modeling_van.py

# coding=utf-8
# Copyright 2022 BNRist (Tsinghua University), TKLNDST (Nankai University) and 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
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" PyTorch Visual Attention Network (VAN) model."""

import math
from collections import OrderedDict
from typing import Optional, Tuple, Union

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

from ...activations import ACT2FN
from ...modeling_outputs import (
    BaseModelOutputWithNoAttention,
    BaseModelOutputWithPoolingAndNoAttention,
    ImageClassifierOutputWithNoAttention,
)
from ...modeling_utils import PreTrainedModel
from ...utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward, logging
from .configuration_van import VanConfig


logger = logging.get_logger(__name__)

# General docstring
_CONFIG_FOR_DOC = "VanConfig"

# Base docstring
_CHECKPOINT_FOR_DOC = "Visual-Attention-Network/van-base"
_EXPECTED_OUTPUT_SHAPE = [1, 512, 7, 7]

# Image classification docstring
_IMAGE_CLASS_CHECKPOINT = "Visual-Attention-Network/van-base"
_IMAGE_CLASS_EXPECTED_OUTPUT = "tabby, tabby cat"

VAN_PRETRAINED_MODEL_ARCHIVE_LIST = [
    "Visual-Attention-Network/van-base",
    # See all VAN models at https://huggingface.co/models?filter=van
]


# Copied from transformers.models.convnext.modeling_convnext.drop_path
def drop_path(input, drop_prob: float = 0.0, training: bool = False):
    """
    Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).

    Comment by Ross Wightman: This is the same as the DropConnect impl I created for EfficientNet, etc networks,
    however, the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
    See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for changing the
    layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use 'survival rate' as the
    argument.
    """
    if drop_prob == 0.0 or not training:
        return input
    keep_prob = 1 - drop_prob
    shape = (input.shape[0],) + (1,) * (input.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
    random_tensor = keep_prob + torch.rand(shape, dtype=input.dtype, device=input.device)
    random_tensor.floor_()  # binarize
    output = input.div(keep_prob) * random_tensor
    return output


# Copied from transformers.models.convnext.modeling_convnext.ConvNextDropPath with ConvNext->Van
class VanDropPath(nn.Module):
    """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks)."""

    def __init__(self, drop_prob: Optional[float] = None) -> None:
        super().__init__()
        self.drop_prob = drop_prob

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        return drop_path(hidden_states, self.drop_prob, self.training)

    def extra_repr(self) -> str:
        return "p={}".format(self.drop_prob)


class VanOverlappingPatchEmbedder(nn.Module):
    """
    Downsamples the input using a patchify operation with a `stride` of 4 by default making adjacent windows overlap by
    half of the area. From [PVTv2: Improved Baselines with Pyramid Vision
    Transformer](https://arxiv.org/abs/2106.13797).
    """

    def __init__(self, in_channels: int, hidden_size: int, patch_size: int = 7, stride: int = 4):
        super().__init__()
        self.convolution = nn.Conv2d(
            in_channels, hidden_size, kernel_size=patch_size, stride=stride, padding=patch_size // 2
        )
        self.normalization = nn.BatchNorm2d(hidden_size)

    def forward(self, input: torch.Tensor) -> torch.Tensor:
        hidden_state = self.convolution(input)
        hidden_state = self.normalization(hidden_state)
        return hidden_state


class VanMlpLayer(nn.Module):
    """
    MLP with depth-wise convolution, from [PVTv2: Improved Baselines with Pyramid Vision
    Transformer](https://arxiv.org/abs/2106.13797).
    """

    def __init__(
        self,
        in_channels: int,
        hidden_size: int,
        out_channels: int,
        hidden_act: str = "gelu",
        dropout_rate: float = 0.5,
    ):
        super().__init__()
        self.in_dense = nn.Conv2d(in_channels, hidden_size, kernel_size=1)
        self.depth_wise = nn.Conv2d(hidden_size, hidden_size, kernel_size=3, padding=1, groups=hidden_size)
        self.activation = ACT2FN[hidden_act]
        self.dropout1 = nn.Dropout(dropout_rate)
        self.out_dense = nn.Conv2d(hidden_size, out_channels, kernel_size=1)
        self.dropout2 = nn.Dropout(dropout_rate)

    def forward(self, hidden_state: torch.Tensor) -> torch.Tensor:
        hidden_state = self.in_dense(hidden_state)
        hidden_state = self.depth_wise(hidden_state)
        hidden_state = self.activation(hidden_state)
        hidden_state = self.dropout1(hidden_state)
        hidden_state = self.out_dense(hidden_state)
        hidden_state = self.dropout2(hidden_state)
        return hidden_state


class VanLargeKernelAttention(nn.Module):
    """
    Basic Large Kernel Attention (LKA).
    """

    def __init__(self, hidden_size: int):
        super().__init__()
        self.depth_wise = nn.Conv2d(hidden_size, hidden_size, kernel_size=5, padding=2, groups=hidden_size)
        self.depth_wise_dilated = nn.Conv2d(
            hidden_size, hidden_size, kernel_size=7, dilation=3, padding=9, groups=hidden_size
        )
        self.point_wise = nn.Conv2d(hidden_size, hidden_size, kernel_size=1)

    def forward(self, hidden_state: torch.Tensor) -> torch.Tensor:
        hidden_state = self.depth_wise(hidden_state)
        hidden_state = self.depth_wise_dilated(hidden_state)
        hidden_state = self.point_wise(hidden_state)
        return hidden_state


class VanLargeKernelAttentionLayer(nn.Module):
    """
    Computes attention using Large Kernel Attention (LKA) and attends the input.
    """

    def __init__(self, hidden_size: int):
        super().__init__()
        self.attention = VanLargeKernelAttention(hidden_size)

    def forward(self, hidden_state: torch.Tensor) -> torch.Tensor:
        attention = self.attention(hidden_state)
        attended = hidden_state * attention
        return attended


class VanSpatialAttentionLayer(nn.Module):
    """
    Van spatial attention layer composed by projection (via conv) -> act -> Large Kernel Attention (LKA) attention ->
    projection (via conv) + residual connection.
    """

    def __init__(self, hidden_size: int, hidden_act: str = "gelu"):
        super().__init__()
        self.pre_projection = nn.Sequential(
            OrderedDict(
                [
                    ("conv", nn.Conv2d(hidden_size, hidden_size, kernel_size=1)),
                    ("act", ACT2FN[hidden_act]),
                ]
            )
        )
        self.attention_layer = VanLargeKernelAttentionLayer(hidden_size)
        self.post_projection = nn.Conv2d(hidden_size, hidden_size, kernel_size=1)

    def forward(self, hidden_state: torch.Tensor) -> torch.Tensor:
        residual = hidden_state
        hidden_state = self.pre_projection(hidden_state)
        hidden_state = self.attention_layer(hidden_state)
        hidden_state = self.post_projection(hidden_state)
        hidden_state = hidden_state + residual
        return hidden_state


class VanLayerScaling(nn.Module):
    """
    Scales the inputs by a learnable parameter initialized by `initial_value`.
    """

    def __init__(self, hidden_size: int, initial_value: float = 1e-2):
        super().__init__()
        self.weight = nn.Parameter(initial_value * torch.ones((hidden_size)), requires_grad=True)

    def forward(self, hidden_state: torch.Tensor) -> torch.Tensor:
        # unsqueezing for broadcasting
        hidden_state = self.weight.unsqueeze(-1).unsqueeze(-1) * hidden_state
        return hidden_state


class VanLayer(nn.Module):
    """
    Van layer composed by normalization layers, large kernel attention (LKA) and a multi layer perceptron (MLP).
    """

    def __init__(
        self,
        config: VanConfig,
        hidden_size: int,
        mlp_ratio: int = 4,
        drop_path_rate: float = 0.5,
    ):
        super().__init__()
        self.drop_path = VanDropPath(drop_path_rate) if drop_path_rate > 0.0 else nn.Identity()
        self.pre_normomalization = nn.BatchNorm2d(hidden_size)
        self.attention = VanSpatialAttentionLayer(hidden_size, config.hidden_act)
        self.attention_scaling = VanLayerScaling(hidden_size, config.layer_scale_init_value)
        self.post_normalization = nn.BatchNorm2d(hidden_size)
        self.mlp = VanMlpLayer(
            hidden_size, hidden_size * mlp_ratio, hidden_size, config.hidden_act, config.dropout_rate
        )
        self.mlp_scaling = VanLayerScaling(hidden_size, config.layer_scale_init_value)

    def forward(self, hidden_state: torch.Tensor) -> torch.Tensor:
        residual = hidden_state
        # attention
        hidden_state = self.pre_normomalization(hidden_state)
        hidden_state = self.attention(hidden_state)
        hidden_state = self.attention_scaling(hidden_state)
        hidden_state = self.drop_path(hidden_state)
        # residual connection
        hidden_state = residual + hidden_state
        residual = hidden_state
        # mlp
        hidden_state = self.post_normalization(hidden_state)
        hidden_state = self.mlp(hidden_state)
        hidden_state = self.mlp_scaling(hidden_state)
        hidden_state = self.drop_path(hidden_state)
        # residual connection
        hidden_state = residual + hidden_state
        return hidden_state


class VanStage(nn.Module):
    """
    VanStage, consisting of multiple layers.
    """

    def __init__(
        self,
        config: VanConfig,
        in_channels: int,
        hidden_size: int,
        patch_size: int,
        stride: int,
        depth: int,
        mlp_ratio: int = 4,
        drop_path_rate: float = 0.0,
    ):
        super().__init__()
        self.embeddings = VanOverlappingPatchEmbedder(in_channels, hidden_size, patch_size, stride)
        self.layers = nn.Sequential(
            *[
                VanLayer(
                    config,
                    hidden_size,
                    mlp_ratio=mlp_ratio,
                    drop_path_rate=drop_path_rate,
                )
                for _ in range(depth)
            ]
        )
        self.normalization = nn.LayerNorm(hidden_size, eps=config.layer_norm_eps)

    def forward(self, hidden_state: torch.Tensor) -> torch.Tensor:
        hidden_state = self.embeddings(hidden_state)
        hidden_state = self.layers(hidden_state)
        # rearrange b c h w -> b (h w) c
        batch_size, hidden_size, height, width = hidden_state.shape
        hidden_state = hidden_state.flatten(2).transpose(1, 2)
        hidden_state = self.normalization(hidden_state)
        # rearrange  b (h w) c- > b c h w
        hidden_state = hidden_state.view(batch_size, height, width, hidden_size).permute(0, 3, 1, 2)
        return hidden_state


class VanEncoder(nn.Module):
    """
    VanEncoder, consisting of multiple stages.
    """

    def __init__(self, config: VanConfig):
        super().__init__()
        self.stages = nn.ModuleList([])
        patch_sizes = config.patch_sizes
        strides = config.strides
        hidden_sizes = config.hidden_sizes
        depths = config.depths
        mlp_ratios = config.mlp_ratios
        drop_path_rates = [x.item() for x in torch.linspace(0, config.drop_path_rate, sum(config.depths))]

        for num_stage, (patch_size, stride, hidden_size, depth, mlp_expantion, drop_path_rate) in enumerate(
            zip(patch_sizes, strides, hidden_sizes, depths, mlp_ratios, drop_path_rates)
        ):
            is_first_stage = num_stage == 0
            in_channels = hidden_sizes[num_stage - 1]
            if is_first_stage:
                in_channels = config.num_channels
            self.stages.append(
                VanStage(
                    config,
                    in_channels,
                    hidden_size,
                    patch_size=patch_size,
                    stride=stride,
                    depth=depth,
                    mlp_ratio=mlp_expantion,
                    drop_path_rate=drop_path_rate,
                )
            )

    def forward(
        self,
        hidden_state: torch.Tensor,
        output_hidden_states: Optional[bool] = False,
        return_dict: Optional[bool] = True,
    ) -> Union[Tuple, BaseModelOutputWithNoAttention]:
        all_hidden_states = () if output_hidden_states else None

        for _, stage_module in enumerate(self.stages):
            hidden_state = stage_module(hidden_state)

            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_state,)

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

        return BaseModelOutputWithNoAttention(last_hidden_state=hidden_state, hidden_states=all_hidden_states)


class VanPreTrainedModel(PreTrainedModel):
    """
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
    models.
    """

    config_class = VanConfig
    base_model_prefix = "van"
    main_input_name = "pixel_values"
    supports_gradient_checkpointing = True

    def _init_weights(self, module):
        """Initialize the weights"""
        if isinstance(module, nn.Linear):
            nn.init.trunc_normal_(module.weight, std=self.config.initializer_range)
            if isinstance(module, nn.Linear) and module.bias is not None:
                nn.init.constant_(module.bias, 0)
        elif isinstance(module, nn.LayerNorm):
            nn.init.constant_(module.bias, 0)
            nn.init.constant_(module.weight, 1.0)
        elif isinstance(module, nn.Conv2d):
            fan_out = module.kernel_size[0] * module.kernel_size[1] * module.out_channels
            fan_out //= module.groups
            module.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
            if module.bias is not None:
                module.bias.data.zero_()

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


VAN_START_DOCSTRING = r"""
    This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#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
    behavior.

    Parameters:
        config ([`VanConfig`]): 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 [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""

VAN_INPUTS_DOCSTRING = r"""
    Args:
        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
            Pixel values. Pixel values can be obtained using [`AutoImageProcessor`]. See
            [`ConvNextImageProcessor.__call__`] for details.

        output_hidden_states (`bool`, *optional*):
            Whether or not to return the hidden states of all stages. See `hidden_states` under returned tensors for
            more detail.
        return_dict (`bool`, *optional*):
            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""


@add_start_docstrings(
    "The bare VAN model outputting raw features without any specific head on top. Note, VAN does not have an embedding"
    " layer.",
    VAN_START_DOCSTRING,
)
class VanModel(VanPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.config = config
        self.encoder = VanEncoder(config)
        # final layernorm layer
        self.layernorm = nn.LayerNorm(config.hidden_sizes[-1], eps=config.layer_norm_eps)
        # Initialize weights and apply final processing
        self.post_init()

    @add_start_docstrings_to_model_forward(VAN_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(
        checkpoint=_CHECKPOINT_FOR_DOC,
        output_type=BaseModelOutputWithPoolingAndNoAttention,
        config_class=_CONFIG_FOR_DOC,
        modality="vision",
        expected_output=_EXPECTED_OUTPUT_SHAPE,
    )
    def forward(
        self,
        pixel_values: Optional[torch.FloatTensor],
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, BaseModelOutputWithPoolingAndNoAttention]:
        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

        encoder_outputs = self.encoder(
            pixel_values,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        last_hidden_state = encoder_outputs[0]
        # global average pooling, n c w h -> n c
        pooled_output = last_hidden_state.mean(dim=[-2, -1])

        if not return_dict:
            return (last_hidden_state, pooled_output) + encoder_outputs[1:]

        return BaseModelOutputWithPoolingAndNoAttention(
            last_hidden_state=last_hidden_state,
            pooler_output=pooled_output,
            hidden_states=encoder_outputs.hidden_states,
        )


@add_start_docstrings(
    """
    VAN Model with an image classification head on top (a linear layer on top of the pooled features), e.g. for
    ImageNet.
    """,
    VAN_START_DOCSTRING,
)
class VanForImageClassification(VanPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.van = VanModel(config)
        # Classifier head
        self.classifier = (
            nn.Linear(config.hidden_sizes[-1], config.num_labels) if config.num_labels > 0 else nn.Identity()
        )

        # Initialize weights and apply final processing
        self.post_init()

    @add_start_docstrings_to_model_forward(VAN_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(
        checkpoint=_IMAGE_CLASS_CHECKPOINT,
        output_type=ImageClassifierOutputWithNoAttention,
        config_class=_CONFIG_FOR_DOC,
        expected_output=_IMAGE_CLASS_EXPECTED_OUTPUT,
    )
    def forward(
        self,
        pixel_values: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, ImageClassifierOutputWithNoAttention]:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the image classification/regression loss. Indices should be in `[0, ...,
            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        outputs = self.van(pixel_values, output_hidden_states=output_hidden_states, return_dict=return_dict)

        pooled_output = outputs.pooler_output if return_dict else outputs[1]

        logits = self.classifier(pooled_output)

        loss = None
        if labels is not None:
            if self.config.problem_type is None:
                if self.config.num_labels == 1:
                    self.config.problem_type = "regression"
                elif self.config.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
                    self.config.problem_type = "single_label_classification"
                else:
                    self.config.problem_type = "multi_label_classification"

            if self.config.problem_type == "regression":
                loss_fct = MSELoss()
                if self.config.num_labels == 1:
                    loss = loss_fct(logits.squeeze(), labels.squeeze())
                else:
                    loss = loss_fct(logits, labels)
            elif self.config.problem_type == "single_label_classification":
                loss_fct = CrossEntropyLoss()
                loss = loss_fct(logits.view(-1, self.config.num_labels), labels.view(-1))
            elif self.config.problem_type == "multi_label_classification":
                loss_fct = BCEWithLogitsLoss()
                loss = loss_fct(logits, labels)

        if not return_dict:
            output = (logits,) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

        return ImageClassifierOutputWithNoAttention(loss=loss, logits=logits, hidden_states=outputs.hidden_states)