vision_transformer.py

import math
from collections import OrderedDict
from functools import partial
from typing import Any, Callable, List, NamedTuple, Optional, Dict

import torch
import torch.nn as nn

from ..ops.misc import Conv2dNormActivation
from ..transforms._presets import ImageClassification, InterpolationMode
from ..utils import _log_api_usage_once
from ._api import WeightsEnum, Weights
from ._meta import _IMAGENET_CATEGORIES
from ._utils import handle_legacy_interface, _ovewrite_named_param


__all__ = [
    "VisionTransformer",
    "ViT_B_16_Weights",
    "ViT_B_32_Weights",
    "ViT_L_16_Weights",
    "ViT_L_32_Weights",
    "ViT_H_14_Weights",
    "vit_b_16",
    "vit_b_32",
    "vit_l_16",
    "vit_l_32",
    "vit_h_14",
]


class ConvStemConfig(NamedTuple):
    out_channels: int
    kernel_size: int
    stride: int
    norm_layer: Callable[..., nn.Module] = nn.BatchNorm2d
    activation_layer: Callable[..., nn.Module] = nn.ReLU


class MLPBlock(nn.Sequential):
    """Transformer MLP block."""

    def __init__(self, in_dim: int, mlp_dim: int, dropout: float):
        super().__init__()
        self.linear_1 = nn.Linear(in_dim, mlp_dim)
        self.act = nn.GELU()
        self.dropout_1 = nn.Dropout(dropout)
        self.linear_2 = nn.Linear(mlp_dim, in_dim)
        self.dropout_2 = nn.Dropout(dropout)

        nn.init.xavier_uniform_(self.linear_1.weight)
        nn.init.xavier_uniform_(self.linear_2.weight)
        nn.init.normal_(self.linear_1.bias, std=1e-6)
        nn.init.normal_(self.linear_2.bias, std=1e-6)


class EncoderBlock(nn.Module):
    """Transformer encoder block."""

    def __init__(
        self,
        num_heads: int,
        hidden_dim: int,
        mlp_dim: int,
        dropout: float,
        attention_dropout: float,
        norm_layer: Callable[..., torch.nn.Module] = partial(nn.LayerNorm, eps=1e-6),
    ):
        super().__init__()
        self.num_heads = num_heads

        # Attention block
        self.ln_1 = norm_layer(hidden_dim)
        self.self_attention = nn.MultiheadAttention(hidden_dim, num_heads, dropout=attention_dropout, batch_first=True)
        self.dropout = nn.Dropout(dropout)

        # MLP block
        self.ln_2 = norm_layer(hidden_dim)
        self.mlp = MLPBlock(hidden_dim, mlp_dim, dropout)

    def forward(self, input: torch.Tensor):
        torch._assert(input.dim() == 3, f"Expected (batch_size, seq_length, hidden_dim) got {input.shape}")
        x = self.ln_1(input)
        x, _ = self.self_attention(query=x, key=x, value=x, need_weights=False)
        x = self.dropout(x)
        x = x + input

        y = self.ln_2(x)
        y = self.mlp(y)
        return x + y


class Encoder(nn.Module):
    """Transformer Model Encoder for sequence to sequence translation."""

    def __init__(
        self,
        seq_length: int,
        num_layers: int,
        num_heads: int,
        hidden_dim: int,
        mlp_dim: int,
        dropout: float,
        attention_dropout: float,
        norm_layer: Callable[..., torch.nn.Module] = partial(nn.LayerNorm, eps=1e-6),
    ):
        super().__init__()
        # Note that batch_size is on the first dim because
        # we have batch_first=True in nn.MultiAttention() by default
        self.pos_embedding = nn.Parameter(torch.empty(1, seq_length, hidden_dim).normal_(std=0.02))  # from BERT
        self.dropout = nn.Dropout(dropout)
        layers: OrderedDict[str, nn.Module] = OrderedDict()
        for i in range(num_layers):
            layers[f"encoder_layer_{i}"] = EncoderBlock(
                num_heads,
                hidden_dim,
                mlp_dim,
                dropout,
                attention_dropout,
                norm_layer,
            )
        self.layers = nn.Sequential(layers)
        self.ln = norm_layer(hidden_dim)

    def forward(self, input: torch.Tensor):
        torch._assert(input.dim() == 3, f"Expected (batch_size, seq_length, hidden_dim) got {input.shape}")
        input = input + self.pos_embedding
        return self.ln(self.layers(self.dropout(input)))


class VisionTransformer(nn.Module):
    """Vision Transformer as per https://arxiv.org/abs/2010.11929."""

    def __init__(
        self,
        image_size: int,
        patch_size: int,
        num_layers: int,
        num_heads: int,
        hidden_dim: int,
        mlp_dim: int,
        dropout: float = 0.0,
        attention_dropout: float = 0.0,
        num_classes: int = 1000,
        representation_size: Optional[int] = None,
        norm_layer: Callable[..., torch.nn.Module] = partial(nn.LayerNorm, eps=1e-6),
        conv_stem_configs: Optional[List[ConvStemConfig]] = None,
    ):
        super().__init__()
        _log_api_usage_once(self)
        torch._assert(image_size % patch_size == 0, "Input shape indivisible by patch size!")
        self.image_size = image_size
        self.patch_size = patch_size
        self.hidden_dim = hidden_dim
        self.mlp_dim = mlp_dim
        self.attention_dropout = attention_dropout
        self.dropout = dropout
        self.num_classes = num_classes
        self.representation_size = representation_size
        self.norm_layer = norm_layer

        if conv_stem_configs is not None:
            # As per https://arxiv.org/abs/2106.14881
            seq_proj = nn.Sequential()
            prev_channels = 3
            for i, conv_stem_layer_config in enumerate(conv_stem_configs):
                seq_proj.add_module(
                    f"conv_bn_relu_{i}",
                    Conv2dNormActivation(
                        in_channels=prev_channels,
                        out_channels=conv_stem_layer_config.out_channels,
                        kernel_size=conv_stem_layer_config.kernel_size,
                        stride=conv_stem_layer_config.stride,
                        norm_layer=conv_stem_layer_config.norm_layer,
                        activation_layer=conv_stem_layer_config.activation_layer,
                    ),
                )
                prev_channels = conv_stem_layer_config.out_channels
            seq_proj.add_module(
                "conv_last", nn.Conv2d(in_channels=prev_channels, out_channels=hidden_dim, kernel_size=1)
            )
            self.conv_proj: nn.Module = seq_proj
        else:
            self.conv_proj = nn.Conv2d(
                in_channels=3, out_channels=hidden_dim, kernel_size=patch_size, stride=patch_size
            )

        seq_length = (image_size // patch_size) ** 2

        # Add a class token
        self.class_token = nn.Parameter(torch.zeros(1, 1, hidden_dim))
        seq_length += 1

        self.encoder = Encoder(
            seq_length,
            num_layers,
            num_heads,
            hidden_dim,
            mlp_dim,
            dropout,
            attention_dropout,
            norm_layer,
        )
        self.seq_length = seq_length

        heads_layers: OrderedDict[str, nn.Module] = OrderedDict()
        if representation_size is None:
            heads_layers["head"] = nn.Linear(hidden_dim, num_classes)
        else:
            heads_layers["pre_logits"] = nn.Linear(hidden_dim, representation_size)
            heads_layers["act"] = nn.Tanh()
            heads_layers["head"] = nn.Linear(representation_size, num_classes)

        self.heads = nn.Sequential(heads_layers)

        if isinstance(self.conv_proj, nn.Conv2d):
            # Init the patchify stem
            fan_in = self.conv_proj.in_channels * self.conv_proj.kernel_size[0] * self.conv_proj.kernel_size[1]
            nn.init.trunc_normal_(self.conv_proj.weight, std=math.sqrt(1 / fan_in))
            if self.conv_proj.bias is not None:
                nn.init.zeros_(self.conv_proj.bias)
        elif self.conv_proj.conv_last is not None and isinstance(self.conv_proj.conv_last, nn.Conv2d):
            # Init the last 1x1 conv of the conv stem
            nn.init.normal_(
                self.conv_proj.conv_last.weight, mean=0.0, std=math.sqrt(2.0 / self.conv_proj.conv_last.out_channels)
            )
            if self.conv_proj.conv_last.bias is not None:
                nn.init.zeros_(self.conv_proj.conv_last.bias)

        if hasattr(self.heads, "pre_logits") and isinstance(self.heads.pre_logits, nn.Linear):
            fan_in = self.heads.pre_logits.in_features
            nn.init.trunc_normal_(self.heads.pre_logits.weight, std=math.sqrt(1 / fan_in))
            nn.init.zeros_(self.heads.pre_logits.bias)

        if isinstance(self.heads.head, nn.Linear):
            nn.init.zeros_(self.heads.head.weight)
            nn.init.zeros_(self.heads.head.bias)

    def _process_input(self, x: torch.Tensor) -> torch.Tensor:
        n, c, h, w = x.shape
        p = self.patch_size
        torch._assert(h == self.image_size, "Wrong image height!")
        torch._assert(w == self.image_size, "Wrong image width!")
        n_h = h // p
        n_w = w // p

        # (n, c, h, w) -> (n, hidden_dim, n_h, n_w)
        x = self.conv_proj(x)
        # (n, hidden_dim, n_h, n_w) -> (n, hidden_dim, (n_h * n_w))
        x = x.reshape(n, self.hidden_dim, n_h * n_w)

        # (n, hidden_dim, (n_h * n_w)) -> (n, (n_h * n_w), hidden_dim)
        # The self attention layer expects inputs in the format (N, S, E)
        # where S is the source sequence length, N is the batch size, E is the
        # embedding dimension
        x = x.permute(0, 2, 1)

        return x

    def forward(self, x: torch.Tensor):
        # Reshape and permute the input tensor
        x = self._process_input(x)
        n = x.shape[0]

        # Expand the class token to the full batch
        batch_class_token = self.class_token.expand(n, -1, -1)
        x = torch.cat([batch_class_token, x], dim=1)

        x = self.encoder(x)

        # Classifier "token" as used by standard language architectures
        x = x[:, 0]

        x = self.heads(x)

        return x


def _vision_transformer(
    patch_size: int,
    num_layers: int,
    num_heads: int,
    hidden_dim: int,
    mlp_dim: int,
    weights: Optional[WeightsEnum],
    progress: bool,
    **kwargs: Any,
) -> VisionTransformer:
    if weights is not None:
        _ovewrite_named_param(kwargs, "num_classes", len(weights.meta["categories"]))
        assert weights.meta["min_size"][0] == weights.meta["min_size"][1]
        _ovewrite_named_param(kwargs, "image_size", weights.meta["min_size"][0])
    image_size = kwargs.pop("image_size", 224)

    model = VisionTransformer(
        image_size=image_size,
        patch_size=patch_size,
        num_layers=num_layers,
        num_heads=num_heads,
        hidden_dim=hidden_dim,
        mlp_dim=mlp_dim,
        **kwargs,
    )

    if weights:
        model.load_state_dict(weights.get_state_dict(progress=progress))

    return model


_COMMON_META: Dict[str, Any] = {
    "categories": _IMAGENET_CATEGORIES,
}

_COMMON_SWAG_META = {
    **_COMMON_META,
    "recipe": "https://github.com/facebookresearch/SWAG",
    "license": "https://github.com/facebookresearch/SWAG/blob/main/LICENSE",
}


class ViT_B_16_Weights(WeightsEnum):
    IMAGENET1K_V1 = Weights(
        url="https://download.pytorch.org/models/vit_b_16-c867db91.pth",
        transforms=partial(ImageClassification, crop_size=224),
        meta={
            **_COMMON_META,
            "num_params": 86567656,
            "min_size": (224, 224),
            "recipe": "https://github.com/pytorch/vision/tree/main/references/classification#vit_b_16",
            "acc@1": 81.072,
            "acc@5": 95.318,
        },
    )
    IMAGENET1K_SWAG_E2E_V1 = Weights(
        url="https://download.pytorch.org/models/vit_b_16_swag-9ac1b537.pth",
        transforms=partial(
            ImageClassification,
            crop_size=384,
            resize_size=384,
            interpolation=InterpolationMode.BICUBIC,
        ),
        meta={
            **_COMMON_SWAG_META,
            "num_params": 86859496,
            "min_size": (384, 384),
            "acc@1": 85.304,
            "acc@5": 97.650,
        },
    )
    IMAGENET1K_SWAG_LINEAR_V1 = Weights(
        url="https://download.pytorch.org/models/vit_b_16_lc_swag-4e70ced5.pth",
        transforms=partial(
            ImageClassification,
            crop_size=224,
            resize_size=224,
            interpolation=InterpolationMode.BICUBIC,
        ),
        meta={
            **_COMMON_SWAG_META,
            "recipe": "https://github.com/pytorch/vision/pull/5793",
            "num_params": 86567656,
            "min_size": (224, 224),
            "acc@1": 81.886,
            "acc@5": 96.180,
        },
    )
    DEFAULT = IMAGENET1K_V1


class ViT_B_32_Weights(WeightsEnum):
    IMAGENET1K_V1 = Weights(
        url="https://download.pytorch.org/models/vit_b_32-d86f8d99.pth",
        transforms=partial(ImageClassification, crop_size=224),
        meta={
            **_COMMON_META,
            "num_params": 88224232,
            "min_size": (224, 224),
            "recipe": "https://github.com/pytorch/vision/tree/main/references/classification#vit_b_32",
            "acc@1": 75.912,
            "acc@5": 92.466,
        },
    )
    DEFAULT = IMAGENET1K_V1


class ViT_L_16_Weights(WeightsEnum):
    IMAGENET1K_V1 = Weights(
        url="https://download.pytorch.org/models/vit_l_16-852ce7e3.pth",
        transforms=partial(ImageClassification, crop_size=224, resize_size=242),
        meta={
            **_COMMON_META,
            "num_params": 304326632,
            "min_size": (224, 224),
            "recipe": "https://github.com/pytorch/vision/tree/main/references/classification#vit_l_16",
            "acc@1": 79.662,
            "acc@5": 94.638,
        },
    )
    IMAGENET1K_SWAG_E2E_V1 = Weights(
        url="https://download.pytorch.org/models/vit_l_16_swag-4f3808c9.pth",
        transforms=partial(
            ImageClassification,
            crop_size=512,
            resize_size=512,
            interpolation=InterpolationMode.BICUBIC,
        ),
        meta={
            **_COMMON_SWAG_META,
            "num_params": 305174504,
            "min_size": (512, 512),
            "acc@1": 88.064,
            "acc@5": 98.512,
        },
    )
    IMAGENET1K_SWAG_LINEAR_V1 = Weights(
        url="https://download.pytorch.org/models/vit_l_16_lc_swag-4d563306.pth",
        transforms=partial(
            ImageClassification,
            crop_size=224,
            resize_size=224,
            interpolation=InterpolationMode.BICUBIC,
        ),
        meta={
            **_COMMON_SWAG_META,
            "recipe": "https://github.com/pytorch/vision/pull/5793",
            "num_params": 304326632,
            "min_size": (224, 224),
            "acc@1": 85.146,
            "acc@5": 97.422,
        },
    )
    DEFAULT = IMAGENET1K_V1


class ViT_L_32_Weights(WeightsEnum):
    IMAGENET1K_V1 = Weights(
        url="https://download.pytorch.org/models/vit_l_32-c7638314.pth",
        transforms=partial(ImageClassification, crop_size=224),
        meta={
            **_COMMON_META,
            "num_params": 306535400,
            "min_size": (224, 224),
            "recipe": "https://github.com/pytorch/vision/tree/main/references/classification#vit_l_32",
            "acc@1": 76.972,
            "acc@5": 93.07,
        },
    )
    DEFAULT = IMAGENET1K_V1


class ViT_H_14_Weights(WeightsEnum):
    IMAGENET1K_SWAG_E2E_V1 = Weights(
        url="https://download.pytorch.org/models/vit_h_14_swag-80465313.pth",
        transforms=partial(
            ImageClassification,
            crop_size=518,
            resize_size=518,
            interpolation=InterpolationMode.BICUBIC,
        ),
        meta={
            **_COMMON_SWAG_META,
            "num_params": 633470440,
            "min_size": (518, 518),
            "acc@1": 88.552,
            "acc@5": 98.694,
        },
    )
    IMAGENET1K_SWAG_LINEAR_V1 = Weights(
        url="https://download.pytorch.org/models/vit_h_14_lc_swag-c1eb923e.pth",
        transforms=partial(
            ImageClassification,
            crop_size=224,
            resize_size=224,
            interpolation=InterpolationMode.BICUBIC,
        ),
        meta={
            **_COMMON_SWAG_META,
            "recipe": "https://github.com/pytorch/vision/pull/5793",
            "num_params": 632045800,
            "min_size": (224, 224),
            "acc@1": 85.708,
            "acc@5": 97.730,
        },
    )
    DEFAULT = IMAGENET1K_SWAG_E2E_V1


@handle_legacy_interface(weights=("pretrained", ViT_B_16_Weights.IMAGENET1K_V1))
def vit_b_16(*, weights: Optional[ViT_B_16_Weights] = None, progress: bool = True, **kwargs: Any) -> VisionTransformer:
    """
    Constructs a vit_b_16 architecture from
    `"An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale" <https://arxiv.org/abs/2010.11929>`_.

    Args:
        weights (ViT_B_16_Weights, optional): The pretrained weights for the model
        progress (bool): If True, displays a progress bar of the download to stderr
    """
    weights = ViT_B_16_Weights.verify(weights)

    return _vision_transformer(
        patch_size=16,
        num_layers=12,
        num_heads=12,
        hidden_dim=768,
        mlp_dim=3072,
        weights=weights,
        progress=progress,
        **kwargs,
    )


@handle_legacy_interface(weights=("pretrained", ViT_B_32_Weights.IMAGENET1K_V1))
def vit_b_32(*, weights: Optional[ViT_B_32_Weights] = None, progress: bool = True, **kwargs: Any) -> VisionTransformer:
    """
    Constructs a vit_b_32 architecture from
    `"An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale" <https://arxiv.org/abs/2010.11929>`_.

    Args:
        weights (ViT_B_32_Weights, optional): The pretrained weights for the model
        progress (bool): If True, displays a progress bar of the download to stderr
    """
    weights = ViT_B_32_Weights.verify(weights)

    return _vision_transformer(
        patch_size=32,
        num_layers=12,
        num_heads=12,
        hidden_dim=768,
        mlp_dim=3072,
        weights=weights,
        progress=progress,
        **kwargs,
    )


@handle_legacy_interface(weights=("pretrained", ViT_L_16_Weights.IMAGENET1K_V1))
def vit_l_16(*, weights: Optional[ViT_L_16_Weights] = None, progress: bool = True, **kwargs: Any) -> VisionTransformer:
    """
    Constructs a vit_l_16 architecture from
    `"An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale" <https://arxiv.org/abs/2010.11929>`_.

    Args:
        weights (ViT_L_16_Weights, optional): The pretrained weights for the model
        progress (bool): If True, displays a progress bar of the download to stderr
    """
    weights = ViT_L_16_Weights.verify(weights)

    return _vision_transformer(
        patch_size=16,
        num_layers=24,
        num_heads=16,
        hidden_dim=1024,
        mlp_dim=4096,
        weights=weights,
        progress=progress,
        **kwargs,
    )


@handle_legacy_interface(weights=("pretrained", ViT_L_32_Weights.IMAGENET1K_V1))
def vit_l_32(*, weights: Optional[ViT_L_32_Weights] = None, progress: bool = True, **kwargs: Any) -> VisionTransformer:
    """
    Constructs a vit_l_32 architecture from
    `"An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale" <https://arxiv.org/abs/2010.11929>`_.

    Args:
        weights (ViT_L_32_Weights, optional): The pretrained weights for the model
        progress (bool): If True, displays a progress bar of the download to stderr
    """
    weights = ViT_L_32_Weights.verify(weights)

    return _vision_transformer(
        patch_size=32,
        num_layers=24,
        num_heads=16,
        hidden_dim=1024,
        mlp_dim=4096,
        weights=weights,
        progress=progress,
        **kwargs,
    )


def vit_h_14(*, weights: Optional[ViT_H_14_Weights] = None, progress: bool = True, **kwargs: Any) -> VisionTransformer:
    """
    Constructs a vit_h_14 architecture from
    `"An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale" <https://arxiv.org/abs/2010.11929>`_.

    Args:
        weights (ViT_H_14_Weights, optional): The pretrained weights for the model
        progress (bool): If True, displays a progress bar of the download to stderr
    """
    weights = ViT_H_14_Weights.verify(weights)

    return _vision_transformer(
        patch_size=14,
        num_layers=32,
        num_heads=16,
        hidden_dim=1280,
        mlp_dim=5120,
        weights=weights,
        progress=progress,
        **kwargs,
    )


def interpolate_embeddings(
    image_size: int,
    patch_size: int,
    model_state: "OrderedDict[str, torch.Tensor]",
    interpolation_mode: str = "bicubic",
    reset_heads: bool = False,
) -> "OrderedDict[str, torch.Tensor]":
    """This function helps interpolating positional embeddings during checkpoint loading,
    especially when you want to apply a pre-trained model on images with different resolution.

    Args:
        image_size (int): Image size of the new model.
        patch_size (int): Patch size of the new model.
        model_state (OrderedDict[str, torch.Tensor]): State dict of the pre-trained model.
        interpolation_mode (str): The algorithm used for upsampling. Default: bicubic.
        reset_heads (bool): If true, not copying the state of heads. Default: False.

    Returns:
        OrderedDict[str, torch.Tensor]: A state dict which can be loaded into the new model.
    """
    # Shape of pos_embedding is (1, seq_length, hidden_dim)
    pos_embedding = model_state["encoder.pos_embedding"]
    n, seq_length, hidden_dim = pos_embedding.shape
    if n != 1:
        raise ValueError(f"Unexpected position embedding shape: {pos_embedding.shape}")

    new_seq_length = (image_size // patch_size) ** 2 + 1

    # Need to interpolate the weights for the position embedding.
    # We do this by reshaping the positions embeddings to a 2d grid, performing
    # an interpolation in the (h, w) space and then reshaping back to a 1d grid.
    if new_seq_length != seq_length:
        # The class token embedding shouldn't be interpolated so we split it up.
        seq_length -= 1
        new_seq_length -= 1
        pos_embedding_token = pos_embedding[:, :1, :]
        pos_embedding_img = pos_embedding[:, 1:, :]

        # (1, seq_length, hidden_dim) -> (1, hidden_dim, seq_length)
        pos_embedding_img = pos_embedding_img.permute(0, 2, 1)
        seq_length_1d = int(math.sqrt(seq_length))
        if seq_length_1d * seq_length_1d != seq_length:
            raise ValueError(
                f"seq_length is not a perfect square! Instead got seq_length_1d * seq_length_1d = {seq_length_1d * seq_length_1d } and seq_length = {seq_length}"
            )

        # (1, hidden_dim, seq_length) -> (1, hidden_dim, seq_l_1d, seq_l_1d)
        pos_embedding_img = pos_embedding_img.reshape(1, hidden_dim, seq_length_1d, seq_length_1d)
        new_seq_length_1d = image_size // patch_size

        # Perform interpolation.
        # (1, hidden_dim, seq_l_1d, seq_l_1d) -> (1, hidden_dim, new_seq_l_1d, new_seq_l_1d)
        new_pos_embedding_img = nn.functional.interpolate(
            pos_embedding_img,
            size=new_seq_length_1d,
            mode=interpolation_mode,
            align_corners=True,
        )

        # (1, hidden_dim, new_seq_l_1d, new_seq_l_1d) -> (1, hidden_dim, new_seq_length)
        new_pos_embedding_img = new_pos_embedding_img.reshape(1, hidden_dim, new_seq_length)

        # (1, hidden_dim, new_seq_length) -> (1, new_seq_length, hidden_dim)
        new_pos_embedding_img = new_pos_embedding_img.permute(0, 2, 1)
        new_pos_embedding = torch.cat([pos_embedding_token, new_pos_embedding_img], dim=1)

        model_state["encoder.pos_embedding"] = new_pos_embedding

        if reset_heads:
            model_state_copy: "OrderedDict[str, torch.Tensor]" = OrderedDict()
            for k, v in model_state.items():
                if not k.startswith("heads"):
                    model_state_copy[k] = v
            model_state = model_state_copy

    return model_state