Merge pull request #1955 from myhloli/dev

Dev push

magic_pdf/model/sub_modules/mfr/unimernet/unimernet_hf/unimer_swin/__init__.py

+from .configuration_unimer_swin import UnimerSwinConfig
+from .modeling_unimer_swin import UnimerSwinModel
+from .image_processing_unimer_swin import UnimerSwinImageProcessor
+
+__all__ = [
+    "UnimerSwinConfig",
+    "UnimerSwinModel",
+    "UnimerSwinImageProcessor",
+]

magic_pdf/model/sub_modules/mfr/unimernet/unimernet_hf/unimer_swin/configuration_unimer_swin.py

+# coding=utf-8
+# Copyright 2022 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.
+"""Donut Swin Transformer model configuration"""
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class UnimerSwinConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`UnimerSwinModel`]. It is used to instantiate a
+    Donut model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of the Donut
+    [naver-clova-ix/donut-base](https://huggingface.co/naver-clova-ix/donut-base) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        image_size (`int`, *optional*, defaults to 224):
+            The size (resolution) of each image.
+        patch_size (`int`, *optional*, defaults to 4):
+            The size (resolution) of each patch.
+        num_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        embed_dim (`int`, *optional*, defaults to 96):
+            Dimensionality of patch embedding.
+        depths (`list(int)`, *optional*, defaults to `[2, 2, 6, 2]`):
+            Depth of each layer in the Transformer encoder.
+        num_heads (`list(int)`, *optional*, defaults to `[3, 6, 12, 24]`):
+            Number of attention heads in each layer of the Transformer encoder.
+        window_size (`int`, *optional*, defaults to 7):
+            Size of windows.
+        mlp_ratio (`float`, *optional*, defaults to 4.0):
+            Ratio of MLP hidden dimensionality to embedding dimensionality.
+        qkv_bias (`bool`, *optional*, defaults to `True`):
+            Whether or not a learnable bias should be added to the queries, keys and values.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.0):
+            The dropout probability for all fully connected layers in the embeddings and encoder.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        drop_path_rate (`float`, *optional*, defaults to 0.1):
+            Stochastic depth rate.
+        hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder. If string, `"gelu"`, `"relu"`,
+            `"selu"` and `"gelu_new"` are supported.
+        use_absolute_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether or not to add absolute position embeddings to the patch embeddings.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-05):
+            The epsilon used by the layer normalization layers.
+
+    Example:
+
+    ```python
+    >>> from transformers import UnimerSwinConfig, UnimerSwinModel
+
+    >>> # Initializing a Donut naver-clova-ix/donut-base style configuration
+    >>> configuration = UnimerSwinConfig()
+
+    >>> # Randomly initializing a model from the naver-clova-ix/donut-base style configuration
+    >>> model = UnimerSwinModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "unimer-swin"
+
+    attribute_map = {
+        "num_attention_heads": "num_heads",
+        "num_hidden_layers": "num_layers",
+    }
+
+    def __init__(
+        self,
+        image_size=224,
+        patch_size=4,
+        num_channels=3,
+        embed_dim=96,
+        depths=[2, 2, 6, 2],
+        num_heads=[3, 6, 12, 24],
+        window_size=7,
+        mlp_ratio=4.0,
+        qkv_bias=True,
+        hidden_dropout_prob=0.0,
+        attention_probs_dropout_prob=0.0,
+        drop_path_rate=0.1,
+        hidden_act="gelu",
+        use_absolute_embeddings=False,
+        initializer_range=0.02,
+        layer_norm_eps=1e-5,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+        self.embed_dim = embed_dim
+        self.depths = depths
+        self.num_layers = len(depths)
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.mlp_ratio = mlp_ratio
+        self.qkv_bias = qkv_bias
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.drop_path_rate = drop_path_rate
+        self.hidden_act = hidden_act
+        self.use_absolute_embeddings = use_absolute_embeddings
+        self.layer_norm_eps = layer_norm_eps
+        self.initializer_range = initializer_range
+        # we set the hidden_size attribute in order to make Swin work with VisionEncoderDecoderModel
+        # this indicates the channel dimension after the last stage of the model
+        self.hidden_size = int(embed_dim * 2 ** (len(depths) - 1))

magic_pdf/model/sub_modules/mfr/unimernet/unimernet_hf/unimer_swin/image_processing_unimer_swin.py

+from transformers.image_processing_utils import BaseImageProcessor
+import numpy as np
+import cv2
+import albumentations as alb
+from albumentations.pytorch import ToTensorV2
+
+
+# TODO: dereference cv2 if possible
+class UnimerSwinImageProcessor(BaseImageProcessor):
+    def __init__(
+            self,
+            image_size = (192, 672),
+        ):
+        self.input_size = [int(_) for _ in image_size]
+        assert len(self.input_size) == 2
+    
+        self.transform = alb.Compose(
+            [
+                alb.ToGray(always_apply=True),
+                alb.Normalize((0.7931, 0.7931, 0.7931), (0.1738, 0.1738, 0.1738)),
+                # alb.Sharpen()
+                ToTensorV2(),
+            ]
+        )
+
+    def __call__(self, item):
+        image = self.prepare_input(item)
+        return self.transform(image=image)['image'][:1]
+
+    @staticmethod
+    def crop_margin_numpy(img: np.ndarray) -> np.ndarray:
+        """Crop margins of image using NumPy operations"""
+        # Convert to grayscale if it's a color image
+        if len(img.shape) == 3 and img.shape[2] == 3:
+            gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
+        else:
+            gray = img.copy()
+
+        # Normalize and threshold
+        if gray.max() == gray.min():
+            return img
+
+        normalized = (((gray - gray.min()) / (gray.max() - gray.min())) * 255).astype(np.uint8)
+        binary = 255 * (normalized < 200).astype(np.uint8)
+
+        # Find bounding box
+        coords = cv2.findNonZero(binary)  # Find all non-zero points (text)
+        x, y, w, h = cv2.boundingRect(coords)  # Find minimum spanning bounding box
+
+        # Return cropped image
+        return img[y:y + h, x:x + w]
+
+    def prepare_input(self, img, random_padding: bool = False):
+        """
+        Convert PIL Image or numpy array to properly sized and padded image after:
+            - crop margins
+            - resize while maintaining aspect ratio
+            - pad to target size
+        """
+        if img is None:
+            return None
+
+        try:
+            img = self.crop_margin_numpy(img)
+        except Exception:
+            # might throw an error for broken files
+            return None
+
+        if img.shape[0] == 0 or img.shape[1] == 0:
+            return None
+
+        # Resize while preserving aspect ratio
+        h, w = img.shape[:2]
+        scale = min(self.input_size[0] / h, self.input_size[1] / w)
+        new_h, new_w = int(h * scale), int(w * scale)
+        resized_img = cv2.resize(img, (new_w, new_h), interpolation=cv2.INTER_AREA)
+
+        # Calculate padding
+        pad_width, pad_height = self._get_padding_values(new_w, new_h, random_padding)
+
+        # Create and apply padding
+        channels = 3 if len(img.shape) == 3 else 1
+        padded_img = np.full((self.input_size[0], self.input_size[1], channels), 255, dtype=np.uint8)
+        padded_img[pad_height:pad_height + new_h, pad_width:pad_width + new_w] = resized_img
+
+        return padded_img
+
+    def _calculate_padding(self, new_w, new_h, random_padding):
+        """Calculate padding values for PIL images"""
+        delta_width = self.input_size[1] - new_w
+        delta_height = self.input_size[0] - new_h
+
+        pad_width, pad_height = self._get_padding_values(new_w, new_h, random_padding)
+
+        return (
+            pad_width,
+            pad_height,
+            delta_width - pad_width,
+            delta_height - pad_height,
+        )
+
+    def _get_padding_values(self, new_w, new_h, random_padding):
+        """Get padding values based on image dimensions and padding strategy"""
+        delta_width = self.input_size[1] - new_w
+        delta_height = self.input_size[0] - new_h
+
+        if random_padding:
+            pad_width = np.random.randint(low=0, high=delta_width + 1)
+            pad_height = np.random.randint(low=0, high=delta_height + 1)
+        else:
+            pad_width = delta_width // 2
+            pad_height = delta_height // 2
+
+        return pad_width, pad_height

magic_pdf/model/sub_modules/mfr/unimernet/unimernet_hf/unimer_swin/modeling_unimer_swin.py

+# coding=utf-8
+# Copyright 2022 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 UnimerSwin Transformer model.
+
+This implementation is identical to a regular Swin Transformer, without final layer norm on top of the final hidden
+states."""
+
+import collections.abc
+import math
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+
+from transformers.activations import ACT2FN
+from transformers.modeling_utils import PreTrainedModel
+from transformers.pytorch_utils import find_pruneable_heads_and_indices, meshgrid, prune_linear_layer
+from transformers.utils import (
+    ModelOutput,
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    torch_int,
+)
+from .configuration_unimer_swin import UnimerSwinConfig
+
+
+logger = logging.get_logger(__name__)
+
+# General docstring
+_CONFIG_FOR_DOC = "UnimerSwinConfig"
+
+# Base docstring
+_CHECKPOINT_FOR_DOC = "https://huggingface.co/naver-clova-ix/donut-base"
+_EXPECTED_OUTPUT_SHAPE = [1, 49, 768]
+
+
+@dataclass
+# Copied from transformers.models.swin.modeling_swin.SwinEncoderOutput with Swin->UnimerSwin
+class UnimerSwinEncoderOutput(ModelOutput):
+    """
+    UnimerSwin encoder's outputs, with potential hidden states and attentions.
+
+    Args:
+        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each stage) of
+            shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each stage) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+        reshaped_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each stage) of
+            shape `(batch_size, hidden_size, height, width)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs reshaped to
+            include the spatial dimensions.
+    """
+
+    last_hidden_state: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
+    attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
+    reshaped_hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
+
+
+@dataclass
+# Copied from transformers.models.swin.modeling_swin.SwinModelOutput with Swin->UnimerSwin
+class UnimerSwinModelOutput(ModelOutput):
+    """
+    UnimerSwin model's outputs that also contains a pooling of the last hidden states.
+
+    Args:
+        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        pooler_output (`torch.FloatTensor` of shape `(batch_size, hidden_size)`, *optional*, returned when `add_pooling_layer=True` is passed):
+            Average pooling of the last layer hidden-state.
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each stage) of
+            shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each stage) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+        reshaped_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each stage) of
+            shape `(batch_size, hidden_size, height, width)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs reshaped to
+            include the spatial dimensions.
+    """
+
+    last_hidden_state: torch.FloatTensor = None
+    pooler_output: Optional[torch.FloatTensor] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
+    attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
+    reshaped_hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
+
+
+# Copied from transformers.models.swin.modeling_swin.window_partition
+def window_partition(input_feature, window_size):
+    """
+    Partitions the given input into windows.
+    """
+    batch_size, height, width, num_channels = input_feature.shape
+    input_feature = input_feature.view(
+        batch_size, height // window_size, window_size, width // window_size, window_size, num_channels
+    )
+    windows = input_feature.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, num_channels)
+    return windows
+
+
+# Copied from transformers.models.swin.modeling_swin.window_reverse
+def window_reverse(windows, window_size, height, width):
+    """
+    Merges windows to produce higher resolution features.
+    """
+    num_channels = windows.shape[-1]
+    windows = windows.view(-1, height // window_size, width // window_size, window_size, window_size, num_channels)
+    windows = windows.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, height, width, num_channels)
+    return windows
+
+
+# Copied from transformers.models.swin.modeling_swin.SwinEmbeddings with Swin->UnimerSwin
+class UnimerSwinEmbeddings(nn.Module):
+    """
+    Construct the patch and position embeddings. Optionally, also the mask token.
+    """
+
+    def __init__(self, config, use_mask_token=False):
+        super().__init__()
+
+        self.patch_embeddings = UnimerSwinPatchEmbeddings(config)
+        num_patches = self.patch_embeddings.num_patches
+        self.patch_grid = self.patch_embeddings.grid_size
+        self.mask_token = nn.Parameter(torch.zeros(1, 1, config.embed_dim)) if use_mask_token else None
+
+        if config.use_absolute_embeddings:
+            self.position_embeddings = nn.Parameter(torch.zeros(1, num_patches + 1, config.embed_dim))
+        else:
+            self.position_embeddings = None
+
+        ### code added. ###
+        if config.use_2d_embeddings:
+            self.row_embeddings = nn.Parameter(torch.zeros(1, self.patch_grid[0] + 1, config.embed_dim))
+            self.column_embeddings = nn.Parameter(torch.zeros(1, self.patch_grid[1] + 1, config.embed_dim))
+        else:
+            self.row_embeddings = None
+            self.column_embeddings = None
+        ######
+
+        self.norm = nn.LayerNorm(config.embed_dim)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def interpolate_pos_encoding(self, embeddings: torch.Tensor, height: int, width: int) -> torch.Tensor:
+        """
+        This method allows to interpolate the pre-trained position encodings, to be able to use the model on higher
+        resolution images.
+
+        Source:
+        https://github.com/facebookresearch/dino/blob/de9ee3df6cf39fac952ab558447af1fa1365362a/vision_transformer.py#L174
+        """
+
+        num_patches = embeddings.shape[1] - 1
+        num_positions = self.position_embeddings.shape[1] - 1
+        if num_patches == num_positions and height == width:
+            return self.position_embeddings
+        class_pos_embed = self.position_embeddings[:, 0]
+        patch_pos_embed = self.position_embeddings[:, 1:]
+        dim = embeddings.shape[-1]
+        h0 = height // self.config.patch_size
+        w0 = width // self.config.patch_size
+        # we add a small number to avoid floating point error in the interpolation
+        # see discussion at https://github.com/facebookresearch/dino/issues/8
+        h0, w0 = h0 + 0.1, w0 + 0.1
+        patch_pos_embed = patch_pos_embed.reshape(1, int(math.sqrt(num_positions)), int(math.sqrt(num_positions)), dim)
+        patch_pos_embed = patch_pos_embed.permute(0, 3, 1, 2)
+        patch_pos_embed = nn.functional.interpolate(
+            patch_pos_embed,
+            scale_factor=(h0 / math.sqrt(num_positions), w0 / math.sqrt(num_positions)),
+            mode="bicubic",
+            align_corners=False,
+        )
+        patch_pos_embed = patch_pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
+        return torch.cat((class_pos_embed.unsqueeze(0), patch_pos_embed), dim=1)
+
+    def forward(
+        self,
+        pixel_values: Optional[torch.FloatTensor],
+        bool_masked_pos: Optional[torch.BoolTensor] = None,
+        interpolate_pos_encoding: bool = False,
+    ) -> Tuple[torch.Tensor]:
+        _, num_channels, height, width = pixel_values.shape
+        embeddings, output_dimensions = self.patch_embeddings(pixel_values)
+        embeddings = self.norm(embeddings)
+        batch_size, seq_len, _ = embeddings.size()
+
+        if bool_masked_pos is not None:
+            mask_tokens = self.mask_token.expand(batch_size, seq_len, -1)
+            # replace the masked visual tokens by mask_tokens
+            mask = bool_masked_pos.unsqueeze(-1).type_as(mask_tokens)
+            embeddings = embeddings * (1.0 - mask) + mask_tokens * mask
+
+        if self.position_embeddings is not None:
+            # if interpolate_pos_encoding:
+            #     embeddings = embeddings + self.interpolate_pos_encoding(embeddings, height, width)
+            # else:
+            #     embeddings = embeddings + self.position_embeddings
+            embeddings = embeddings + self.position_embeddings[:, :seq_len, :] # code edited.
+
+        ### code added. ###
+        if self.row_embeddings is not None and self.column_embeddings is not None:
+            # Repeat the x position embeddings across the y axis like 0, 1, 2, 3, 0, 1, 2, 3, ...
+            row_embeddings = self.row_embeddings[:, :output_dimensions[0], :].repeat_interleave(output_dimensions[1], dim=1)
+            column_embeddings = self.column_embeddings[:, :output_dimensions[1], :].repeat(1, output_dimensions[0], 1)
+            embeddings = embeddings + row_embeddings + column_embeddings
+        ######
+
+        embeddings = self.dropout(embeddings)
+
+        return embeddings, output_dimensions
+
+class StemLayer(nn.Module):
+    r""" Stem layer of InternImage
+    Args:
+        in_chans (int): number of input channels
+        out_chans (int): number of output channels
+        act_layer (str): activation layer
+        norm_layer (str): normalization layer
+    """
+
+    def __init__(self, in_chans=3, out_chans=96, act_layer=nn.GELU, norm_layer='BN'):
+        super().__init__()
+        self.conv1 = nn.Conv2d(in_chans, out_chans // 2, kernel_size=3, stride=2, padding=1)
+        self.norm1 = self.build_norm_layer(out_chans // 2, norm_layer)
+        self.act = act_layer()
+        self.conv2 = nn.Conv2d(out_chans // 2, out_chans, kernel_size=3, stride=2, padding=1)
+
+    def build_norm_layer(self, dim, norm_layer):
+        layers = []
+        if norm_layer == 'BN':
+            layers.append(nn.BatchNorm2d(dim))
+        else:
+            raise NotImplementedError(f'build_norm_layer does not support {norm_layer}')
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.norm1(x)
+        x = self.act(x)
+        x = self.conv2(x)
+        return x
+
+# Copied from transformers.models.swin.modeling_swin.SwinPatchEmbeddings with Swin->UnimerSwin
+class UnimerSwinPatchEmbeddings(nn.Module):
+    """
+    This class turns `pixel_values` of shape `(batch_size, num_channels, height, width)` into the initial
+    `hidden_states` (patch embeddings) of shape `(batch_size, seq_length, hidden_size)` to be consumed by a
+    Transformer.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        image_size, patch_size = config.image_size, config.patch_size
+        num_channels, hidden_size = config.num_channels, config.embed_dim
+        image_size = image_size if isinstance(image_size, collections.abc.Iterable) else (image_size, image_size)
+        patch_size = patch_size if isinstance(patch_size, collections.abc.Iterable) else (patch_size, patch_size)
+        num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+        self.num_patches = num_patches
+        self.grid_size = (image_size[0] // patch_size[0], image_size[1] // patch_size[1])
+
+        ### code edited. ###
+        # self.projection = nn.Conv2d(num_channels, hidden_size, kernel_size=patch_size, stride=patch_size)
+        self.projection = StemLayer(in_chans=num_channels, out_chans=hidden_size)
+        ###
+
+    def maybe_pad(self, pixel_values, height, width):
+        if width % self.patch_size[1] != 0:
+            pad_values = (0, self.patch_size[1] - width % self.patch_size[1])
+            pixel_values = nn.functional.pad(pixel_values, pad_values)
+        if height % self.patch_size[0] != 0:
+            pad_values = (0, 0, 0, self.patch_size[0] - height % self.patch_size[0])
+            pixel_values = nn.functional.pad(pixel_values, pad_values)
+        return pixel_values
+
+    def forward(self, pixel_values: Optional[torch.FloatTensor]) -> Tuple[torch.Tensor, Tuple[int]]:
+        _, num_channels, height, width = pixel_values.shape
+        # pad the input to be divisible by self.patch_size, if needed
+        pixel_values = self.maybe_pad(pixel_values, height, width)
+        embeddings = self.projection(pixel_values)
+        _, _, height, width = embeddings.shape
+        output_dimensions = (height, width)
+        embeddings = embeddings.flatten(2).transpose(1, 2)
+
+        return embeddings, output_dimensions
+
+
+# Copied from transformers.models.swin.modeling_swin.SwinPatchMerging
+class UnimerSwinPatchMerging(nn.Module):
+    """
+    Patch Merging Layer.
+
+    Args:
+        input_resolution (`Tuple[int]`):
+            Resolution of input feature.
+        dim (`int`):
+            Number of input channels.
+        norm_layer (`nn.Module`, *optional*, defaults to `nn.LayerNorm`):
+            Normalization layer class.
+    """
+
+    def __init__(self, input_resolution: Tuple[int], dim: int, norm_layer: nn.Module = nn.LayerNorm) -> None:
+        super().__init__()
+        self.input_resolution = input_resolution
+        self.dim = dim
+        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
+        self.norm = norm_layer(4 * dim)
+
+    def maybe_pad(self, input_feature, height, width):
+        should_pad = (height % 2 == 1) or (width % 2 == 1)
+        if should_pad:
+            pad_values = (0, 0, 0, width % 2, 0, height % 2)
+            input_feature = nn.functional.pad(input_feature, pad_values)
+
+        return input_feature
+
+    def forward(self, input_feature: torch.Tensor, input_dimensions: Tuple[int, int]) -> torch.Tensor:
+        height, width = input_dimensions
+        # `dim` is height * width
+        batch_size, dim, num_channels = input_feature.shape
+
+        input_feature = input_feature.view(batch_size, height, width, num_channels)
+        # pad input to be disible by width and height, if needed
+        input_feature = self.maybe_pad(input_feature, height, width)
+        # [batch_size, height/2, width/2, num_channels]
+        input_feature_0 = input_feature[:, 0::2, 0::2, :]
+        # [batch_size, height/2, width/2, num_channels]
+        input_feature_1 = input_feature[:, 1::2, 0::2, :]
+        # [batch_size, height/2, width/2, num_channels]
+        input_feature_2 = input_feature[:, 0::2, 1::2, :]
+        # [batch_size, height/2, width/2, num_channels]
+        input_feature_3 = input_feature[:, 1::2, 1::2, :]
+        # batch_size height/2 width/2 4*num_channels
+        input_feature = torch.cat([input_feature_0, input_feature_1, input_feature_2, input_feature_3], -1)
+        input_feature = input_feature.view(batch_size, -1, 4 * num_channels)  # batch_size height/2*width/2 4*C
+
+        input_feature = self.norm(input_feature)
+        input_feature = self.reduction(input_feature)
+
+        return input_feature
+
+
+# Copied from transformers.models.beit.modeling_beit.drop_path
+def drop_path(input: torch.Tensor, drop_prob: float = 0.0, training: bool = False) -> torch.Tensor:
+    """
+    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.swin.modeling_swin.SwinDropPath
+class UnimerSwinDropPath(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)
+
+
+# Copied from transformers.models.swin.modeling_swin.SwinSelfAttention with Swin->UnimerSwin
+class UnimerSwinSelfAttention(nn.Module):
+    def __init__(self, config, dim, num_heads, window_size):
+        super().__init__()
+        if dim % num_heads != 0:
+            raise ValueError(
+                f"The hidden size ({dim}) is not a multiple of the number of attention heads ({num_heads})"
+            )
+
+        self.num_attention_heads = num_heads
+        self.attention_head_size = int(dim / num_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+        self.window_size = (
+            window_size if isinstance(window_size, collections.abc.Iterable) else (window_size, window_size)
+        )
+
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros((2 * self.window_size[0] - 1) * (2 * self.window_size[1] - 1), num_heads)
+        )
+
+        # get pair-wise relative position index for each token inside the window
+        coords_h = torch.arange(self.window_size[0])
+        coords_w = torch.arange(self.window_size[1])
+        coords = torch.stack(meshgrid([coords_h, coords_w], indexing="ij"))
+        coords_flatten = torch.flatten(coords, 1)
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()
+        relative_coords[:, :, 0] += self.window_size[0] - 1
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)
+        self.register_buffer("relative_position_index", relative_position_index)
+
+        self.query = nn.Linear(self.all_head_size, self.all_head_size, bias=config.qkv_bias)
+        self.key = nn.Linear(self.all_head_size, self.all_head_size, bias=config.qkv_bias)
+        self.value = nn.Linear(self.all_head_size, self.all_head_size, bias=config.qkv_bias)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+
+    def transpose_for_scores(self, x):
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        x = x.view(new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor]:
+        batch_size, dim, num_channels = hidden_states.shape
+        mixed_query_layer = self.query(hidden_states)
+
+        key_layer = self.transpose_for_scores(self.key(hidden_states))
+        value_layer = self.transpose_for_scores(self.value(hidden_states))
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+
+        relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)]
+        relative_position_bias = relative_position_bias.view(
+            self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1
+        )
+
+        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()
+        attention_scores = attention_scores + relative_position_bias.unsqueeze(0)
+
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in UnimerSwinModel forward() function)
+            mask_shape = attention_mask.shape[0]
+            attention_scores = attention_scores.view(
+                batch_size // mask_shape, mask_shape, self.num_attention_heads, dim, dim
+            )
+            attention_scores = attention_scores + attention_mask.unsqueeze(1).unsqueeze(0)
+            attention_scores = attention_scores.view(-1, self.num_attention_heads, dim, dim)
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_layer)
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(new_context_layer_shape)
+
+        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
+
+        return outputs
+
+
+# Copied from transformers.models.swin.modeling_swin.SwinSelfOutput
+class UnimerSwinSelfOutput(nn.Module):
+    def __init__(self, config, dim):
+        super().__init__()
+        self.dense = nn.Linear(dim, dim)
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+
+        return hidden_states
+
+
+# Copied from transformers.models.swin.modeling_swin.SwinAttention with Swin->UnimerSwin
+class UnimerSwinAttention(nn.Module):
+    def __init__(self, config, dim, num_heads, window_size):
+        super().__init__()
+        self.self = UnimerSwinSelfAttention(config, dim, num_heads, window_size)
+        self.output = UnimerSwinSelfOutput(config, dim)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads, self.self.num_attention_heads, self.self.attention_head_size, self.pruned_heads
+        )
+
+        # Prune linear layers
+        self.self.query = prune_linear_layer(self.self.query, index)
+        self.self.key = prune_linear_layer(self.self.key, index)
+        self.self.value = prune_linear_layer(self.self.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
+        self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor]:
+        self_outputs = self.self(hidden_states, attention_mask, head_mask, output_attentions)
+        attention_output = self.output(self_outputs[0], hidden_states)
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+# Copied from transformers.models.swin.modeling_swin.SwinIntermediate
+class UnimerSwinIntermediate(nn.Module):
+    def __init__(self, config, dim):
+        super().__init__()
+        self.dense = nn.Linear(dim, int(config.mlp_ratio * dim))
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.swin.modeling_swin.SwinOutput
+class UnimerSwinOutput(nn.Module):
+    def __init__(self, config, dim):
+        super().__init__()
+        self.dense = nn.Linear(int(config.mlp_ratio * dim), dim)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        return hidden_states
+
+
+class ConvEnhance(nn.Module):
+    """Depth-wise convolution to get the positional information.
+    """
+    def __init__(self, config, dim, k=3):
+        super(ConvEnhance, self).__init__()
+        self.proj = nn.Conv2d(dim,
+                              dim,
+                              (k,k),
+                              (1,1),
+                              (k // 2,k // 2),
+                              groups=dim)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x, size: Tuple[int, int]):
+        B, N, C = x.shape
+        H, W = size
+        assert N == H * W
+
+        feat = x.transpose(1, 2).view(B, C, H, W)
+        feat = self.proj(feat)
+        feat = self.act_fn(feat)
+        feat = feat.flatten(2).transpose(1, 2)
+
+        x = x + feat
+        return x
+
+
+# Copied from transformers.models.swin.modeling_swin.SwinLayer with Swin->UnimerSwin
+class UnimerSwinLayer(nn.Module):
+    def __init__(self, config, dim, input_resolution, num_heads, shift_size=0):
+        super().__init__()
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.shift_size = shift_size
+        self.window_size = config.window_size
+        self.input_resolution = input_resolution
+        self.layernorm_before = nn.LayerNorm(dim, eps=config.layer_norm_eps)
+
+        self.ce = nn.ModuleList([ConvEnhance(config, dim=dim, k=3),
+                                  ConvEnhance(config, dim=dim, k=3)])
+
+        self.attention = UnimerSwinAttention(config, dim, num_heads, window_size=self.window_size)
+        self.drop_path = UnimerSwinDropPath(config.drop_path_rate) if config.drop_path_rate > 0.0 else nn.Identity()
+        self.layernorm_after = nn.LayerNorm(dim, eps=config.layer_norm_eps)
+        self.intermediate = UnimerSwinIntermediate(config, dim)
+        self.output = UnimerSwinOutput(config, dim)
+
+    def set_shift_and_window_size(self, input_resolution):
+        if min(input_resolution) <= self.window_size:
+            # if window size is larger than input resolution, we don't partition windows
+            self.shift_size = torch_int(0)
+            self.window_size = (
+                torch.min(torch.tensor(input_resolution)) if torch.jit.is_tracing() else min(input_resolution)
+            )
+
+    def get_attn_mask(self, height, width, dtype, device):
+        if self.shift_size > 0:
+            # calculate attention mask for SW-MSA
+            img_mask = torch.zeros((1, height, width, 1), dtype=dtype, device=device)
+            height_slices = (
+                slice(0, -self.window_size),
+                slice(-self.window_size, -self.shift_size),
+                slice(-self.shift_size, None),
+            )
+            width_slices = (
+                slice(0, -self.window_size),
+                slice(-self.window_size, -self.shift_size),
+                slice(-self.shift_size, None),
+            )
+            count = 0
+            for height_slice in height_slices:
+                for width_slice in width_slices:
+                    img_mask[:, height_slice, width_slice, :] = count
+                    count += 1
+
+            mask_windows = window_partition(img_mask, self.window_size)
+            mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+            attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+            attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
+        else:
+            attn_mask = None
+        return attn_mask
+
+    def maybe_pad(self, hidden_states, height, width):
+        pad_right = (self.window_size - width % self.window_size) % self.window_size
+        pad_bottom = (self.window_size - height % self.window_size) % self.window_size
+        pad_values = (0, 0, 0, pad_right, 0, pad_bottom)
+        hidden_states = nn.functional.pad(hidden_states, pad_values)
+        return hidden_states, pad_values
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        input_dimensions: Tuple[int, int],
+        head_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = False,
+        always_partition: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        if not always_partition:
+            self.set_shift_and_window_size(input_dimensions)
+        else:
+            pass
+        height, width = input_dimensions
+        batch_size, _, channels = hidden_states.size()
+        
+
+
+        hidden_states = self.ce[0](hidden_states, input_dimensions)
+        shortcut = hidden_states
+
+
+        hidden_states = self.layernorm_before(hidden_states)
+        hidden_states = hidden_states.view(batch_size, height, width, channels)
+
+        # pad hidden_states to multiples of window size
+        hidden_states, pad_values = self.maybe_pad(hidden_states, height, width)
+
+        _, height_pad, width_pad, _ = hidden_states.shape
+        # cyclic shift
+        if self.shift_size > 0:
+            shifted_hidden_states = torch.roll(hidden_states, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
+        else:
+            shifted_hidden_states = hidden_states
+
+        # partition windows
+        hidden_states_windows = window_partition(shifted_hidden_states, self.window_size)
+        hidden_states_windows = hidden_states_windows.view(-1, self.window_size * self.window_size, channels)
+        attn_mask = self.get_attn_mask(
+            height_pad, width_pad, dtype=hidden_states.dtype, device=hidden_states_windows.device
+        )
+
+        attention_outputs = self.attention(
+            hidden_states_windows, attn_mask, head_mask, output_attentions=output_attentions
+        )
+
+        attention_output = attention_outputs[0]
+
+        attention_windows = attention_output.view(-1, self.window_size, self.window_size, channels)
+        shifted_windows = window_reverse(attention_windows, self.window_size, height_pad, width_pad)
+
+        # reverse cyclic shift
+        if self.shift_size > 0:
+            attention_windows = torch.roll(shifted_windows, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
+        else:
+            attention_windows = shifted_windows
+
+        was_padded = pad_values[3] > 0 or pad_values[5] > 0
+        if was_padded:
+            attention_windows = attention_windows[:, :height, :width, :].contiguous()
+
+        attention_windows = attention_windows.view(batch_size, height * width, channels)
+
+        hidden_states = shortcut + self.drop_path(attention_windows)
+
+
+
+        hidden_states = self.ce[1](hidden_states, input_dimensions)
+        layer_output = self.layernorm_after(hidden_states)
+        layer_output = self.intermediate(layer_output)
+        layer_output = hidden_states + self.output(layer_output)
+
+        layer_outputs = (layer_output, attention_outputs[1]) if output_attentions else (layer_output,)
+        return layer_outputs
+
+
+# Copied from transformers.models.swin.modeling_swin.SwinStage with Swin->UnimerSwin
+class UnimerSwinStage(nn.Module):
+    def __init__(self, config, dim, input_resolution, depth, num_heads, drop_path, downsample):
+        super().__init__()
+        self.config = config
+        self.dim = dim
+        self.blocks = nn.ModuleList(
+            [
+                UnimerSwinLayer(
+                    config=config,
+                    dim=dim,
+                    input_resolution=input_resolution,
+                    num_heads=num_heads,
+                    shift_size=0,
+                )
+                for i in range(depth)
+            ]
+        )
+
+        # patch merging layer
+        if downsample is not None:
+            self.downsample = downsample(input_resolution, dim=dim, norm_layer=nn.LayerNorm)
+        else:
+            self.downsample = None
+
+        self.pointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        input_dimensions: Tuple[int, int],
+        head_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = False,
+        always_partition: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor]:
+        height, width = input_dimensions
+        for i, layer_module in enumerate(self.blocks):
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+
+            layer_outputs = layer_module(
+                hidden_states, input_dimensions, layer_head_mask, output_attentions, always_partition
+            )
+
+            hidden_states = layer_outputs[0]
+
+        hidden_states_before_downsampling = hidden_states
+        if self.downsample is not None:
+            height_downsampled, width_downsampled = (height + 1) // 2, (width + 1) // 2
+            output_dimensions = (height, width, height_downsampled, width_downsampled)
+            hidden_states = self.downsample(hidden_states_before_downsampling, input_dimensions)
+        else:
+            output_dimensions = (height, width, height, width)
+
+        stage_outputs = (hidden_states, hidden_states_before_downsampling, output_dimensions)
+
+        if output_attentions:
+            stage_outputs += layer_outputs[1:]
+        return stage_outputs
+
+
+# Copied from transformers.models.swin.modeling_swin.SwinEncoder with Swin->UnimerSwin
+class UnimerSwinEncoder(nn.Module):
+    def __init__(self, config, grid_size):
+        super().__init__()
+        self.num_layers = len(config.depths)
+        self.config = config
+        dpr = [x.item() for x in torch.linspace(0, config.drop_path_rate, sum(config.depths))]
+        self.layers = nn.ModuleList(
+            [
+                UnimerSwinStage(
+                    config=config,
+                    dim=int(config.embed_dim * 2**i_layer),
+                    input_resolution=(grid_size[0] // (2**i_layer), grid_size[1] // (2**i_layer)),
+                    depth=config.depths[i_layer],
+                    num_heads=config.num_heads[i_layer],
+                    drop_path=dpr[sum(config.depths[:i_layer]) : sum(config.depths[: i_layer + 1])],
+                    downsample=UnimerSwinPatchMerging if (i_layer < self.num_layers - 1) else None,
+                )
+                for i_layer in range(self.num_layers)
+            ]
+        )
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        input_dimensions: Tuple[int, int],
+        head_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = False,
+        output_hidden_states: Optional[bool] = False,
+        output_hidden_states_before_downsampling: Optional[bool] = False,
+        always_partition: Optional[bool] = False,
+        return_dict: Optional[bool] = True,
+    ) -> Union[Tuple, UnimerSwinEncoderOutput]:
+        all_hidden_states = () if output_hidden_states else None
+        all_reshaped_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+
+        if output_hidden_states:
+            batch_size, _, hidden_size = hidden_states.shape
+            # rearrange b (h w) c -> b c h w
+            reshaped_hidden_state = hidden_states.view(batch_size, *input_dimensions, hidden_size)
+            reshaped_hidden_state = reshaped_hidden_state.permute(0, 3, 1, 2)
+            all_hidden_states += (hidden_states,)
+            all_reshaped_hidden_states += (reshaped_hidden_state,)
+
+        for i, layer_module in enumerate(self.layers):
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    layer_module.__call__,
+                    hidden_states,
+                    input_dimensions,
+                    layer_head_mask,
+                    output_attentions,
+                    always_partition,
+                )
+            else:
+                layer_outputs = layer_module(
+                    hidden_states, input_dimensions, layer_head_mask, output_attentions, always_partition
+                )
+
+            hidden_states = layer_outputs[0]
+            hidden_states_before_downsampling = layer_outputs[1]
+            output_dimensions = layer_outputs[2]
+
+            input_dimensions = (output_dimensions[-2], output_dimensions[-1])
+
+            if output_hidden_states and output_hidden_states_before_downsampling:
+                batch_size, _, hidden_size = hidden_states_before_downsampling.shape
+                # rearrange b (h w) c -> b c h w
+                # here we use the original (not downsampled) height and width
+                reshaped_hidden_state = hidden_states_before_downsampling.view(
+                    batch_size, *(output_dimensions[0], output_dimensions[1]), hidden_size
+                )
+                reshaped_hidden_state = reshaped_hidden_state.permute(0, 3, 1, 2)
+                all_hidden_states += (hidden_states_before_downsampling,)
+                all_reshaped_hidden_states += (reshaped_hidden_state,)
+            elif output_hidden_states and not output_hidden_states_before_downsampling:
+                batch_size, _, hidden_size = hidden_states.shape
+                # rearrange b (h w) c -> b c h w
+                reshaped_hidden_state = hidden_states.view(batch_size, *input_dimensions, hidden_size)
+                reshaped_hidden_state = reshaped_hidden_state.permute(0, 3, 1, 2)
+                all_hidden_states += (hidden_states,)
+                all_reshaped_hidden_states += (reshaped_hidden_state,)
+
+            if output_attentions:
+                all_self_attentions += layer_outputs[3:]
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, all_hidden_states, all_self_attentions] if v is not None)
+
+        return UnimerSwinEncoderOutput(
+            last_hidden_state=hidden_states,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+            reshaped_hidden_states=all_reshaped_hidden_states,
+        )
+
+
+# Copied from transformers.models.swin.modeling_swin.SwinPreTrainedModel with Swin->UnimerSwin
+class UnimerSwinPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = UnimerSwinConfig
+    base_model_prefix = "unimer-swin"
+    main_input_name = "pixel_values"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["UnimerSwinStage"]
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, (nn.Linear, nn.Conv2d)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+
+SWIN_START_DOCSTRING = r"""
+    This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) sub-class. Use
+    it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and
+    behavior.
+
+    Parameters:
+        config ([`UnimerSwinConfig`]): 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.
+"""
+
+SWIN_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
+            [`DonutImageProcessor.__call__`] for details.
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        interpolate_pos_encoding (`bool`, *optional*, defaults to `False`):
+            Whether to interpolate the pre-trained position encodings.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare UnimerSwin Model transformer outputting raw hidden-states without any specific head on top.",
+    SWIN_START_DOCSTRING,
+)
+class UnimerSwinModel(UnimerSwinPreTrainedModel):
+    def __init__(self, config, add_pooling_layer=True, use_mask_token=False):
+        super().__init__(config)
+        self.config = config
+        self.num_layers = len(config.depths)
+        self.num_features = int(config.embed_dim * 2 ** (self.num_layers - 1))
+
+        self.embeddings = UnimerSwinEmbeddings(config, use_mask_token=use_mask_token)
+        self.encoder = UnimerSwinEncoder(config, self.embeddings.patch_grid)
+        self.pooler = nn.AdaptiveAvgPool1d(1) if add_pooling_layer else None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embeddings.patch_embeddings
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    @add_start_docstrings_to_model_forward(SWIN_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=UnimerSwinModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+        modality="vision",
+        expected_output=_EXPECTED_OUTPUT_SHAPE,
+    )
+    def forward(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        bool_masked_pos: Optional[torch.BoolTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        interpolate_pos_encoding: bool = False,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, UnimerSwinModelOutput]:
+        r"""
+        bool_masked_pos (`torch.BoolTensor` of shape `(batch_size, num_patches)`):
+            Boolean masked positions. Indicates which patches are masked (1) and which aren't (0).
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        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
+
+        if pixel_values is None:
+            raise ValueError("You have to specify pixel_values")
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, len(self.config.depths))
+
+        embedding_output, input_dimensions = self.embeddings(
+            pixel_values, bool_masked_pos=bool_masked_pos, interpolate_pos_encoding=interpolate_pos_encoding
+        )
+
+        encoder_outputs = self.encoder(
+            embedding_output,
+            input_dimensions,
+            head_mask=head_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = encoder_outputs[0]
+
+        pooled_output = None
+        if self.pooler is not None:
+            pooled_output = self.pooler(sequence_output.transpose(1, 2))
+            pooled_output = torch.flatten(pooled_output, 1)
+
+        if not return_dict:
+            output = (sequence_output, pooled_output) + encoder_outputs[1:]
+
+            return output
+
+        return UnimerSwinModelOutput(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+            reshaped_hidden_states=encoder_outputs.reshaped_hidden_states,
+        )

magic_pdf/model/sub_modules/model_init.py

-import os
-
 import torch
 from loguru import logger
 
 from magic_pdf.config.constants import MODEL_NAME
 from magic_pdf.model.model_list import AtomicModel
-from magic_pdf.model.sub_modules.language_detection.yolov11.YOLOv11 import \
-    YOLOv11LangDetModel
-from magic_pdf.model.sub_modules.layout.doclayout_yolo.DocLayoutYOLO import \
-    DocLayoutYOLOModel
-from magic_pdf.model.sub_modules.layout.layoutlmv3.model_init import \
-    Layoutlmv3_Predictor
+from magic_pdf.model.sub_modules.language_detection.yolov11.YOLOv11 import YOLOv11LangDetModel
+from magic_pdf.model.sub_modules.layout.doclayout_yolo.DocLayoutYOLO import DocLayoutYOLOModel
 from magic_pdf.model.sub_modules.mfd.yolov8.YOLOv8 import YOLOv8MFDModel
 from magic_pdf.model.sub_modules.mfr.unimernet.Unimernet import UnimernetModel
 
@@ -18,10 +12,8 @@ try:
     from magic_pdf_ascend_plugin.libs.license_verifier import (
         LicenseExpiredError, LicenseFormatError, LicenseSignatureError,
         load_license)
-    from magic_pdf_ascend_plugin.model_plugin.ocr.paddleocr.ppocr_273_npu import \
-        ModifiedPaddleOCR
-    from magic_pdf_ascend_plugin.model_plugin.table.rapidtable.rapid_table_npu import \
-        RapidTableModel
+    from magic_pdf_ascend_plugin.model_plugin.ocr.paddleocr.ppocr_273_npu import ModifiedPaddleOCR
+    from magic_pdf_ascend_plugin.model_plugin.table.rapidtable.rapid_table_npu import RapidTableModel
     license_key = load_license()
     logger.info(f'Using Ascend Plugin Success, License id is {license_key["payload"]["id"]},'
                 f' License expired at {license_key["payload"]["date"]["end_date"]}')
@@ -42,16 +34,13 @@ except Exception as e:
     # from magic_pdf.model.sub_modules.ocr.paddleocr.ppocr_291_mod import ModifiedPaddleOCR
     from magic_pdf.model.sub_modules.table.rapidtable.rapid_table import RapidTableModel
 
-from magic_pdf.model.sub_modules.table.structeqtable.struct_eqtable import \
-    StructTableModel
-from magic_pdf.model.sub_modules.table.tablemaster.tablemaster_paddle import \
-    TableMasterPaddleModel
-
 
 def table_model_init(table_model_type, model_path, max_time, _device_='cpu', ocr_engine=None, table_sub_model_name=None):
     if table_model_type == MODEL_NAME.STRUCT_EQTABLE:
+        from magic_pdf.model.sub_modules.table.structeqtable.struct_eqtable import StructTableModel
         table_model = StructTableModel(model_path, max_new_tokens=2048, max_time=max_time)
     elif table_model_type == MODEL_NAME.TABLE_MASTER:
+        from magic_pdf.model.sub_modules.table.tablemaster.tablemaster_paddle import TableMasterPaddleModel
         config = {
             'model_dir': model_path,
             'device': _device_
@@ -79,6 +68,7 @@ def mfr_model_init(weight_dir, cfg_path, device='cpu'):
 
 
 def layout_model_init(weight, config_file, device):
+    from magic_pdf.model.sub_modules.layout.layoutlmv3.model_init import Layoutlmv3_Predictor
     model = Layoutlmv3_Predictor(weight, config_file, device)
     return model
 

magic_pdf/model/sub_modules/model_utils.py

 import time
-
 import torch
-from PIL import Image
 from loguru import logger
-
+import numpy as np
 from magic_pdf.libs.clean_memory import clean_memory
 
 
-def crop_img(input_res, input_pil_img, crop_paste_x=0, crop_paste_y=0):
+def crop_img(input_res, input_np_img, crop_paste_x=0, crop_paste_y=0):
+
     crop_xmin, crop_ymin = int(input_res['poly'][0]), int(input_res['poly'][1])
     crop_xmax, crop_ymax = int(input_res['poly'][4]), int(input_res['poly'][5])
-    # Create a white background with an additional width and height of 50
+
+    # Calculate new dimensions
     crop_new_width = crop_xmax - crop_xmin + crop_paste_x * 2
     crop_new_height = crop_ymax - crop_ymin + crop_paste_y * 2
-    return_image = Image.new('RGB', (crop_new_width, crop_new_height), 'white')
 
-    # Crop image
-    crop_box = (crop_xmin, crop_ymin, crop_xmax, crop_ymax)
-    cropped_img = input_pil_img.crop(crop_box)
-    return_image.paste(cropped_img, (crop_paste_x, crop_paste_y))
-    return_list = [crop_paste_x, crop_paste_y, crop_xmin, crop_ymin, crop_xmax, crop_ymax, crop_new_width, crop_new_height]
+    # Create a white background array
+    return_image = np.ones((crop_new_height, crop_new_width, 3), dtype=np.uint8) * 255
+
+    # Crop the original image using numpy slicing
+    cropped_img = input_np_img[crop_ymin:crop_ymax, crop_xmin:crop_xmax]
+
+    # Paste the cropped image onto the white background
+    return_image[crop_paste_y:crop_paste_y + (crop_ymax - crop_ymin),
+    crop_paste_x:crop_paste_x + (crop_xmax - crop_xmin)] = cropped_img
+
+    return_list = [crop_paste_x, crop_paste_y, crop_xmin, crop_ymin, crop_xmax, crop_ymax, crop_new_width,
+                   crop_new_height]
     return return_image, return_list
 
 

magic_pdf/pdf_parse_union_core_v2.py

@@ -492,7 +492,7 @@ def insert_lines_into_block(block_bbox, line_height, page_w, page_h):
     else:
         return [[x0, y0, x1, y1]]
 
-# @measure_time
+
 def sort_lines_by_model(fix_blocks, page_w, page_h, line_height):
     page_line_list = []
 

magic_pdf/resources/model_config/model_configs.yaml

@@ -2,7 +2,7 @@ weights:
   layoutlmv3: Layout/LayoutLMv3/model_final.pth
   doclayout_yolo: Layout/YOLO/doclayout_yolo_docstructbench_imgsz1280_2501.pt
   yolo_v8_mfd: MFD/YOLO/yolo_v8_ft.pt
-  unimernet_small: MFR/unimernet_small_2501
+  unimernet_small: MFR/unimernet_hf_small_2503
   struct_eqtable: TabRec/StructEqTable
   tablemaster: TabRec/TableMaster
   rapid_table: TabRec/RapidTable
\ No newline at end of file

requirements.txt

@@ -7,7 +7,8 @@ numpy>=1.21.6,<2.0.0
 pydantic>=2.7.2
 PyMuPDF>=1.24.9,<=1.24.14
 scikit-learn>=1.0.2
-torch>=2.2.2
-transformers
+torch>=2.2.2,!=2.5.0,!=2.5.1,<=2.6.0
+torchvision
+transformers>=4.49.0
 pdfminer.six==20231228
 # The requirements.txt must ensure that only necessary external dependencies are introduced. If there are new dependencies to add, please contact the project administrator.

scripts/download_models.py

 import json
+import shutil
 import os
 
 import requests
@@ -16,7 +17,7 @@ def download_and_modify_json(url, local_filename, modifications):
     if os.path.exists(local_filename):
         data = json.load(open(local_filename))
         config_version = data.get('config_version', '0.0.0')
-        if config_version < '1.1.1':
+        if config_version < '1.2.0':
             data = download_json(url)
     else:
         data = download_json(url)
@@ -32,12 +33,13 @@ def download_and_modify_json(url, local_filename, modifications):
 
 if __name__ == '__main__':
     mineru_patterns = [
-        "models/Layout/LayoutLMv3/*",
+        # "models/Layout/LayoutLMv3/*",
         "models/Layout/YOLO/*",
         "models/MFD/YOLO/*",
-        "models/MFR/unimernet_small_2501/*",
-        "models/TabRec/TableMaster/*",
-        "models/TabRec/StructEqTable/*",
+        "models/MFR/unimernet_hf_small_2503/*",
+        "models/OCR/paddleocr/*",
+        # "models/TabRec/TableMaster/*",
+        # "models/TabRec/StructEqTable/*",
     ]
     model_dir = snapshot_download('opendatalab/PDF-Extract-Kit-1.0', allow_patterns=mineru_patterns)
     layoutreader_model_dir = snapshot_download('ppaanngggg/layoutreader')
@@ -45,6 +47,12 @@ if __name__ == '__main__':
     print(f'model_dir is: {model_dir}')
     print(f'layoutreader_model_dir is: {layoutreader_model_dir}')
 
+    paddleocr_model_dir = model_dir + '/OCR/paddleocr'
+    user_paddleocr_dir = os.path.expanduser('~/.paddleocr')
+    if os.path.exists(user_paddleocr_dir):
+        shutil.rmtree(user_paddleocr_dir)
+    shutil.copytree(paddleocr_model_dir, user_paddleocr_dir)
+
     json_url = 'https://gcore.jsdelivr.net/gh/opendatalab/MinerU@master/magic-pdf.template.json'
     config_file_name = 'magic-pdf.json'
     home_dir = os.path.expanduser('~')

scripts/download_models_hf.py

 import json
 import os
+import shutil
 
 import requests
 from huggingface_hub import snapshot_download
@@ -16,7 +17,7 @@ def download_and_modify_json(url, local_filename, modifications):
     if os.path.exists(local_filename):
         data = json.load(open(local_filename))
         config_version = data.get('config_version', '0.0.0')
-        if config_version < '1.1.1':
+        if config_version < '1.2.0':
             data = download_json(url)
     else:
         data = download_json(url)
@@ -33,12 +34,13 @@ def download_and_modify_json(url, local_filename, modifications):
 if __name__ == '__main__':
 
     mineru_patterns = [
-        "models/Layout/LayoutLMv3/*",
+        # "models/Layout/LayoutLMv3/*",
         "models/Layout/YOLO/*",
         "models/MFD/YOLO/*",
-        "models/MFR/unimernet_small_2501/*",
-        "models/TabRec/TableMaster/*",
-        "models/TabRec/StructEqTable/*",
+        "models/MFR/unimernet_hf_small_2503/*",
+        "models/OCR/paddleocr/*",
+        # "models/TabRec/TableMaster/*",
+        # "models/TabRec/StructEqTable/*",
     ]
     model_dir = snapshot_download('opendatalab/PDF-Extract-Kit-1.0', allow_patterns=mineru_patterns)
 
@@ -52,6 +54,12 @@ if __name__ == '__main__':
     print(f'model_dir is: {model_dir}')
     print(f'layoutreader_model_dir is: {layoutreader_model_dir}')
 
+    paddleocr_model_dir = model_dir + '/OCR/paddleocr'
+    user_paddleocr_dir = os.path.expanduser('~/.paddleocr')
+    if os.path.exists(user_paddleocr_dir):
+        shutil.rmtree(user_paddleocr_dir)
+    shutil.copytree(paddleocr_model_dir, user_paddleocr_dir)
+
     json_url = 'https://github.com/opendatalab/MinerU/raw/master/magic-pdf.template.json'
     config_file_name = 'magic-pdf.json'
     home_dir = os.path.expanduser('~')

setup.py

@@ -36,29 +36,31 @@ if __name__ == '__main__':
                      "paddlepaddle==3.0.0b1;platform_system=='Linux'",
                      "paddlepaddle==2.6.1;platform_system=='Windows' or platform_system=='Darwin'",
                      ],
-            "full": ["unimernet==0.2.3",  # unimernet升级0.2.3,移除torchtext/eva-decord的依赖
-                     "torch>=2.2.2,<=2.3.1",  # torch2.4.0及之后版本未测试，先卡住版本上限
-                     "torchvision>=0.17.2,<=0.18.1",  # torchvision 受torch版本约束
+            "full": [
                      "matplotlib<=3.9.0;platform_system=='Windows'",  # 3.9.1及之后不提供windows的预编译包，避免一些没有编译环境的windows设备安装失败
                      "matplotlib;platform_system=='Linux' or platform_system=='Darwin'",  # linux 和 macos 不应限制matplotlib的最高版本，以避免无法更新导致的一些bug
                      "ultralytics>=8.3.48",  # yolov8,公式检测
                      "paddleocr==2.7.3",  # 2.8.0及2.8.1版本与detectron2有冲突，需锁定2.7.3
                      "paddlepaddle==3.0.0rc1;platform_system=='Linux' or platform_system=='Darwin'",  # 解决linux的段异常问题
                      "paddlepaddle==2.6.1;platform_system=='Windows'",  # windows版本3.0.0效率下降，需锁定2.6.1
-                     "struct-eqtable==0.3.2",  # 表格解析
-                     "einops",  # struct-eqtable依赖
-                     "accelerate",  # struct-eqtable依赖
                      "doclayout_yolo==0.0.2b1",  # doclayout_yolo
                      "rapidocr-paddle>=1.4.5,<2.0.0",  # rapidocr-paddle
                      "rapidocr_onnxruntime>=1.4.4,<2.0.0",
                      "rapid_table>=1.0.3,<2.0.0",  # rapid_table
                      "PyYAML",  # yaml
                      "openai",  # openai SDK
-                     "detectron2"
                      ],
             "old_linux":[
                 "albumentations<=1.4.20", # 1.4.21引入的simsimd不支持2019年及更早的linux系统
-            ]
+            ],
+            "layoutlmv3":[
+                "detectron2"
+            ],
+            "struct_eqtable":[
+                "struct-eqtable==0.3.2",  # 表格解析
+                "einops",  # struct-eqtable依赖
+                "accelerate",  # struct-eqtable依赖
+            ],
         },
         description="A practical tool for converting PDF to Markdown",  # 简短描述
         long_description=long_description,  # 详细描述