第一次提交Yolo项目

mobileclip/modules/common/__init__.py

@@ -0,0 +1,6 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+#
+# For licensing see accompanying LICENSE file.
+# Copyright (C) 2024 Apple Inc. All rights reserved.
+#

mobileclip/modules/common/mobileone.py

@@ -0,0 +1,330 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+# For licensing see accompanying LICENSE file.
+# Copyright (C) 2024 Apple Inc. All Rights Reserved.
+
+from __future__ import annotations
+
+import copy
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+__all__ = ["MobileOneBlock", "reparameterize_model"]
+
+
+class SEBlock(nn.Module):
+    """Squeeze and Excite module.
+
+    Pytorch implementation of `Squeeze-and-Excitation Networks` - https://arxiv.org/pdf/1709.01507.pdf
+    """
+
+    def __init__(self, in_channels: int, rd_ratio: float = 0.0625) -> None:
+        """Construct a Squeeze and Excite Module.
+
+        Args:
+            in_channels: Number of input channels.
+            rd_ratio: Input channel reduction ratio.
+        """
+        super().__init__()
+        self.reduce = nn.Conv2d(
+            in_channels=in_channels,
+            out_channels=int(in_channels * rd_ratio),
+            kernel_size=1,
+            stride=1,
+            bias=True,
+        )
+        self.expand = nn.Conv2d(
+            in_channels=int(in_channels * rd_ratio),
+            out_channels=in_channels,
+            kernel_size=1,
+            stride=1,
+            bias=True,
+        )
+
+    def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+        """Apply forward pass."""
+        _b, c, h, w = inputs.size()
+        x = F.avg_pool2d(inputs, kernel_size=[h, w])
+        x = self.reduce(x)
+        x = F.relu(x)
+        x = self.expand(x)
+        x = torch.sigmoid(x)
+        x = x.view(-1, c, 1, 1)
+        return inputs * x
+
+
+class MobileOneBlock(nn.Module):
+    """MobileOne building block.
+
+    This block has a multi-branched architecture at train-time and plain-CNN style architecture at inference time For
+    more details, please refer to our paper: `An Improved One millisecond Mobile Backbone` -
+    https://arxiv.org/pdf/2206.04040.pdf
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        kernel_size: int,
+        stride: int = 1,
+        padding: int = 0,
+        dilation: int = 1,
+        groups: int = 1,
+        inference_mode: bool = False,
+        use_se: bool = False,
+        use_act: bool = True,
+        use_scale_branch: bool = True,
+        num_conv_branches: int = 1,
+        activation: nn.Module = nn.GELU(),
+    ) -> None:
+        """Construct a MobileOneBlock module.
+
+        Args:
+            in_channels: Number of channels in the input.
+            out_channels: Number of channels produced by the block.
+            kernel_size: Size of the convolution kernel.
+            stride: Stride size.
+            padding: Zero-padding size.
+            dilation: Kernel dilation factor.
+            groups: Group number.
+            inference_mode: If True, instantiates model in inference mode.
+            use_se: Whether to use SE-ReLU activations.
+            use_act: Whether to use activation. Default: ``True``
+            use_scale_branch: Whether to use scale branch. Default: ``True``
+            num_conv_branches: Number of linear conv branches.
+        """
+        super().__init__()
+        self.inference_mode = inference_mode
+        self.groups = groups
+        self.stride = stride
+        self.padding = padding
+        self.dilation = dilation
+        self.kernel_size = kernel_size
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.num_conv_branches = num_conv_branches
+
+        # Check if SE-ReLU is requested
+        if use_se:
+            self.se = SEBlock(out_channels)
+        else:
+            self.se = nn.Identity()
+
+        if use_act:
+            self.activation = activation
+        else:
+            self.activation = nn.Identity()
+
+        if inference_mode:
+            self.reparam_conv = nn.Conv2d(
+                in_channels=in_channels,
+                out_channels=out_channels,
+                kernel_size=kernel_size,
+                stride=stride,
+                padding=padding,
+                dilation=dilation,
+                groups=groups,
+                bias=True,
+            )
+        else:
+            # Re-parameterizable skip connection
+            self.rbr_skip = (
+                nn.BatchNorm2d(num_features=in_channels) if out_channels == in_channels and stride == 1 else None
+            )
+
+            # Re-parameterizable conv branches
+            if num_conv_branches > 0:
+                rbr_conv = list()
+                for _ in range(self.num_conv_branches):
+                    rbr_conv.append(self._conv_bn(kernel_size=kernel_size, padding=padding))
+                self.rbr_conv = nn.ModuleList(rbr_conv)
+            else:
+                self.rbr_conv = None
+
+            # Re-parameterizable scale branch
+            self.rbr_scale = None
+            if not isinstance(kernel_size, int):
+                kernel_size = kernel_size[0]
+            if (kernel_size > 1) and use_scale_branch:
+                self.rbr_scale = self._conv_bn(kernel_size=1, padding=0)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Apply forward pass."""
+        # Inference mode forward pass.
+        if self.inference_mode:
+            return self.activation(self.se(self.reparam_conv(x)))
+
+        # Multi-branched train-time forward pass.
+        # Skip branch output
+        identity_out = 0
+        if self.rbr_skip is not None:
+            identity_out = self.rbr_skip(x)
+
+        # Scale branch output
+        scale_out = 0
+        if self.rbr_scale is not None:
+            scale_out = self.rbr_scale(x)
+
+        # Other branches
+        out = scale_out + identity_out
+        if self.rbr_conv is not None:
+            for ix in range(self.num_conv_branches):
+                out += self.rbr_conv[ix](x)
+
+        return self.activation(self.se(out))
+
+    def reparameterize(self):
+        """Following works like `RepVGG: Making VGG-style ConvNets Great Again` - https://arxiv.org/pdf/2101.03697.pdf.
+        We re-parameterize multi-branched architecture used at training time to obtain a plain CNN-like
+        structure for inference.
+        """
+        if self.inference_mode:
+            return
+        kernel, bias = self._get_kernel_bias()
+        self.reparam_conv = nn.Conv2d(
+            in_channels=self.in_channels,
+            out_channels=self.out_channels,
+            kernel_size=self.kernel_size,
+            stride=self.stride,
+            padding=self.padding,
+            dilation=self.dilation,
+            groups=self.groups,
+            bias=True,
+        )
+        self.reparam_conv.weight.data = kernel
+        self.reparam_conv.bias.data = bias
+
+        # Delete un-used branches
+        for para in self.parameters():
+            para.detach_()
+        self.__delattr__("rbr_conv")
+        self.__delattr__("rbr_scale")
+        if hasattr(self, "rbr_skip"):
+            self.__delattr__("rbr_skip")
+
+        self.inference_mode = True
+
+    def _get_kernel_bias(self) -> tuple[torch.Tensor, torch.Tensor]:
+        """Method to obtain re-parameterized kernel and bias. Reference:
+        https://github.com/DingXiaoH/RepVGG/blob/main/repvgg.py#L83.
+
+        Returns:
+            Tuple of (kernel, bias) after fusing branches.
+        """
+        # get weights and bias of scale branch
+        kernel_scale = 0
+        bias_scale = 0
+        if self.rbr_scale is not None:
+            kernel_scale, bias_scale = self._fuse_bn_tensor(self.rbr_scale)
+            # Pad scale branch kernel to match conv branch kernel size.
+            pad = self.kernel_size // 2
+            kernel_scale = torch.nn.functional.pad(kernel_scale, [pad, pad, pad, pad])
+
+        # get weights and bias of skip branch
+        kernel_identity = 0
+        bias_identity = 0
+        if self.rbr_skip is not None:
+            kernel_identity, bias_identity = self._fuse_bn_tensor(self.rbr_skip)
+
+        # get weights and bias of conv branches
+        kernel_conv = 0
+        bias_conv = 0
+        if self.rbr_conv is not None:
+            for ix in range(self.num_conv_branches):
+                _kernel, _bias = self._fuse_bn_tensor(self.rbr_conv[ix])
+                kernel_conv += _kernel
+                bias_conv += _bias
+
+        kernel_final = kernel_conv + kernel_scale + kernel_identity
+        bias_final = bias_conv + bias_scale + bias_identity
+        return kernel_final, bias_final
+
+    def _fuse_bn_tensor(self, branch: nn.Sequential | nn.BatchNorm2d) -> tuple[torch.Tensor, torch.Tensor]:
+        """Method to fuse batchnorm layer with preceeding conv layer. Reference:
+        https://github.com/DingXiaoH/RepVGG/blob/main/repvgg.py#L95.
+
+        Args:
+            branch: Sequence of ops to be fused.
+
+        Returns:
+            Tuple of (kernel, bias) after fusing batchnorm.
+        """
+        if isinstance(branch, nn.Sequential):
+            kernel = branch.conv.weight
+            running_mean = branch.bn.running_mean
+            running_var = branch.bn.running_var
+            gamma = branch.bn.weight
+            beta = branch.bn.bias
+            eps = branch.bn.eps
+        else:
+            assert isinstance(branch, nn.BatchNorm2d)
+            if not hasattr(self, "id_tensor"):
+                input_dim = self.in_channels // self.groups
+
+                kernel_size = self.kernel_size
+                if isinstance(self.kernel_size, int):
+                    kernel_size = (self.kernel_size, self.kernel_size)
+
+                kernel_value = torch.zeros(
+                    (self.in_channels, input_dim, kernel_size[0], kernel_size[1]),
+                    dtype=branch.weight.dtype,
+                    device=branch.weight.device,
+                )
+                for i in range(self.in_channels):
+                    kernel_value[i, i % input_dim, kernel_size[0] // 2, kernel_size[1] // 2] = 1
+                self.id_tensor = kernel_value
+            kernel = self.id_tensor
+            running_mean = branch.running_mean
+            running_var = branch.running_var
+            gamma = branch.weight
+            beta = branch.bias
+            eps = branch.eps
+        std = (running_var + eps).sqrt()
+        t = (gamma / std).reshape(-1, 1, 1, 1)
+        return kernel * t, beta - running_mean * gamma / std
+
+    def _conv_bn(self, kernel_size: int, padding: int) -> nn.Sequential:
+        """Helper method to construct conv-batchnorm layers.
+
+        Args:
+            kernel_size: Size of the convolution kernel.
+            padding: Zero-padding size.
+
+        Returns:
+            Conv-BN module.
+        """
+        mod_list = nn.Sequential()
+        mod_list.add_module(
+            "conv",
+            nn.Conv2d(
+                in_channels=self.in_channels,
+                out_channels=self.out_channels,
+                kernel_size=kernel_size,
+                stride=self.stride,
+                padding=padding,
+                groups=self.groups,
+                bias=False,
+            ),
+        )
+        mod_list.add_module("bn", nn.BatchNorm2d(num_features=self.out_channels))
+        return mod_list
+
+
+def reparameterize_model(model: torch.nn.Module) -> nn.Module:
+    """Method returns a model where a multi-branched structure used in training is re-parameterized into a single branch
+    for inference.
+
+    Args:
+        model: MobileOne model in train mode.
+
+    Returns:
+        MobileOne model in inference mode.
+    """
+    # Avoid editing original graph
+    model = copy.deepcopy(model)
+    for module in model.modules():
+        if hasattr(module, "reparameterize"):
+            module.reparameterize()
+    return model

mobileclip/modules/common/transformer.py

@@ -0,0 +1,410 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+#
+# For licensing see accompanying LICENSE file.
+# Copyright (C) 2024 Apple Inc. All Rights Reserved.
+#
+"""
+Implementation of the following modules is borrowed from ml-cvnets repo:
+https://github.com/apple/ml-cvnets/blob/main/cvnets/layers/multi_head_attention.py
+https://github.com/apple/ml-cvnets/blob/main/cvnets/text_encoders/transformer.py.
+
+Please see ACKNOWLEDGMENTS for license details.
+"""
+
+from __future__ import annotations
+
+import torch
+from torch import Size, Tensor, nn
+from torch.nn import functional as F
+from torchvision.ops import StochasticDepth
+
+from mobileclip import logger
+
+
+class LayerNormFP32(nn.LayerNorm):
+    """Applies `Layer Normalization <https://arxiv.org/abs/1607.06450>`_ over a input tensor with FP32 precision."""
+
+    def __init__(
+        self,
+        normalized_shape: int | list[int] | Size,
+        eps: float | None = 1e-5,
+        elementwise_affine: bool | None = True,
+        *args,
+        **kwargs,
+    ):
+        super().__init__(
+            normalized_shape=normalized_shape,
+            eps=eps,
+            elementwise_affine=elementwise_affine,
+            *args,
+            **kwargs,
+        )
+
+    def forward(self, x: Tensor) -> Tensor:
+        # Convert input from dtype X to FP32 and perform normalization operation.
+        # This may help with underflow/overflow issues that we typically see with normalization layers
+        inp_dtype = x.dtype
+        return super().forward(x.to(torch.float32)).to(inp_dtype)
+
+
+def get_normalization_layer(norm_type, num_features):
+    if norm_type == "layer_norm":
+        return nn.LayerNorm(num_features)
+    elif norm_type == "layer_norm_fp32":
+        return LayerNormFP32(num_features)
+    else:
+        raise NotImplementedError(f"Option: {norm_type} not supported.")
+
+
+class PositionalEmbedding(nn.Module):
+    def __init__(
+        self,
+        num_embeddings: int,
+        embedding_dim: int,
+        padding_idx: int | None = None,
+        is_learnable: bool | None = False,
+        interpolation_mode: str | None = "bilinear",
+        *args,
+        **kwargs,
+    ):
+        super().__init__()
+        # Add other pos embedding here and logic to choose between them
+        module = LearnablePositionalEmbedding
+
+        self.pos_embed = module(
+            num_embeddings=num_embeddings,
+            embedding_dim=embedding_dim,
+            padding_idx=padding_idx,
+            interpolation_mode=interpolation_mode,
+            *args,
+            **kwargs,
+        )
+
+    def forward(self, seq_len: int, *args, **kwargs) -> Tensor:
+        return self.pos_embed(seq_len, *args, **kwargs)
+
+    def __repr__(self):
+        return self.pos_embed.__repr__()
+
+
+class LearnablePositionalEmbedding(nn.Module):
+    """Learnable Positional embedding."""
+
+    def __init__(
+        self,
+        num_embeddings: int,
+        embedding_dim: int,
+        padding_idx: int | None = None,
+        interpolation_mode: str | None = "bilinear",
+        *args,
+        **kwargs,
+    ):
+        super().__init__()
+        self.pos_embed = nn.Parameter(torch.empty(1, 1, num_embeddings, embedding_dim))
+        self.embedding_dim = embedding_dim
+        self.num_embeddings = num_embeddings
+        self.padding_idx = padding_idx
+        self.interpolation_mode = interpolation_mode
+
+        self.reset_parameters()
+
+    def reset_parameters(self) -> None:
+        nn.init.trunc_normal_(self.pos_embed, mean=0, std=self.embedding_dim**-0.5)
+        if self.padding_idx is not None:
+            with torch.no_grad():
+                self.pos_embed[:, :, self.padding_idx, ...] = 0.0
+
+    def forward(self, seq_len: int, *args, **kwargs) -> Tensor:
+        # scale pos embedding
+        pos_embed = self.pos_embed
+        if self.padding_idx is not None:
+            with torch.no_grad():
+                pos_embed[:, :, self.padding_idx, ...] = 0.0
+
+        if seq_len != self.num_embeddings:
+            pos_embed = F.interpolate(
+                pos_embed,
+                size=(seq_len, self.embedding_dim),
+                mode=self.interpolation_mode,
+            )
+
+        # Input is of the form [Batch, Seq_len, Embedding_dim]
+        return pos_embed.reshape(1, seq_len, self.embedding_dim)
+
+    def __repr__(self):
+        return f"{self.__class__.__name__}(num_embeddings={self.num_embeddings}, embedding_dim={self.embedding_dim}, padding_idx={self.padding_idx})"
+
+
+class MultiHeadAttention(nn.Module):
+    """This layer applies a multi-head self- or cross-attention as described in `Attention is all you need
+    <https://arxiv.org/abs/1706.03762>`_ paper.
+
+    Args:
+        embed_dim (int): :math:`C_{in}` from an expected input of size :math:`(N, S, C_{in})`
+        num_heads (int): Number of heads in multi-head attention
+        attn_dropout (Optional[float]): Attention dropout. Default: 0.0
+        bias (Optional[bool]): Use bias or not. Default: ``True``
+
+    Notes:
+        - Input:
+           - Query tensor (x_q) :math:`(N, S, C_{in})` where :math:`N` is batch size, :math:`S` is number of source tokens,
+        and: math:`C_{in}` is input embedding dim
+           - Optional Key-Value tensor (x_kv) :math:`(N, T, C_{in})` where :math:`T` is number of target tokens
+        - Output: same shape as the input
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        attn_dropout: float | None = 0.0,
+        bias: bool | None = True,
+        output_dim: int | None = None,
+        *args,
+        **kwargs,
+    ) -> None:
+        if output_dim is None:
+            output_dim = embed_dim
+        super().__init__()
+        if embed_dim % num_heads != 0:
+            logger.error(
+                f"Embedding dim must be divisible by number of heads in {self.__class__.__name__}. Got: embed_dim={embed_dim} and num_heads={num_heads}"
+            )
+
+        self.qkv_proj = nn.Linear(in_features=embed_dim, out_features=3 * embed_dim, bias=bias)
+
+        self.attn_dropout = nn.Dropout(p=attn_dropout)
+        self.out_proj = nn.Linear(in_features=embed_dim, out_features=output_dim, bias=bias)
+
+        self.head_dim = embed_dim // num_heads
+        self.scaling = self.head_dim**-0.5
+        self.softmax = nn.Softmax(dim=-1)
+        self.num_heads = num_heads
+        self.embed_dim = embed_dim
+        self.use_separate_proj_weight = embed_dim != output_dim
+
+    def __repr__(self):
+        return f"{self.__class__.__name__}(head_dim={self.head_dim}, num_heads={self.num_heads}, attn_dropout={self.attn_dropout.p})"
+
+    def _forward_impl(
+        self,
+        x_q: Tensor,
+        x_kv: Tensor | None = None,
+        key_padding_mask: Tensor | None = None,
+        attn_mask: Tensor | None = None,
+    ) -> Tensor:
+        # [N, S, C]
+        b_sz, S_len, _in_channels = x_q.shape
+
+        if x_kv is None:
+            # self-attention
+            # [N, S, C] --> [N, S, 3C] --> [N, S, 3, h, c] where C = hc
+            qkv = self.qkv_proj(x_q).reshape(b_sz, S_len, 3, self.num_heads, -1)
+            # [N, S, 3, h, c] --> [N, h, 3, S, C]
+            qkv = qkv.transpose(1, 3).contiguous()
+
+            # [N, h, 3, S, C] --> [N, h, S, C] x 3
+            query, key, value = qkv[:, :, 0], qkv[:, :, 1], qkv[:, :, 2]
+        else:
+            T_len = x_kv.shape[1]
+
+            # cross-attention
+            # [N, S, C]
+            query = F.linear(
+                x_q,
+                weight=self.qkv_proj.weight[: self.embed_dim, ...],
+                bias=self.qkv_proj.bias[: self.embed_dim] if self.qkv_proj.bias is not None else None,
+            )
+            # [N, S, C] --> [N, S, h, c] --> [N, h, S, c]
+            query = query.reshape(b_sz, S_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+            # [N, T, C] --> [N, T, 2C]
+            kv = F.linear(
+                x_kv,
+                weight=self.qkv_proj.weight[self.embed_dim :, ...],
+                bias=self.qkv_proj.bias[self.embed_dim :] if self.qkv_proj.bias is not None else None,
+            )
+            # [N, T, 2C] --> [N, T, 2, h, c]
+            kv = kv.reshape(b_sz, T_len, 2, self.num_heads, self.head_dim)
+            # [N, T, 2, h, c] --> [N, h, 2, T, c]
+            kv = kv.transpose(1, 3).contiguous()
+            key, value = kv[:, :, 0], kv[:, :, 1]
+
+        query = query * self.scaling
+
+        # [N h, T, c] --> [N, h, c, T]
+        key = key.transpose(-1, -2)
+
+        # QK^T
+        # [N, h, S, c] x [N, h, c, T] --> [N, h, S, T]
+        attn = torch.matmul(query, key)
+
+        batch_size, _num_heads, num_src_tokens, num_tgt_tokens = attn.shape
+        if attn_mask is not None:
+            # attn_mask shape should be the same as attn
+            assert list(attn_mask.shape) == [
+                batch_size,
+                num_src_tokens,
+                num_tgt_tokens,
+            ], (
+                f"Shape of attention mask should be [{batch_size}, {num_src_tokens}, {num_tgt_tokens}]. Got: {attn_mask.shape}"
+            )
+            # [N, S, T] --> [N, 1, S, T]
+            attn_mask = attn_mask.unsqueeze(1)
+            attn = attn + attn_mask
+
+        if key_padding_mask is not None:
+            # Do not attend to padding positions
+            # key padding mask size is [N, T]
+            assert key_padding_mask.dim() == 2 and list(key_padding_mask.shape) == [
+                batch_size,
+                num_tgt_tokens,
+            ], (
+                f"Key_padding_mask should be 2-dimension with shape [{batch_size}, {num_tgt_tokens}]. Got: {key_padding_mask.shape}"
+            )
+            attn = attn.masked_fill(
+                key_padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool),  # [N, T] --> [N, 1, 1, T]
+                float("-inf"),
+            )
+
+        attn_dtype = attn.dtype
+        attn_as_float = self.softmax(attn.float())
+        attn = attn_as_float.to(attn_dtype)
+        attn = self.attn_dropout(attn)
+
+        # weighted sum
+        # [N, h, S, T] x [N, h, T, c] --> [N, h, S, c]
+        out = torch.matmul(attn, value)
+
+        # [N, h, S, c] --> [N, S, h, c] --> [N, S, C]
+        out = out.transpose(1, 2).reshape(b_sz, S_len, -1)
+        out = self.out_proj(out)
+
+        return out
+
+    def forward(
+        self,
+        x_q: Tensor,
+        x_kv: Tensor | None = None,
+        key_padding_mask: Tensor | None = None,
+        attn_mask: Tensor | None = None,
+        *args,
+        **kwargs,
+    ) -> Tensor:
+        # [Batch , Sequence, Hidden_dim]
+        return self._forward_impl(
+            x_q=x_q,
+            x_kv=x_kv,
+            key_padding_mask=key_padding_mask,
+            attn_mask=attn_mask,
+        )
+
+
+class TransformerEncoder(nn.Module):
+    """This class defines the pre-norm `Transformer encoder <https://arxiv.org/abs/1706.03762>`_.
+
+    Args:
+        embed_dim: :math:`C_{in}` from an expected input of size :math:`(N, P, C_{in})`.
+        ffn_latent_dim: Inner dimension of the FFN.
+        num_heads: Number of heads in multi-head attention. Default: 8.
+        attn_dropout: Dropout rate for attention in multi-head attention. Default: 0.0
+        dropout: Dropout rate. Default: 0.0.
+        ffn_dropout: Dropout between FFN layers. Default: 0.0.
+        transformer_norm_layer: Normalization layer. Default: layer_norm.
+        stochastic_dropout: Stochastic dropout setting. Default: 0.0.
+
+    Notes:
+        - Input: :math:`(N, P, C_{in})` where :math:`N` is batch size, :math:`P` is number of patches,
+        and: math:`C_{in}` is input embedding dim
+        - Output: same shape as the input
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        ffn_latent_dim: int,
+        num_heads: int | None = 8,
+        attn_dropout: float | None = 0.0,
+        dropout: float | None = 0.0,
+        ffn_dropout: float | None = 0.0,
+        transformer_norm_layer: str | None = "layer_norm",
+        stochastic_dropout: float | None = 0.0,
+        *args,
+        **kwargs,
+    ) -> None:
+        super().__init__()
+
+        # Build attention layer
+        attn_unit = MultiHeadAttention(
+            embed_dim,
+            num_heads,
+            attn_dropout=attn_dropout,
+            bias=True,
+        )
+
+        self.pre_norm_mha = nn.Sequential(
+            get_normalization_layer(norm_type=transformer_norm_layer, num_features=embed_dim),
+            attn_unit,
+            nn.Dropout(p=dropout),
+        )
+
+        act_name = nn.GELU()
+        self.pre_norm_ffn = nn.Sequential(
+            get_normalization_layer(norm_type=transformer_norm_layer, num_features=embed_dim),
+            nn.Linear(in_features=embed_dim, out_features=ffn_latent_dim, bias=True),
+            act_name,
+            nn.Dropout(p=ffn_dropout),
+            nn.Linear(in_features=ffn_latent_dim, out_features=embed_dim, bias=True),
+            nn.Dropout(p=dropout),
+        )
+
+        self.drop_path = nn.Identity()
+        if stochastic_dropout > 0.0:
+            if dropout > 0.0:
+                logger.error(
+                    "Stochastic dropout and dropout are mutually exclusive. "
+                    "Use either of them, but not both."
+                    f"Got: {stochastic_dropout} and {dropout}"
+                )
+            self.drop_path = StochasticDepth(p=stochastic_dropout, mode="row")
+
+        self.embed_dim = embed_dim
+        self.ffn_dim = ffn_latent_dim
+        self.ffn_dropout = ffn_dropout
+        self.stochastic_dropout = stochastic_dropout
+        self.std_dropout = dropout
+        self.attn_fn_name = attn_unit.__class__.__name__
+        self.act_fn_name = act_name.__class__.__name__
+        self.norm_type = transformer_norm_layer
+
+    def __repr__(self) -> str:
+        return f"{self.__class__.__name__}(embed_dim={self.embed_dim}, ffn_dim={self.ffn_dim}, dropout={self.std_dropout}, ffn_dropout={self.ffn_dropout}, stochastic_dropout={self.stochastic_dropout}, attn_fn={self.attn_fn_name}, act_fn={self.act_fn_name}, norm_fn={self.norm_type})"
+
+    def forward(
+        self,
+        x: Tensor,
+        x_prev: Tensor | None = None,
+        key_padding_mask: Tensor | None = None,
+        attn_mask: Tensor | None = None,
+        *args,
+        **kwargs,
+    ) -> Tensor:
+        # Multi-head attention
+        res = x
+        x = self.pre_norm_mha[0](x)  # norm
+        x = self.pre_norm_mha[1](
+            x_q=x,
+            x_kv=x_prev,
+            key_padding_mask=key_padding_mask,
+            attn_mask=attn_mask,
+            *args,
+            **kwargs,
+        )  # mha
+
+        x = self.drop_path(self.pre_norm_mha[2](x))  # applying stochastic depth
+        x = x + res
+
+        # Feed forward network
+        x = x + self.drop_path(self.pre_norm_ffn(x))
+        return x