoder_blocks,611            out_channels=config.decoder_hidden_size,612            kernel_size=1,613            bias=False,614        )615        self.batch_norm = nn.BatchNorm2d(config.decoder_hidden_size)616        self.activation = nn.ReLU()617 618        self.dropout = nn.Dropout(config.classifier_dropout_prob)619        self.classifier = nn.Conv2d(config.decoder_hidden_size, config.num_labels, kernel_size=1)620 621        self.config = config622 623    def forward(self, encoder_hidden_states: torch.FloatTensor) -> torch.Tensor:624        batch_size = encoder_hidden_states[-1].shape[0]625 626        all_hidden_states = ()627        for encoder_hidden_state, mlp in zip(encoder_hidden_states, self.linear_c):628            if self.config.reshape_last_stage is False and encoder_hidden_state.ndim == 3:629                height = width = int(math.sqrt(encoder_hidden_state.shape[-1]))630                encoder_hidden_state = (631                    encoder_hidden_state.reshape(batch_size, height, width, -1).permute(0, 3, 1, 2).contiguous()632                )633 634            # unify channel dimension635            height, width = encoder_hidden_state.shape[2], encoder_hidden_state.shape[3]636            encoder_hidden_state = mlp(encoder_hidden_state)637            encoder_hidden_state = encoder_hidden_state.permute(0, 2, 1)638            encoder_hidden_state = encoder_hidden_state.reshape(batch_size, -1, height, width)639            # upsample640            encoder_hidden_state = nn.functional.interpolate(641                encoder_hidden_state, size=encoder_hidden_states[0].size()[2:], mode="bilinear", align_corners=False642            )643            all_hidden_states += (encoder_hidden_state,)644 645        hidden_states = self.linear_fuse(torch.cat(all_hidden_states[::-1], dim=1))646        hidden_states = self.batch_norm(hidden_states)647        hidden_states = self.activation(hidden_states)648        hidden_states = self.dropout(hidden_states)649 650        # logits are of shape (batch_size, num_labels, height/4, width/4)651        logits = self.classifier(hidden_states)652 653        return logits654 655 656@auto_docstring(657    custom_intro="""658    SegFormer Model transformer with an all-MLP decode head on top e.g. for ADE20k, CityScapes.659    """660)661class SegformerForSemanticSegmentation(SegformerPreTrainedModel):662    def __init__(self, config):663        super().__init__(config)664        self.segformer = SegformerModel(config)665        self.decode_head = SegformerDecodeHead(config)666 667        # Initialize weights and apply final processing668        self.post_init()669 670    @auto_docstring671    def forward(672        self,673        pixel_values: torch.FloatTensor,674        labels: Optional[torch.LongTensor] = None,675        output_attentions: Optional[bool] = None,676        output_hidden_states: Optional[bool] = None,677        return_dict: Optional[bool] = None,678    ) -> Union[tuple, SemanticSegmenterOutput]:679        r"""680        labels (`torch.LongTensor` of shape `(batch_size, height, width)`, *optional*):681            Ground truth semantic segmentation maps for computing the loss. Indices should be in `[0, ...,682            config.num_labels - 1]`. If `config.num_labels > 1`, a classification loss is computed (Cross-Entropy).683 684        Examples:685 686        ```python687        >>> from transformers import AutoImageProcessor, SegformerForSemanticSegmentation688        >>> from PIL import Image689        >>> import requests690 691        >>> image_processor = AutoImageProcessor.from_pretrained("nvidia/segformer-b0-finetuned-ade-512-512")692        >>> model = SegformerForSemanticSegmentation.from_pretrained("nvidia/segformer-b0-finetuned-ade-512-512")693 694        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"695        >>> image = Image.open(requests.get(url, stream=True).raw)696 697        >>> inputs = image_processor(images=image, return_tensors="pt")698        >>> outputs = model(**inputs)699        >>> logits = outputs.logits  # shape (batch_size, num_labels, height/4, width/4)700        >>> list(logits.shape)701        [1, 150, 128, 128]702        ```"""703        return_dict = return_dict if return_dict is not None else self.config.use_return_dict704        output_hidden_states = (705            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states706        )707 708        if labels is not None and self.config.num_labels < 1:709            raise ValueError(f"Number of labels should be >=0: {self.config.num_labels}")710 711        outputs = self.segformer(712            pixel_values,713            output_attentions=output_attentions,714            output_hidden_states=True,  # we need the intermediate hidden states715            return_dict=return_dict,716        )717 718        encoder_hidden_states = outputs.hidden_states if return_dict else outputs[1]719 720        logits = self.decode_head(encoder_hidden_states)721 722        loss = None723        if labels is not None:724            # upsample logits to the images' original size725            upsampled_logits = nn.functional.interpolate(726                logits, size=labels.shape[-2:], mode="bilinear", align_corners=False727            )728            if self.config.num_labels > 1:729                loss_fct = CrossEntropyLoss(ignore_index=self.config.semantic_loss_ignore_index)730                loss = loss_fct(upsampled_logits, labels)731            elif self.config.num_labels == 1:732                valid_mask = ((labels >= 0) & (labels != self.config.semantic_loss_ignore_index)).float()733                loss_fct = BCEWithLogitsLoss(reduction="none")734                loss = loss_fct(upsampled_logits.squeeze(1), labels.float())735                loss = (loss * valid_mask).mean()736 737        if not return_dict:738            if output_hidden_states:739                output = (logits,) + outputs[1:]740            else:741                output = (logits,) + outputs[2:]742            return ((loss,) + output) if loss is not None else output743 744        return SemanticSegmenterOutput(745            loss=loss,746            logits=logits,747            hidden_states=outputs.hidden_states if output_hidden_states else None,748            attentions=outputs.attentions,749        )750 751 752__all__ = [753    "SegformerDecodeHead",754    "SegformerForImageClassification",755    "SegformerForSemanticSegmentation",756    "SegformerLayer",757    "SegformerModel",758    "SegformerPreTrainedModel",759]760