Aluode/PerceptionLabPortable

1# coding=utf-82# Copyright 2022 KAIST and 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 GLPN model."""16 17import math18from typing import Optional, Union19 20import torch21from torch import nn22 23from ...activations import ACT2FN24from ...modeling_outputs import BaseModelOutput, DepthEstimatorOutput25from ...modeling_utils import PreTrainedModel26from ...pytorch_utils import find_pruneable_heads_and_indices, prune_linear_layer27from ...utils import auto_docstring, logging28from .configuration_glpn import GLPNConfig29 30 31logger = logging.get_logger(__name__)32 33 34# Copied from transformers.models.beit.modeling_beit.drop_path35def drop_path(input: torch.Tensor, drop_prob: float = 0.0, training: bool = False) -> torch.Tensor:36    """37    Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).38 39    Comment by Ross Wightman: This is the same as the DropConnect impl I created for EfficientNet, etc networks,40    however, the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...41    See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for changing the42    layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use 'survival rate' as the43    argument.44    """45    if drop_prob == 0.0 or not training:46        return input47    keep_prob = 1 - drop_prob48    shape = (input.shape[0],) + (1,) * (input.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets49    random_tensor = keep_prob + torch.rand(shape, dtype=input.dtype, device=input.device)50    random_tensor.floor_()  # binarize51    output = input.div(keep_prob) * random_tensor52    return output53 54 55# Copied from transformers.models.segformer.modeling_segformer.SegformerDropPath56class GLPNDropPath(nn.Module):57    """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks)."""58 59    def __init__(self, drop_prob: Optional[float] = None) -> None:60        super().__init__()61        self.drop_prob = drop_prob62 63    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:64        return drop_path(hidden_states, self.drop_prob, self.training)65 66    def extra_repr(self) -> str:67        return f"p={self.drop_prob}"68 69 70# Copied from transformers.models.segformer.modeling_segformer.SegformerOverlapPatchEmbeddings71class GLPNOverlapPatchEmbeddings(nn.Module):72    """Construct the overlapping patch embeddings."""73 74    def __init__(self, patch_size, stride, num_channels, hidden_size):75        super().__init__()76        self.proj = nn.Conv2d(77            num_channels,78            hidden_size,79            kernel_size=patch_size,80            stride=stride,81            padding=patch_size // 2,82        )83 84        self.layer_norm = nn.LayerNorm(hidden_size)85 86    def forward(self, pixel_values):87        embeddings = self.proj(pixel_values)88        _, _, height, width = embeddings.shape89        # (batch_size, num_channels, height, width) -> (batch_size, num_channels, height*width) -> (batch_size, height*width, num_channels)90        # this can be fed to a Transformer layer91        embeddings = embeddings.flatten(2).transpose(1, 2)92        embeddings = self.layer_norm(embeddings)93        return embeddings, height, width94 95 96# Copied from transformers.models.segformer.modeling_segformer.SegformerEfficientSelfAttention97class GLPNEfficientSelfAttention(nn.Module):98    """SegFormer's efficient self-attention mechanism. Employs the sequence reduction process introduced in the [PvT99    paper](https://huggingface.co/papers/2102.12122)."""100 101    def __init__(self, config, hidden_size, num_attention_heads, sequence_reduction_ratio):102        super().__init__()103        self.hidden_size = hidden_size104        self.num_attention_heads = num_attention_heads105 106        if self.hidden_size % self.num_attention_heads != 0:107            raise ValueError(108                f"The hidden size ({self.hidden_size}) is not a multiple of the number of attention "109                f"heads ({self.num_attention_heads})"110            )111 112        self.attention_head_size = int(self.hidden_size / self.num_attention_heads)113        self.all_head_size = self.num_attention_heads * self.attention_head_size114 115        self.query = nn.Linear(self.hidden_size, self.all_head_size)116        self.key = nn.Linear(self.hidden_size, self.all_head_size)117        self.value = nn.Linear(self.hidden_size, self.all_head_size)118 119        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)120 121        self.sr_ratio = sequence_reduction_ratio122        if sequence_reduction_ratio > 1:123            self.sr = nn.Conv2d(124                hidden_size, hidden_size, kernel_size=sequence_reduction_ratio, stride=sequence_reduction_ratio125            )126            self.layer_norm = nn.LayerNorm(hidden_size)127 128    def forward(129        self,130        hidden_states,131        height,132        width,133        output_attentions=False,134    ):135        batch_size, seq_length, _ = hidden_states.shape136        query_layer = (137            self.query(hidden_states)138            .view(batch_size, -1, self.num_attention_heads, self.attention_head_size)139            .transpose(1, 2)140        )141 142        if self.sr_ratio > 1:143            batch_size, seq_len, num_channels = hidden_states.shape144            # Reshape to (batch_size, num_channels, height, width)145            hidden_states = hidden_states.permute(0, 2, 1).reshape(batch_size, num_channels, height, width)146            # Apply sequence reduction147            hidden_states = self.sr(hidden_states)148            # Reshape back to (batch_size, seq_len, num_channels)149            hidden_states = hidden_states.reshape(batch_size, num_channels, -1).permute(0, 2, 1)150            hidden_states = self.layer_norm(hidden_states)151 152        key_layer = (153            self.key(hidden_states)154            .view(batch_size, -1, self.num_attention_heads, self.attention_head_size)155            .transpose(1, 2)156        )157        value_layer = (158            self.value(hidden_states)159            .view(batch_size, -1, self.num_attention_heads, self.attention_head_size)160            .transpose(1, 2)161        )162 163        # Take the dot product between "query" and "key" to get the raw attention scores.164        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))165 166        attention_scores = attention_scores / math.sqrt(self.attention_head_size)167 168        # Normalize the attention scores to probabilities.169        attention_probs = nn.functional.softmax(attention_scores, dim=-1)170 171        # This is actually dropping out entire tokens to attend to, which might172        # seem a bit unusual, but is taken from the original Transformer paper.173        attention_probs = self.dropout(attention_probs)174 175        context_layer = torch.matmul(attention_probs, value_layer)176 177        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()178        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)179        context_layer = context_layer.view(new_context_layer_shape)180 181        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)182 183        return outputs184 185 186# Copied from transformers.models.segformer.modeling_segformer.SegformerSelfOutput187class GLPNSelfOutput(nn.Module):188    def __init__(self, config, hidden_size):189        super().__init__()190        self.dense = nn.Linear(hidden_size, hidden_size)191        self.dropout = nn.Dropout(config.hidden_dropout_prob)192 193    def forward(self, hidden_states, input_tensor):194        hidden_states = self.dense(hidden_states)195        hidden_states = self.dropout(hidden_states)196        return hidden_states197 198 199# Copied from transformers.models.segformer.modeling_segformer.SegformerAttention with Segformer->GLPN200class GLPNAttention(nn.Module):201    def __init__(self, config, hidden_size, num_attention_heads, sequence_reduction_ratio):202        super().__init__()203        self.self = GLPNEfficientSelfAttention(204            config=config,205            hidden_size=hidden_size,206            num_attention_heads=num_attention_heads,207            sequence_reduction_ratio=sequence_reduction_ratio,208        )209        self.output = GLPNSelfOutput(config, hidden_size=hidden_size)210        self.pruned_heads = set()211 212    def prune_heads(self, heads):213        if len(heads) == 0:214            return215        heads, index = find_pruneable_heads_and_indices(216            heads, self.self.num_attention_heads, self.self.attention_head_size, self.pruned_heads217        )218 219        # Prune linear layers220        self.self.query = prune_linear_layer(self.self.query, index)221        self.self.key = prune_linear_layer(self.self.key, index)222        self.self.value = prune_linear_layer(self.self.value, index)223        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)224 225        # Update hyper params and store pruned heads226        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)227        self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads228        self.pruned_heads = self.pruned_heads.union(heads)229 230    def forward(self, hidden_states, height, width, output_attentions=False):231        self_outputs = self.self(hidden_states, height, width, output_attentions)232 233        attention_output = self.output(self_outputs[0], hidden_states)234        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them235        return outputs236 237 238# Copied from transformers.models.segformer.modeling_segformer.SegformerDWConv239class GLPNDWConv(nn.Module):240    def __init__(self, dim=768):241        super().__init__()242        self.dwconv = nn.Conv2d(dim, dim, 3, 1, 1, bias=True, groups=dim)243 244    def forward(self, hidden_states, height, width):245        batch_size, seq_len, num_channels = hidden_states.shape246        hidden_states = hidden_states.transpose(1, 2).view(batch_size, num_channels, height, width)247        hidden_states = self.dwconv(hidden_states)248        hidden_states = hidden_states.flatten(2).transpose(1, 2)249 250        return hidden_states251 252 253# Copied from transformers.models.segformer.modeling_segformer.SegformerMixFFN with Segformer->GLPN254class GLPNMixFFN(nn.Module):255    def __init__(self, config, in_features, hidden_features=None, out_features=None):256        super().__init__()257        out_features = out_features or in_features258        self.dense1 = nn.Linear(in_features, hidden_features)259        self.dwconv = GLPNDWConv(hidden_features)260        if isinstance(config.hidden_act, str):261            self.intermediate_act_fn = ACT2FN[config.hidden_act]262        else:263            self.intermediate_act_fn = config.hidden_act264        self.dense2 = nn.Linear(hidden_features, out_features)265        self.dropout = nn.Dropout(config.hidden_dropout_prob)266 267    def forward(self, hidden_states, height, width):268        hidden_states = self.dense1(hidden_states)269        hidden_states = self.dwconv(hidden_states, height, width)270        hidden_states = self.intermediate_act_fn(hidden_states)271        hidden_states = self.dropout(hidden_states)272        hidden_states = self.dense2(hidden_states)273        hidden_states = self.dropout(hidden_states)274        return hidden_states275 276 277# Copied from transformers.models.segformer.modeling_segformer.SegformerLayer with Segformer->GLPN278class GLPNLayer(nn.Module):279    """This corresponds to the Block class in the original implementation."""280 281    def __init__(self, config, hidden_size, num_attention_heads, drop_path, sequence_reduction_ratio, mlp_ratio):282        super().__init__()283        self.layer_norm_1 = nn.LayerNorm(hidden_size)284        self.attention = GLPNAttention(285            config,286            hidden_size=hidden_size,287            num_attention_heads=num_attention_heads,288            sequence_reduction_ratio=sequence_reduction_ratio,289        )290        self.drop_path = GLPNDropPath(drop_path) if drop_path > 0.0 else nn.Identity()291        self.layer_norm_2 = nn.LayerNorm(hidden_size)292        mlp_hidden_size = int(hidden_size * mlp_ratio)293        self.mlp = GLPNMixFFN(config, in_features=hidden_size, hidden_features=mlp_hidden_size)294 295    def forward(self, hidden_states, height, width, output_attentions=False):296        self_attention_outputs = self.attention(297            self.layer_norm_1(hidden_states),  # in GLPN, layernorm is applied before self-attention298            height,299            width,300            output_attentions=output_attentions,301        )302 303        attention_output = self_attention_outputs[0]304        outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights305 306        # first residual connection (with stochastic depth)307        attention_output = self.drop_path(attention_output)308        hidden_states = attention_output + hidden_states309 310        mlp_output = self.mlp(self.layer_norm_2(hidden_states), height, width)311 312        # second residual connection (with stochastic depth)313        mlp_output = self.drop_path(mlp_output)314        layer_output = mlp_output + hidden_states315 316        outputs = (layer_output,) + outputs317 318        return outputs319 320 321class GLPNEncoder(nn.Module):322    def __init__(self, config):323        super().__init__()324        self.config = config325 326        # stochastic depth decay rule327        dpr = [x.item() for x in torch.linspace(0, config.drop_path_rate, sum(config.depths), device="cpu")]328 329        # patch embeddings330        embeddings = []331        for i in range(config.num_encoder_blocks):332            embeddings.append(333                GLPNOverlapPatchEmbeddings(334                    patch_size=config.patch_sizes[i],335                    stride=config.strides[i],336                    num_channels=config.num_channels if i == 0 else config.hidden_sizes[i - 1],337                    hidden_size=config.hidden_sizes[i],338                )339            )340        self.patch_embeddings = nn.ModuleList(embeddings)341 342        # Transformer blocks343        blocks = []344        cur = 0345        for i in range(config.num_encoder_blocks):346            # each block consists of layers347            layers = []348            if i != 0:349                cur += config.depths[i - 1]350            for j in range(config.depths[i]):351                layers.append(352                    GLPNLayer(353                        config,354                        hidden_size=config.hidden_sizes[i],355                        num_attention_heads=config.num_attention_heads[i],356                        drop_path=dpr[cur + j],357                        sequence_reduction_ratio=config.sr_ratios[i],358                        mlp_ratio=config.mlp_ratios[i],359                    )360                )361            blocks.append(nn.ModuleList(layers))362 363        self.block = nn.ModuleList(blocks)364 365        # Layer norms366        self.layer_norm = nn.ModuleList(367            [nn.LayerNorm(config.hidden_sizes[i]) for i in range(config.num_encoder_blocks)]368        )369 370    def forward(371        self,372        pixel_values,373        output_attentions=False,374        output_hidden_states=False,375        return_dict=True,376    ):377        all_hidden_states = () if output_hidden_states else None378        all_self_attentions = () if output_attentions else None379 380        batch_size = pixel_values.shape[0]381 382        hidden_states = pixel_values383        for idx, x in enumerate(zip(self.patch_embeddings, self.block, self.layer_norm)):384            embedding_layer, block_layer, norm_layer = x385            # first, obtain patch embeddings386            hidden_states, height, width = embedding_layer(hidden_states)387            # second, send embeddings through blocks388            for i, blk in enumerate(block_layer):389                layer_outputs = blk(hidden_states, height, width, output_attentions)390                hidden_states = layer_outputs[0]391                if output_attentions:392                    all_self_attentions = all_self_attentions + (layer_outputs[1],)393            # third, apply layer norm394            hidden_states = norm_layer(hidden_states)395            # fourth, optionally reshape back to (batch_size, num_channels, height, width)396            hidden_states = hidden_states.reshape(batch_size, height, width, -1).permute(0, 3, 1, 2).contiguous()397            if output_hidden_states:398                all_hidden_states = all_hidden_states + (hidden_states,)399 400        if not return_dict:401            return tuple(v for v in [hidden_states, all_hidden_states, all_self_attentions] if v is not None)402        return BaseModelOutput(403            last_hidden_state=hidden_states,404            hidden_states=all_hidden_states,405            attentions=all_self_attentions,406        )407 408 409@auto_docstring410class GLPNPreTrainedModel(PreTrainedModel):411    config: GLPNConfig412    base_model_prefix = "glpn"413    main_input_name = "pixel_values"414    _no_split_modules = []415 416    # Copied from transformers.models.segformer.modeling_segformer.SegformerPreTrainedModel._init_weights417    def _init_weights(self, module):418        """Initialize the weights"""419        if isinstance(module, (nn.Linear, nn.Conv2d)):420            # Slightly different from the TF version which uses truncated_normal for initialization421            # cf https://github.com/pytorch/pytorch/pull/5617422            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)423            if module.bias is not None:424                module.bias.data.zero_()425        elif isinstance(module, nn.Embedding):426            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)427            if module.padding_idx is not None:428                module.weight.data[module.padding_idx].zero_()429        elif isinstance(module, (nn.LayerNorm, nn.BatchNorm2d)):430            module.bias.data.zero_()431            module.weight.data.fill_(1.0)432 433 434@auto_docstring435class GLPNModel(GLPNPreTrainedModel):436    # Copied from transformers.models.segformer.modeling_segformer.SegformerModel.__init__ with Segformer->GLPN437    def __init__(self, config):438        super().__init__(config)439        self.config = config440 441        # hierarchical Transformer encoder442        self.encoder = GLPNEncoder(config)443 444        # Initialize weights and apply final processing445        self.post_init()446 447    def _prune_heads(self, heads_to_prune):448        """449        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base450        class PreTrainedModel451        """452        for layer, heads in heads_to_prune.items():453            self.encoder.layer[layer].attention.prune_heads(heads)454 455    @auto_docstring456    # Copied from transformers.models.segformer.modeling_segformer.SegformerModel.forward457    def forward(458        self,459        pixel_values: torch.FloatTensor,460        output_attentions: Optional[bool] = None,461        output_hidden_states: Optional[bool] = None,462        return_dict: Optional[bool] = None,463    ) -> Union[tuple, BaseModelOutput]:464        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions465        output_hidden_states = (466            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states467        )468        return_dict = return_dict if return_dict is not None else self.config.use_return_dict469 470        encoder_outputs = self.encoder(471            pixel_values,472            output_attentions=output_attentions,473            output_hidden_states=output_hidden_states,474            return_dict=return_dict,475        )476        sequence_output = encoder_outputs[0]477 478        if not return_dict:479            return (sequence_output,) + encoder_outputs[1:]480 481        return BaseModelOutput(482            last_hidden_state=sequence_output,483            hidden_states=encoder_outputs.hidden_states,484            attentions=encoder_outputs.attentions,485        )486 487 488class GLPNSelectiveFeatureFusion(nn.Module):489    """490    Selective Feature Fusion module, as explained in the [paper](https://huggingface.co/papers/2201.07436) (section 3.4). This491    module adaptively selects and integrates local and global features by attaining an attention map for each feature.492    """493 494    def __init__(self, in_channel=64):495        super().__init__()496 497        self.convolutional_layer1 = nn.Sequential(498            nn.Conv2d(in_channels=int(in_channel * 2), out_channels=in_channel, kernel_size=3, stride=1, padding=1),499            nn.BatchNorm2d(in_channel),500            nn.ReLU(),501        )502 503        self.convolutional_layer2 = nn.Sequential(504            nn.Conv2d(in_channels=in_channel, out_channels=int(in_channel / 2), kernel_size=3, stride=1, padding=1),505            nn.BatchNorm2d(int(in_channel / 2)),506            nn.ReLU(),507        )508 509        self.convolutional_layer3 = nn.Conv2d(510            in_channels=int(in_channel / 2), out_channels=2, kernel_size=3, stride=1, padding=1511        )512 513        self.sigmoid = nn.Sigmoid()514 515    def forward(self, local_features, global_features):516        # concatenate features along the channel dimension517        features = torch.cat((local_features, global_features), dim=1)518        # pass through convolutional layers519        features = self.convolutional_layer1(features)520        features = self.convolutional_layer2(features)521        features = self.convolutional_layer3(features)522        # apply sigmoid to get two-channel attention map523        attn = self.sigmoid(features)524        # construct hybrid features by adding element-wise525        hybrid_features = local_features * attn[:, 0, :, :].unsqueeze(1) + global_features * attn[526            :, 1, :, :527        ].unsqueeze(1)528 529        return hybrid_features530 531 532class GLPNDecoderStage(nn.Module):533    def __init__(self, in_channels, out_channels):534        super().__init__()535        should_skip = in_channels == out_channels536        self.convolution = nn.Conv2d(in_channels, out_channels, kernel_size=1) if not should_skip else nn.Identity()537        self.fusion = GLPNSelectiveFeatureFusion(out_channels)538        self.upsample = nn.Upsample(scale_factor=2, mode="bilinear", align_corners=False)539 540    def forward(self, hidden_state, residual=None):541        hidden_state = self.convolution(hidden_state)542        if residual is not None:543            hidden_state = self.fusion(hidden_state, residual)544        hidden_state = self.upsample(hidden_state)545 546        return hidden_state547 548        hidden_state = self.upsample(hidden_state)549        return hidden_state550 551 552class GLPNDecoder(nn.Module):553    def __init__(self, config):554        super().__init__()555        # we use features from end -> start556        reserved_hidden_sizes = config.hidden_sizes[::-1]557        out_channels = config.decoder_hidden_size558 559        self.stages = nn.ModuleList(560            [GLPNDecoderStage(hidden_size, out_channels) for hidden_size in reserved_hidden_sizes]561        )562        # don't fuse in first stage563        self.stages[0].fusion = None564 565        self.final_upsample = nn.Upsample(scale_factor=2, mode="bilinear", align_corners=False)566 567    def forward(self, hidden_states: list[torch.Tensor]) -> list[torch.Tensor]:568        stage_hidden_states = []569        stage_hidden_state = None570        for hidden_state, stage in zip(hidden_states[::-1], self.stages):571            stage_hidden_state = stage(hidden_state, stage_hidden_state)572            stage_hidden_states.append(stage_hidden_state)573 574        stage_hidden_states[-1] = self.final_upsample(stage_hidden_state)575 576        return stage_hidden_states577 578 579class SiLogLoss(nn.Module):580    r"""581    Implements the Scale-invariant log scale loss [Eigen et al., 2014](https://huggingface.co/papers/1406.2283).582 583    $$L=\frac{1}{n} \sum_{i} d_{i}^{2}-\frac{1}{2 n^{2}}\left(\sum_{i} d_{i}^{2}\right)$$ where $d_{i}=\log y_{i}-\log584    y_{i}^{*}$.585 586    """587 588    def __init__(self, lambd=0.5):589        super().__init__()590        self.lambd = lambd591 592    def forward(self, pred, target):593        valid_mask = (target > 0).detach()594        diff_log = torch.log(target[valid_mask]) - torch.log(pred[valid_mask])595        loss = torch.sqrt(torch.pow(diff_log, 2).mean() - self.lambd * torch.pow(diff_log.mean(), 2))596 597        return loss598 599 600class GLPNDepthEstimationHead(nn.Module):601    def __init__(self, config):602        super().__init__()603 604        self.config = config605 606        channels = config.decoder_hidden_size607        self.head = nn.Sequential(608            nn.Conv2d(channels, channels, kernel_size=3, stride=1, padding=1),609            nn.ReLU(inplace=False),610            nn.Conv2d(channels, 1, kernel_size=3, stride=1, padding=1),611        )612 613    def forward(self, hidden_states: list[torch.Tensor]) -> torch.Tensor:614        # use last features of the decoder615        hidden_states = hidden_states[self.config.head_in_index]616 617        hidden_states = self.head(hidden_states)618 619        predicted_depth = torch.sigmoid(hidden_states) * self.config.max_depth620        predicted_depth = predicted_depth.squeeze(dim=1)621 622        return predicted_depth623 624 625@auto_docstring(626    custom_intro="""627    GLPN Model transformer with a lightweight depth estimation head on top e.g. for KITTI, NYUv2.628    """629)630class GLPNForDepthEstimation(GLPNPreTrainedModel):631    def __init__(self, config):632        super().__init__(config)633 634        self.glpn = GLPNModel(config)635        self.decoder = GLPNDecoder(config)636        self.head = GLPNDepthEstimationHead(config)637 638        # Initialize weights and apply final processing639        self.post_init()640 641    @auto_docstring642    def forward(643        self,644        pixel_values: torch.FloatTensor,645        labels: Optional[torch.FloatTensor] = None,646        output_attentions: Optional[bool] = None,647        output_hidden_states: Optional[bool] = None,648        return_dict: Optional[bool] = None,649    ) -> Union[tuple[torch.Tensor], DepthEstimatorOutput]:650        r"""651        labels (`torch.FloatTensor` of shape `(batch_size, height, width)`, *optional*):652            Ground truth depth estimation maps for computing the loss.653 654        Examples:655 656        ```python657        >>> from transformers import AutoImageProcessor, GLPNForDepthEstimation658        >>> import torch659        >>> import numpy as np660        >>> from PIL import Image661        >>> import requests662 663        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"664        >>> image = Image.open(requests.get(url, stream=True).raw)665 666        >>> image_processor = AutoImageProcessor.from_pretrained("vinvino02/glpn-kitti")667        >>> model = GLPNForDepthEstimation.from_pretrained("vinvino02/glpn-kitti")668 669        >>> # prepare image for the model670        >>> inputs = image_processor(images=image, return_tensors="pt")671 672        >>> with torch.no_grad():673        ...     outputs = model(**inputs)674 675        >>> # interpolate to original size676        >>> post_processed_output = image_processor.post_process_depth_estimation(677        ...     outputs,678        ...     target_sizes=[(image.height, image.width)],679        ... )680 681        >>> # visualize the prediction682        >>> predicted_depth = post_processed_output[0]["predicted_depth"]683        >>> depth = predicted_depth * 255 / predicted_depth.max()684        >>> depth = depth.detach().cpu().numpy()685        >>> depth = Image.fromarray(depth.astype("uint8"))686        ```"""687        return_dict = return_dict if return_dict is not None else self.config.use_return_dict688        output_hidden_states = (689            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states690        )691 692        outputs = self.glpn(693            pixel_values,694            output_attentions=output_attentions,695            output_hidden_states=True,  # we need the intermediate hidden states696            return_dict=return_dict,697        )698 699        hidden_states = outputs.hidden_states if return_dict else outputs[1]700 701        out = self.decoder(hidden_states)702        predicted_depth = self.head(out)703 704        loss = None705        if labels is not None:706            loss_fct = SiLogLoss()707            loss = loss_fct(predicted_depth, labels)708 709        if not return_dict:710            if output_hidden_states:711                output = (predicted_depth,) + outputs[1:]712            else:713                output = (predicted_depth,) + outputs[2:]714            return ((loss,) + output) if loss is not None else output715 716        return DepthEstimatorOutput(717            loss=loss,718            predicted_depth=predicted_depth,719            hidden_states=outputs.hidden_states if output_hidden_states else None,720            attentions=outputs.attentions,721        )722 723 724__all__ = ["GLPNForDepthEstimation", "GLPNLayer", "GLPNModel", "GLPNPreTrainedModel"]725