Merge pull request #2065 from opendatalab/release-1.3.0

Release 1.3.0

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(),
+                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
+
+        # Get current dimensions
+        h, w = img.shape[:2]
+        target_h, target_w = self.input_size
+
+        # Calculate scale to preserve aspect ratio (equivalent to resize + thumbnail)
+        scale = min(target_h / h, target_w / w)
+
+        # Calculate new dimensions
+        new_h, new_w = int(h * scale), int(w * scale)
+
+        # Resize the image while preserving aspect ratio
+        resized_img = cv2.resize(img, (new_w, new_h))
+
+        # Calculate padding values using the existing method
+        delta_width = target_w - new_w
+        delta_height = target_h - new_h
+
+        pad_width, pad_height = self._get_padding_values(new_w, new_h, random_padding)
+
+        # Apply padding (convert PIL padding format to OpenCV format)
+        padding_color = [0, 0, 0] if len(img.shape) == 3 else [0]
+
+        padded_img = cv2.copyMakeBorder(
+            resized_img,
+            pad_height,  # top
+            delta_height - pad_height,  # bottom
+            pad_width,  # left
+            delta_width - pad_width,  # right
+            cv2.BORDER_CONSTANT,
+            value=padding_color
+        )
+
+        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

@@ -5,47 +5,57 @@ 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.mfd.yolov8.YOLOv8 import YOLOv8MFDModel
 from magic_pdf.model.sub_modules.mfr.unimernet.Unimernet import UnimernetModel
-
-try:
-    from magic_pdf_ascend_plugin.libs.license_verifier import load_license, LicenseFormatError, LicenseSignatureError, LicenseExpiredError
-    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"]}')
-except Exception as e:
-    if isinstance(e, ImportError):
-        pass
-    elif isinstance(e, LicenseFormatError):
-        logger.error("Ascend Plugin: Invalid license format. Please check the license file.")
-    elif isinstance(e, LicenseSignatureError):
-        logger.error("Ascend Plugin: Invalid signature. The license may be tampered with.")
-    elif isinstance(e, LicenseExpiredError):
-        logger.error("Ascend Plugin: License has expired. Please renew your license.")
-    elif isinstance(e, FileNotFoundError):
-        logger.error("Ascend Plugin: Not found License file.")
-    else:
-        logger.error(f"Ascend Plugin: {e}")
-    from magic_pdf.model.sub_modules.ocr.paddleocr.ppocr_273_mod import ModifiedPaddleOCR
-    # 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):
+from magic_pdf.model.sub_modules.ocr.paddleocr2pytorch.pytorch_paddle import PytorchPaddleOCR
+from magic_pdf.model.sub_modules.table.rapidtable.rapid_table import RapidTableModel
+# 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
+#     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"]}')
+# except Exception as e:
+#     if isinstance(e, ImportError):
+#         pass
+#     elif isinstance(e, LicenseFormatError):
+#         logger.error('Ascend Plugin: Invalid license format. Please check the license file.')
+#     elif isinstance(e, LicenseSignatureError):
+#         logger.error('Ascend Plugin: Invalid signature. The license may be tampered with.')
+#     elif isinstance(e, LicenseExpiredError):
+#         logger.error('Ascend Plugin: License has expired. Please renew your license.')
+#     elif isinstance(e, FileNotFoundError):
+#         logger.error('Ascend Plugin: Not found License file.')
+#     else:
+#         logger.error(f'Ascend Plugin: {e}')
+#     from magic_pdf.model.sub_modules.ocr.paddleocr.ppocr_273_mod import ModifiedPaddleOCR
+#     # 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
+
+
+def table_model_init(table_model_type, model_path, max_time, _device_='cpu', lang=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_
         }
         table_model = TableMasterPaddleModel(config)
     elif table_model_type == MODEL_NAME.RAPID_TABLE:
+        atom_model_manager = AtomModelSingleton()
+        ocr_engine = atom_model_manager.get_atom_model(
+            atom_model_name='ocr',
+            ocr_show_log=False,
+            det_db_box_thresh=0.5,
+            det_db_unclip_ratio=1.6,
+            lang=lang
+        )
         table_model = RapidTableModel(ocr_engine, table_sub_model_name)
     else:
         logger.error('table model type not allow')
@@ -55,7 +65,7 @@ def table_model_init(table_model_type, model_path, max_time, _device_='cpu', ocr
 
 
 def mfd_model_init(weight, device='cpu'):
-    if str(device).startswith("npu"):
+    if str(device).startswith('npu'):
         device = torch.device(device)
     mfd_model = YOLOv8MFDModel(weight, device)
     return mfd_model
@@ -67,19 +77,20 @@ 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
 
 
 def doclayout_yolo_model_init(weight, device='cpu'):
-    if str(device).startswith("npu"):
+    if str(device).startswith('npu'):
         device = torch.device(device)
     model = DocLayoutYOLOModel(weight, device)
     return model
 
 
 def langdetect_model_init(langdetect_model_weight, device='cpu'):
-    if str(device).startswith("npu"):
+    if str(device).startswith('npu'):
         device = torch.device(device)
     model = YOLOv11LangDetModel(langdetect_model_weight, device)
     return model
@@ -92,7 +103,8 @@ def ocr_model_init(show_log: bool = False,
                    det_db_unclip_ratio=1.8,
                    ):
     if lang is not None and lang != '':
-        model = ModifiedPaddleOCR(
+        # model = ModifiedPaddleOCR(
+        model = PytorchPaddleOCR(
             show_log=show_log,
             det_db_box_thresh=det_db_box_thresh,
             lang=lang,
@@ -100,7 +112,8 @@ def ocr_model_init(show_log: bool = False,
             det_db_unclip_ratio=det_db_unclip_ratio,
         )
     else:
-        model = ModifiedPaddleOCR(
+        # model = ModifiedPaddleOCR(
+        model = PytorchPaddleOCR(
             show_log=show_log,
             det_db_box_thresh=det_db_box_thresh,
             use_dilation=use_dilation,
@@ -129,7 +142,7 @@ class AtomModelSingleton:
         elif atom_model_name in [AtomicModel.Layout]:
             key = (atom_model_name, layout_model_name)
         elif atom_model_name in [AtomicModel.Table]:
-            key = (atom_model_name, table_model_name)
+            key = (atom_model_name, table_model_name, lang)
         else:
             key = atom_model_name
 
@@ -177,7 +190,7 @@ def atom_model_init(model_name: str, **kwargs):
             kwargs.get('table_model_path'),
             kwargs.get('table_max_time'),
             kwargs.get('device'),
-            kwargs.get('ocr_engine'),
+            kwargs.get('lang'),
             kwargs.get('table_sub_model_name')
         )
     elif model_name == AtomicModel.LangDetect:

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/model/sub_modules/ocr/paddleocr/ppocr_273_mod.py

-import copy
-import platform
-import time
-import cv2
-import numpy as np
-import torch
-
-from paddleocr import PaddleOCR
-from ppocr.utils.logging import get_logger
-from ppocr.utils.utility import alpha_to_color, binarize_img
-from tools.infer.predict_system import sorted_boxes
-from tools.infer.utility import get_rotate_crop_image, get_minarea_rect_crop
-
-from magic_pdf.model.sub_modules.ocr.paddleocr.ocr_utils import update_det_boxes, merge_det_boxes, check_img, \
-    ONNXModelSingleton
-
-logger = get_logger()
-
-
-class ModifiedPaddleOCR(PaddleOCR):
-    def __init__(self, *args, **kwargs):
-
-        super().__init__(*args, **kwargs)
-        self.lang = kwargs.get('lang', 'ch')
-        # 在cpu架构为arm且不支持cuda时调用onnx、
-        if not torch.cuda.is_available() and platform.machine() in ['arm64', 'aarch64']:
-            self.use_onnx = True
-            onnx_model_manager = ONNXModelSingleton()
-            self.additional_ocr = onnx_model_manager.get_onnx_model(**kwargs)
-        else:
-            self.use_onnx = False
-
-    def ocr(self,
-            img,
-            det=True,
-            rec=True,
-            cls=True,
-            bin=False,
-            inv=False,
-            alpha_color=(255, 255, 255),
-            mfd_res=None,
-            ):
-        """
-        OCR with PaddleOCR
-        args：
-            img: img for OCR, support ndarray, img_path and list or ndarray
-            det: use text detection or not. If False, only rec will be exec. Default is True
-            rec: use text recognition or not. If False, only det will be exec. Default is True
-            cls: use angle classifier or not. Default is True. If True, the text with rotation of 180 degrees can be recognized. If no text is rotated by 180 degrees, use cls=False to get better performance. Text with rotation of 90 or 270 degrees can be recognized even if cls=False.
-            bin: binarize image to black and white. Default is False.
-            inv: invert image colors. Default is False.
-            alpha_color: set RGB color Tuple for transparent parts replacement. Default is pure white.
-        """
-        assert isinstance(img, (np.ndarray, list, str, bytes))
-        if isinstance(img, list) and det == True:
-            logger.error('When input a list of images, det must be false')
-            exit(0)
-        if cls == True and self.use_angle_cls == False:
-            pass
-            # logger.warning(
-            #     'Since the angle classifier is not initialized, it will not be used during the forward process'
-            # )
-
-        img = check_img(img)
-        # for infer pdf file
-        if isinstance(img, list):
-            if self.page_num > len(img) or self.page_num == 0:
-                self.page_num = len(img)
-            imgs = img[:self.page_num]
-        else:
-            imgs = [img]
-
-        def preprocess_image(_image):
-            _image = alpha_to_color(_image, alpha_color)
-            if inv:
-                _image = cv2.bitwise_not(_image)
-            if bin:
-                _image = binarize_img(_image)
-            return _image
-
-        if det and rec:
-            ocr_res = []
-            for img in imgs:
-                img = preprocess_image(img)
-                dt_boxes, rec_res, _ = self.__call__(img, cls, mfd_res=mfd_res)
-                if not dt_boxes and not rec_res:
-                    ocr_res.append(None)
-                    continue
-                tmp_res = [[box.tolist(), res]
-                           for box, res in zip(dt_boxes, rec_res)]
-                ocr_res.append(tmp_res)
-            return ocr_res
-        elif det and not rec:
-            ocr_res = []
-            for img in imgs:
-                img = preprocess_image(img)
-                if self.lang in ['ch'] and self.use_onnx:
-                    dt_boxes, elapse = self.additional_ocr.text_detector(img)
-                else:
-                    dt_boxes, elapse = self.text_detector(img)
-                if dt_boxes is None:
-                    ocr_res.append(None)
-                    continue
-                dt_boxes = sorted_boxes(dt_boxes)
-                # merge_det_boxes 和 update_det_boxes 都会把poly转成bbox再转回poly，因此需要过滤所有倾斜程度较大的文本框
-                dt_boxes = merge_det_boxes(dt_boxes)
-                if mfd_res:
-                    bef = time.time()
-                    dt_boxes = update_det_boxes(dt_boxes, mfd_res)
-                    aft = time.time()
-                    logger.debug("split text box by formula, new dt_boxes num : {}, elapsed : {}".format(
-                        len(dt_boxes), aft - bef))
-                tmp_res = [box.tolist() for box in dt_boxes]
-                ocr_res.append(tmp_res)
-            return ocr_res
-        else:
-            ocr_res = []
-            cls_res = []
-            for img in imgs:
-                if not isinstance(img, list):
-                    img = preprocess_image(img)
-                    img = [img]
-                if self.use_angle_cls and cls:
-                    img, cls_res_tmp, elapse = self.text_classifier(img)
-                    if not rec:
-                        cls_res.append(cls_res_tmp)
-                if self.lang in ['ch'] and self.use_onnx:
-                    rec_res, elapse = self.additional_ocr.text_recognizer(img)
-                else:
-                    rec_res, elapse = self.text_recognizer(img)
-                ocr_res.append(rec_res)
-            if not rec:
-                return cls_res
-            return ocr_res
-
-    def __call__(self, img, cls=True, mfd_res=None):
-        time_dict = {'det': 0, 'rec': 0, 'cls': 0, 'all': 0}
-
-        if img is None:
-            logger.debug("no valid image provided")
-            return None, None, time_dict
-
-        start = time.time()
-        ori_im = img.copy()
-        if self.lang in ['ch'] and self.use_onnx:
-            dt_boxes, elapse = self.additional_ocr.text_detector(img)
-        else:
-            dt_boxes, elapse = self.text_detector(img)
-        time_dict['det'] = elapse
-
-        if dt_boxes is None:
-            logger.debug("no dt_boxes found, elapsed : {}".format(elapse))
-            end = time.time()
-            time_dict['all'] = end - start
-            return None, None, time_dict
-        else:
-            logger.debug("dt_boxes num : {}, elapsed : {}".format(
-                len(dt_boxes), elapse))
-        img_crop_list = []
-
-        dt_boxes = sorted_boxes(dt_boxes)
-
-        # merge_det_boxes 和 update_det_boxes 都会把poly转成bbox再转回poly，因此需要过滤所有倾斜程度较大的文本框
-        dt_boxes = merge_det_boxes(dt_boxes)
-
-        if mfd_res:
-            bef = time.time()
-            dt_boxes = update_det_boxes(dt_boxes, mfd_res)
-            aft = time.time()
-            logger.debug("split text box by formula, new dt_boxes num : {}, elapsed : {}".format(
-                len(dt_boxes), aft - bef))
-
-        for bno in range(len(dt_boxes)):
-            tmp_box = copy.deepcopy(dt_boxes[bno])
-            if self.args.det_box_type == "quad":
-                img_crop = get_rotate_crop_image(ori_im, tmp_box)
-            else:
-                img_crop = get_minarea_rect_crop(ori_im, tmp_box)
-            img_crop_list.append(img_crop)
-        if self.use_angle_cls and cls:
-            img_crop_list, angle_list, elapse = self.text_classifier(
-                img_crop_list)
-            time_dict['cls'] = elapse
-            logger.debug("cls num  : {}, elapsed : {}".format(
-                len(img_crop_list), elapse))
-        if self.lang in ['ch'] and self.use_onnx:
-            rec_res, elapse = self.additional_ocr.text_recognizer(img_crop_list)
-        else:
-            rec_res, elapse = self.text_recognizer(img_crop_list)
-        time_dict['rec'] = elapse
-        logger.debug("rec_res num  : {}, elapsed : {}".format(
-            len(rec_res), elapse))
-        if self.args.save_crop_res:
-            self.draw_crop_rec_res(self.args.crop_res_save_dir, img_crop_list,
-                                   rec_res)
-        filter_boxes, filter_rec_res = [], []
-        for box, rec_result in zip(dt_boxes, rec_res):
-            text, score = rec_result
-            if score >= self.drop_score:
-                filter_boxes.append(box)
-                filter_rec_res.append(rec_result)
-        end = time.time()
-        time_dict['all'] = end - start
-        return filter_boxes, filter_rec_res, time_dict
\ No newline at end of file

magic_pdf/model/sub_modules/ocr/paddleocr/ppocr_291_mod.py

-import copy
-import time
-
-
-import cv2
-import numpy as np
-from paddleocr import PaddleOCR
-from paddleocr.paddleocr import check_img, logger
-from paddleocr.ppocr.utils.utility import alpha_to_color, binarize_img
-from paddleocr.tools.infer.predict_system import sorted_boxes
-from paddleocr.tools.infer.utility import slice_generator, merge_fragmented, get_rotate_crop_image, \
-    get_minarea_rect_crop
-
-from magic_pdf.model.sub_modules.ocr.paddleocr.ocr_utils import update_det_boxes
-
-
-class ModifiedPaddleOCR(PaddleOCR):
-
-    def ocr(
-        self,
-        img,
-        det=True,
-        rec=True,
-        cls=True,
-        bin=False,
-        inv=False,
-        alpha_color=(255, 255, 255),
-        slice={},
-        mfd_res=None,
-    ):
-        """
-        OCR with PaddleOCR
-
-        Args:
-            img: Image for OCR. It can be an ndarray, img_path, or a list of ndarrays.
-            det: Use text detection or not. If False, only text recognition will be executed. Default is True.
-            rec: Use text recognition or not. If False, only text detection will be executed. Default is True.
-            cls: Use angle classifier or not. Default is True. If True, the text with a rotation of 180 degrees can be recognized. If no text is rotated by 180 degrees, use cls=False to get better performance.
-            bin: Binarize image to black and white. Default is False.
-            inv: Invert image colors. Default is False.
-            alpha_color: Set RGB color Tuple for transparent parts replacement. Default is pure white.
-            slice: Use sliding window inference for large images. Both det and rec must be True. Requires int values for slice["horizontal_stride"], slice["vertical_stride"], slice["merge_x_thres"], slice["merge_y_thres"] (See doc/doc_en/slice_en.md). Default is {}.
-
-        Returns:
-            If both det and rec are True, returns a list of OCR results for each image. Each OCR result is a list of bounding boxes and recognized text for each detected text region.
-            If det is True and rec is False, returns a list of detected bounding boxes for each image.
-            If det is False and rec is True, returns a list of recognized text for each image.
-            If both det and rec are False, returns a list of angle classification results for each image.
-
-        Raises:
-            AssertionError: If the input image is not of type ndarray, list, str, or bytes.
-            SystemExit: If det is True and the input is a list of images.
-
-        Note:
-            - If the angle classifier is not initialized (use_angle_cls=False), it will not be used during the forward process.
-            - For PDF files, if the input is a list of images and the page_num is specified, only the first page_num images will be processed.
-            - The preprocess_image function is used to preprocess the input image by applying alpha color replacement, inversion, and binarization if specified.
-        """
-        assert isinstance(img, (np.ndarray, list, str, bytes))
-        if isinstance(img, list) and det == True:
-            logger.error("When input a list of images, det must be false")
-            exit(0)
-        if cls == True and self.use_angle_cls == False:
-            logger.warning(
-                "Since the angle classifier is not initialized, it will not be used during the forward process"
-            )
-
-        img, flag_gif, flag_pdf = check_img(img, alpha_color)
-        # for infer pdf file
-        if isinstance(img, list) and flag_pdf:
-            if self.page_num > len(img) or self.page_num == 0:
-                imgs = img
-            else:
-                imgs = img[: self.page_num]
-        else:
-            imgs = [img]
-
-        def preprocess_image(_image):
-            _image = alpha_to_color(_image, alpha_color)
-            if inv:
-                _image = cv2.bitwise_not(_image)
-            if bin:
-                _image = binarize_img(_image)
-            return _image
-
-        if det and rec:
-            ocr_res = []
-            for img in imgs:
-                img = preprocess_image(img)
-                dt_boxes, rec_res, _ = self.__call__(img, cls, slice, mfd_res=mfd_res)
-                if not dt_boxes and not rec_res:
-                    ocr_res.append(None)
-                    continue
-                tmp_res = [[box.tolist(), res] for box, res in zip(dt_boxes, rec_res)]
-                ocr_res.append(tmp_res)
-            return ocr_res
-        elif det and not rec:
-            ocr_res = []
-            for img in imgs:
-                img = preprocess_image(img)
-                dt_boxes, elapse = self.text_detector(img)
-                if dt_boxes.size == 0:
-                    ocr_res.append(None)
-                    continue
-                tmp_res = [box.tolist() for box in dt_boxes]
-                ocr_res.append(tmp_res)
-            return ocr_res
-        else:
-            ocr_res = []
-            cls_res = []
-            for img in imgs:
-                if not isinstance(img, list):
-                    img = preprocess_image(img)
-                    img = [img]
-                if self.use_angle_cls and cls:
-                    img, cls_res_tmp, elapse = self.text_classifier(img)
-                    if not rec:
-                        cls_res.append(cls_res_tmp)
-                rec_res, elapse = self.text_recognizer(img)
-                ocr_res.append(rec_res)
-            if not rec:
-                return cls_res
-            return ocr_res
-
-    def __call__(self, img, cls=True, slice={}, mfd_res=None):
-        time_dict = {"det": 0, "rec": 0, "cls": 0, "all": 0}
-
-        if img is None:
-            logger.debug("no valid image provided")
-            return None, None, time_dict
-
-        start = time.time()
-        ori_im = img.copy()
-        if slice:
-            slice_gen = slice_generator(
-                img,
-                horizontal_stride=slice["horizontal_stride"],
-                vertical_stride=slice["vertical_stride"],
-            )
-            elapsed = []
-            dt_slice_boxes = []
-            for slice_crop, v_start, h_start in slice_gen:
-                dt_boxes, elapse = self.text_detector(slice_crop, use_slice=True)
-                if dt_boxes.size:
-                    dt_boxes[:, :, 0] += h_start
-                    dt_boxes[:, :, 1] += v_start
-                    dt_slice_boxes.append(dt_boxes)
-                    elapsed.append(elapse)
-            dt_boxes = np.concatenate(dt_slice_boxes)
-
-            dt_boxes = merge_fragmented(
-                boxes=dt_boxes,
-                x_threshold=slice["merge_x_thres"],
-                y_threshold=slice["merge_y_thres"],
-            )
-            elapse = sum(elapsed)
-        else:
-            dt_boxes, elapse = self.text_detector(img)
-
-        time_dict["det"] = elapse
-
-        if dt_boxes is None:
-            logger.debug("no dt_boxes found, elapsed : {}".format(elapse))
-            end = time.time()
-            time_dict["all"] = end - start
-            return None, None, time_dict
-        else:
-            logger.debug(
-                "dt_boxes num : {}, elapsed : {}".format(len(dt_boxes), elapse)
-            )
-        img_crop_list = []
-
-        dt_boxes = sorted_boxes(dt_boxes)
-
-        if mfd_res:
-            bef = time.time()
-            dt_boxes = update_det_boxes(dt_boxes, mfd_res)
-            aft = time.time()
-            logger.debug("split text box by formula, new dt_boxes num : {}, elapsed : {}".format(
-                len(dt_boxes), aft - bef))
-
-        for bno in range(len(dt_boxes)):
-            tmp_box = copy.deepcopy(dt_boxes[bno])
-            if self.args.det_box_type == "quad":
-                img_crop = get_rotate_crop_image(ori_im, tmp_box)
-            else:
-                img_crop = get_minarea_rect_crop(ori_im, tmp_box)
-            img_crop_list.append(img_crop)
-        if self.use_angle_cls and cls:
-            img_crop_list, angle_list, elapse = self.text_classifier(img_crop_list)
-            time_dict["cls"] = elapse
-            logger.debug(
-                "cls num  : {}, elapsed : {}".format(len(img_crop_list), elapse)
-            )
-        if len(img_crop_list) > 1000:
-            logger.debug(
-                f"rec crops num: {len(img_crop_list)}, time and memory cost may be large."
-            )
-
-        rec_res, elapse = self.text_recognizer(img_crop_list)
-        time_dict["rec"] = elapse
-        logger.debug("rec_res num  : {}, elapsed : {}".format(len(rec_res), elapse))
-        if self.args.save_crop_res:
-            self.draw_crop_rec_res(self.args.crop_res_save_dir, img_crop_list, rec_res)
-        filter_boxes, filter_rec_res = [], []
-        for box, rec_result in zip(dt_boxes, rec_res):
-            text, score = rec_result[0], rec_result[1]
-            if score >= self.drop_score:
-                filter_boxes.append(box)
-                filter_rec_res.append(rec_result)
-        end = time.time()
-        time_dict["all"] = end - start
-        return filter_boxes, filter_rec_res, time_dict

magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/__init__.py

+# Copyright (c) Opendatalab. All rights reserved.

magic_pdf/model/sub_modules/ocr/paddleocr/ocr_utils.py → magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/ocr_utils.py

+# Copyright (c) Opendatalab. All rights reserved.
+import copy
+
 import cv2
 import numpy as np
-from loguru import logger
-from io import BytesIO
-from PIL import Image
-import base64
-from magic_pdf.libs.boxbase import __is_overlaps_y_exceeds_threshold
 from magic_pdf.pre_proc.ocr_dict_merge import merge_spans_to_line
-
-import importlib.resources
-from paddleocr import PaddleOCR
-from ppocr.utils.utility import check_and_read
+from magic_pdf.libs.boxbase import __is_overlaps_y_exceeds_threshold
 
 
 def img_decode(content: bytes):
     np_arr = np.frombuffer(content, dtype=np.uint8)
     return cv2.imdecode(np_arr, cv2.IMREAD_UNCHANGED)
 
-
 def check_img(img):
     if isinstance(img, bytes):
         img = img_decode(img)
-    if isinstance(img, str):
-        image_file = img
-        img, flag_gif, flag_pdf = check_and_read(image_file)
-        if not flag_gif and not flag_pdf:
-            with open(image_file, 'rb') as f:
-                img_str = f.read()
-                img = img_decode(img_str)
-            if img is None:
-                try:
-                    buf = BytesIO()
-                    image = BytesIO(img_str)
-                    im = Image.open(image)
-                    rgb = im.convert('RGB')
-                    rgb.save(buf, 'jpeg')
-                    buf.seek(0)
-                    image_bytes = buf.read()
-                    data_base64 = str(base64.b64encode(image_bytes),
-                                      encoding="utf-8")
-                    image_decode = base64.b64decode(data_base64)
-                    img_array = np.frombuffer(image_decode, np.uint8)
-                    img = cv2.imdecode(img_array, cv2.IMREAD_COLOR)
-                except:
-                    logger.error("error in loading image:{}".format(image_file))
-                    return None
-        if img is None:
-            logger.error("error in loading image:{}".format(image_file))
-            return None
     if isinstance(img, np.ndarray) and len(img.shape) == 2:
         img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)
+    return img
 
+
+def alpha_to_color(img, alpha_color=(255, 255, 255)):
+    if len(img.shape) == 3 and img.shape[2] == 4:
+        B, G, R, A = cv2.split(img)
+        alpha = A / 255
+
+        R = (alpha_color[0] * (1 - alpha) + R * alpha).astype(np.uint8)
+        G = (alpha_color[1] * (1 - alpha) + G * alpha).astype(np.uint8)
+        B = (alpha_color[2] * (1 - alpha) + B * alpha).astype(np.uint8)
+
+        img = cv2.merge((B, G, R))
     return img
 
 
+def preprocess_image(_image):
+    alpha_color = (255, 255, 255)
+    _image = alpha_to_color(_image, alpha_color)
+    return _image
+
+
+def sorted_boxes(dt_boxes):
+    """
+    Sort text boxes in order from top to bottom, left to right
+    args:
+        dt_boxes(array):detected text boxes with shape [4, 2]
+    return:
+        sorted boxes(array) with shape [4, 2]
+    """
+    num_boxes = dt_boxes.shape[0]
+    sorted_boxes = sorted(dt_boxes, key=lambda x: (x[0][1], x[0][0]))
+    _boxes = list(sorted_boxes)
+
+    for i in range(num_boxes - 1):
+        for j in range(i, -1, -1):
+            if abs(_boxes[j + 1][0][1] - _boxes[j][0][1]) < 10 and \
+                    (_boxes[j + 1][0][0] < _boxes[j][0][0]):
+                tmp = _boxes[j]
+                _boxes[j] = _boxes[j + 1]
+                _boxes[j + 1] = tmp
+            else:
+                break
+    return _boxes
+
+
 def bbox_to_points(bbox):
     """ 将bbox格式转换为四个顶点的数组 """
     x0, y0, x1, y1 = bbox
@@ -252,9 +261,10 @@ def get_adjusted_mfdetrec_res(single_page_mfdetrec_res, useful_list):
     return adjusted_mfdetrec_res
 
 
-def get_ocr_result_list(ocr_res, useful_list):
+def get_ocr_result_list(ocr_res, useful_list, ocr_enable, new_image, lang):
     paste_x, paste_y, xmin, ymin, xmax, ymax, new_width, new_height = useful_list
     ocr_result_list = []
+    ori_im = new_image.copy()
     for box_ocr_res in ocr_res:
 
         if len(box_ocr_res) == 2:
@@ -266,6 +276,11 @@ def get_ocr_result_list(ocr_res, useful_list):
         else:
             p1, p2, p3, p4 = box_ocr_res
             text, score = "", 1
+
+            if ocr_enable:
+                tmp_box = copy.deepcopy(np.array([p1, p2, p3, p4]).astype('float32'))
+                img_crop = get_rotate_crop_image(ori_im, tmp_box)
+
         # average_angle_degrees = calculate_angle_degrees(box_ocr_res[0])
         # if average_angle_degrees > 0.5:
         poly = [p1, p2, p3, p4]
@@ -288,12 +303,22 @@ def get_ocr_result_list(ocr_res, useful_list):
         p3 = [p3[0] - paste_x + xmin, p3[1] - paste_y + ymin]
         p4 = [p4[0] - paste_x + xmin, p4[1] - paste_y + ymin]
 
-        ocr_result_list.append({
-            'category_id': 15,
-            'poly': p1 + p2 + p3 + p4,
-            'score': float(round(score, 2)),
-            'text': text,
-        })
+        if ocr_enable:
+            ocr_result_list.append({
+                'category_id': 15,
+                'poly': p1 + p2 + p3 + p4,
+                'score': 1,
+                'text': text,
+                'np_img': img_crop,
+                'lang': lang,
+            })
+        else:
+            ocr_result_list.append({
+                'category_id': 15,
+                'poly': p1 + p2 + p3 + p4,
+                'score': float(round(score, 2)),
+                'text': text,
+            })
 
     return ocr_result_list
 
@@ -308,56 +333,36 @@ def calculate_is_angle(poly):
         return True
 
 
-class ONNXModelSingleton:
-    _instance = None
-    _models = {}
-
-    def __new__(cls, *args, **kwargs):
-        if cls._instance is None:
-            cls._instance = super().__new__(cls)
-        return cls._instance
-
-    def get_onnx_model(self, **kwargs):
-
-        lang = kwargs.get('lang', None)
-        det_db_box_thresh = kwargs.get('det_db_box_thresh', 0.3)
-        use_dilation = kwargs.get('use_dilation', True)
-        det_db_unclip_ratio = kwargs.get('det_db_unclip_ratio', 1.8)
-        key = (lang, det_db_box_thresh, use_dilation, det_db_unclip_ratio)
-        if key not in self._models:
-            self._models[key] = onnx_model_init(key)
-        return self._models[key]
-
-def onnx_model_init(key):
-    if len(key) < 4:
-        logger.error('Invalid key length, expected at least 4 elements')
-        exit(1)
-
-    try:
-        with importlib.resources.path('rapidocr_onnxruntime.models', '') as resource_path:
-            additional_ocr_params = {
-                "use_onnx": True,
-                "det_model_dir": f'{resource_path}/ch_PP-OCRv4_det_infer.onnx',
-                "rec_model_dir": f'{resource_path}/ch_PP-OCRv4_rec_infer.onnx',
-                "cls_model_dir": f'{resource_path}/ch_ppocr_mobile_v2.0_cls_infer.onnx',
-                "det_db_box_thresh": key[1],
-                "use_dilation": key[2],
+def get_rotate_crop_image(img, points):
+    '''
+    img_height, img_width = img.shape[0:2]
+    left = int(np.min(points[:, 0]))
+    right = int(np.max(points[:, 0]))
+    top = int(np.min(points[:, 1]))
+    bottom = int(np.max(points[:, 1]))
+    img_crop = img[top:bottom, left:right, :].copy()
+    points[:, 0] = points[:, 0] - left
+    points[:, 1] = points[:, 1] - top
+    '''
+    assert len(points) == 4, "shape of points must be 4*2"
+    img_crop_width = int(
+        max(
+            np.linalg.norm(points[0] - points[1]),
+            np.linalg.norm(points[2] - points[3])))
+    img_crop_height = int(
+        max(
+            np.linalg.norm(points[0] - points[3]),
+            np.linalg.norm(points[1] - points[2])))
+    pts_std = np.float32([[0, 0], [img_crop_width, 0],
+                          [img_crop_width, img_crop_height],
+                          [0, img_crop_height]])
+    M = cv2.getPerspectiveTransform(points, pts_std)
+    dst_img = cv2.warpPerspective(
+        img,
+        M, (img_crop_width, img_crop_height),
+        borderMode=cv2.BORDER_REPLICATE,
+        flags=cv2.INTER_CUBIC)
+    dst_img_height, dst_img_width = dst_img.shape[0:2]
+    if dst_img_height * 1.0 / dst_img_width >= 1.5:
+        dst_img = np.rot90(dst_img)
+    return dst_img
\ No newline at end of file
-                "det_db_unclip_ratio": key[3],
-            }
-
-            if key[0] is not None:
-                additional_ocr_params["lang"] = key[0]
-
-            # logger.info(f"additional_ocr_params: {additional_ocr_params}")
-
-            onnx_model = PaddleOCR(**additional_ocr_params)
-
-            if onnx_model is None:
-                logger.error('model init failed')
-                exit(1)
-            else:
-                return onnx_model
-
-    except Exception as e:
-        logger.exception(f'Error initializing model: {e}')
-        exit(1)
\ No newline at end of file

magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorch_paddle.py

+# Copyright (c) Opendatalab. All rights reserved.
+import copy
+import os.path
+import warnings
+from pathlib import Path
+
+import cv2
+import numpy as np
+import yaml
+from loguru import logger
+
+from magic_pdf.libs.config_reader import get_device, get_local_models_dir
+from .ocr_utils import check_img, preprocess_image, sorted_boxes, merge_det_boxes, update_det_boxes, get_rotate_crop_image
+from .tools.infer.predict_system import TextSystem
+from .tools.infer import pytorchocr_utility as utility
+import argparse
+
+
+latin_lang = [
+        'af', 'az', 'bs', 'cs', 'cy', 'da', 'de', 'es', 'et', 'fr', 'ga', 'hr',  # noqa: E126
+        'hu', 'id', 'is', 'it', 'ku', 'la', 'lt', 'lv', 'mi', 'ms', 'mt', 'nl',
+        'no', 'oc', 'pi', 'pl', 'pt', 'ro', 'rs_latin', 'sk', 'sl', 'sq', 'sv',
+        'sw', 'tl', 'tr', 'uz', 'vi', 'french', 'german'
+]
+arabic_lang = ['ar', 'fa', 'ug', 'ur']
+cyrillic_lang = [
+        'ru', 'rs_cyrillic', 'be', 'bg', 'uk', 'mn', 'abq', 'ady', 'kbd', 'ava',  # noqa: E126
+        'dar', 'inh', 'che', 'lbe', 'lez', 'tab'
+]
+devanagari_lang = [
+        'hi', 'mr', 'ne', 'bh', 'mai', 'ang', 'bho', 'mah', 'sck', 'new', 'gom',  # noqa: E126
+        'sa', 'bgc'
+]
+
+
+def get_model_params(lang, config):
+    if lang in config['lang']:
+        params = config['lang'][lang]
+        det = params.get('det')
+        rec = params.get('rec')
+        dict_file = params.get('dict')
+        return det, rec, dict_file
+    else:
+        raise Exception (f'Language {lang} not supported')
+
+
+root_dir = Path(__file__).resolve().parent
+
+
+class PytorchPaddleOCR(TextSystem):
+    def __init__(self, *args, **kwargs):
+        parser = utility.init_args()
+        args = parser.parse_args(args)
+
+        self.lang = kwargs.get('lang', 'ch')
+        if self.lang in latin_lang:
+            self.lang = 'latin'
+        elif self.lang in arabic_lang:
+            self.lang = 'arabic'
+        elif self.lang in cyrillic_lang:
+            self.lang = 'cyrillic'
+        elif self.lang in devanagari_lang:
+            self.lang = 'devanagari'
+        else:
+            pass
+
+        models_config_path = os.path.join(root_dir, 'pytorchocr', 'utils', 'resources', 'models_config.yml')
+        with open(models_config_path) as file:
+            config = yaml.safe_load(file)
+            det, rec, dict_file = get_model_params(self.lang, config)
+        ocr_models_dir = os.path.join(get_local_models_dir(), 'OCR', 'paddleocr_torch')
+        kwargs['det_model_path'] = os.path.join(ocr_models_dir, det)
+        kwargs['rec_model_path'] = os.path.join(ocr_models_dir, rec)
+        kwargs['rec_char_dict_path'] = os.path.join(root_dir, 'pytorchocr', 'utils', 'resources', 'dict', dict_file)
+        # kwargs['rec_batch_num'] = 8
+
+        kwargs['device'] = get_device()
+
+        default_args = vars(args)
+        default_args.update(kwargs)
+        args = argparse.Namespace(**default_args)
+
+        super().__init__(args)
+
+    def ocr(self,
+            img,
+            det=True,
+            rec=True,
+            mfd_res=None,
+            tqdm_enable=False,
+            ):
+        assert isinstance(img, (np.ndarray, list, str, bytes))
+        if isinstance(img, list) and det == True:
+            logger.error('When input a list of images, det must be false')
+            exit(0)
+        img = check_img(img)
+        imgs = [img]
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore", category=RuntimeWarning)
+            if det and rec:
+                ocr_res = []
+                for img in imgs:
+                    img = preprocess_image(img)
+                    dt_boxes, rec_res = self.__call__(img, mfd_res=mfd_res)
+                    if not dt_boxes and not rec_res:
+                        ocr_res.append(None)
+                        continue
+                    tmp_res = [[box.tolist(), res] for box, res in zip(dt_boxes, rec_res)]
+                    ocr_res.append(tmp_res)
+                return ocr_res
+            elif det and not rec:
+                ocr_res = []
+                for img in imgs:
+                    img = preprocess_image(img)
+                    dt_boxes, elapse = self.text_detector(img)
+                    # logger.debug("dt_boxes num : {}, elapsed : {}".format(len(dt_boxes), elapse))
+                    if dt_boxes is None:
+                        ocr_res.append(None)
+                        continue
+                    dt_boxes = sorted_boxes(dt_boxes)
+                    # merge_det_boxes 和 update_det_boxes 都会把poly转成bbox再转回poly，因此需要过滤所有倾斜程度较大的文本框
+                    dt_boxes = merge_det_boxes(dt_boxes)
+                    if mfd_res:
+                        dt_boxes = update_det_boxes(dt_boxes, mfd_res)
+                    tmp_res = [box.tolist() for box in dt_boxes]
+                    ocr_res.append(tmp_res)
+                return ocr_res
+            elif not det and rec:
+                ocr_res = []
+                for img in imgs:
+                    if not isinstance(img, list):
+                        img = preprocess_image(img)
+                        img = [img]
+                    rec_res, elapse = self.text_recognizer(img, tqdm_enable=tqdm_enable)
+                    # logger.debug("rec_res num  : {}, elapsed : {}".format(len(rec_res), elapse))
+                    ocr_res.append(rec_res)
+                return ocr_res
+
+    def __call__(self, img, mfd_res=None):
+
+        if img is None:
+            logger.debug("no valid image provided")
+            return None, None
+
+        ori_im = img.copy()
+        dt_boxes, elapse = self.text_detector(img)
+
+        if dt_boxes is None:
+            logger.debug("no dt_boxes found, elapsed : {}".format(elapse))
+            return None, None
+        else:
+            pass
+            # logger.debug("dt_boxes num : {}, elapsed : {}".format(len(dt_boxes), elapse))
+        img_crop_list = []
+
+        dt_boxes = sorted_boxes(dt_boxes)
+
+        # merge_det_boxes 和 update_det_boxes 都会把poly转成bbox再转回poly，因此需要过滤所有倾斜程度较大的文本框
+        dt_boxes = merge_det_boxes(dt_boxes)
+
+        if mfd_res:
+            dt_boxes = update_det_boxes(dt_boxes, mfd_res)
+
+        for bno in range(len(dt_boxes)):
+            tmp_box = copy.deepcopy(dt_boxes[bno])
+            img_crop = get_rotate_crop_image(ori_im, tmp_box)
+            img_crop_list.append(img_crop)
+
+        rec_res, elapse = self.text_recognizer(img_crop_list)
+        # logger.debug("rec_res num  : {}, elapsed : {}".format(len(rec_res), elapse))
+
+        filter_boxes, filter_rec_res = [], []
+        for box, rec_result in zip(dt_boxes, rec_res):
+            text, score = rec_result
+            if score >= self.drop_score:
+                filter_boxes.append(box)
+                filter_rec_res.append(rec_result)
+
+        return filter_boxes, filter_rec_res
+
+if __name__ == '__main__':
+    pytorch_paddle_ocr = PytorchPaddleOCR()
+    img = cv2.imread("/Users/myhloli/Downloads/screenshot-20250326-194348.png")
+    dt_boxes, rec_res = pytorch_paddle_ocr(img)
+    ocr_res = []
+    if not dt_boxes and not rec_res:
+        ocr_res.append(None)
+    else:
+        tmp_res = [[box.tolist(), res] for box, res in zip(dt_boxes, rec_res)]
+        ocr_res.append(tmp_res)
+    print(ocr_res)
+
+

magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/base_ocr_v20.py

+import os
+import torch
+from .modeling.architectures.base_model import BaseModel
+
+class BaseOCRV20:
+    def __init__(self, config, **kwargs):
+        self.config = config
+        self.build_net(**kwargs)
+        self.net.eval()
+
+
+    def build_net(self, **kwargs):
+        self.net = BaseModel(self.config, **kwargs)
+
+    def read_pytorch_weights(self, weights_path):
+        if not os.path.exists(weights_path):
+            raise FileNotFoundError('{} is not existed.'.format(weights_path))
+        weights = torch.load(weights_path)
+        return weights
+
+    def get_out_channels(self, weights):
+        if list(weights.keys())[-1].endswith('.weight') and len(list(weights.values())[-1].shape) == 2:
+            out_channels = list(weights.values())[-1].numpy().shape[1]
+        else:
+            out_channels = list(weights.values())[-1].numpy().shape[0]
+        return out_channels
+
+    def load_state_dict(self, weights):
+        self.net.load_state_dict(weights)
+        # print('weights is loaded.')
+
+    def load_pytorch_weights(self, weights_path):
+        self.net.load_state_dict(torch.load(weights_path, weights_only=True))
+        # print('model is loaded: {}'.format(weights_path))
+
+    def inference(self, inputs):
+        with torch.no_grad():
+            infer = self.net(inputs)
+        return infer

magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/data/__init__.py

+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+from __future__ import unicode_literals
+
+from .imaug import transform, create_operators
+
+

magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/data/imaug/__init__.py

+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+from __future__ import unicode_literals
+
+# from .iaa_augment import IaaAugment
+# from .make_border_map import MakeBorderMap
+# from .make_shrink_map import MakeShrinkMap
+# from .random_crop_data import EastRandomCropData, PSERandomCrop
+
+# from .rec_img_aug import RecAug, RecResizeImg, ClsResizeImg
+# from .randaugment import RandAugment
+from .operators import *
+# from .label_ops import *
+
+# from .east_process import *
+# from .sast_process import *
+# from .gen_table_mask import *
+
+def transform(data, ops=None):
+    """ transform """
+    if ops is None:
+        ops = []
+    for op in ops:
+        data = op(data)
+        if data is None:
+            return None
+    return data
+
+
+def create_operators(op_param_list, global_config=None):
+    """
+    create operators based on the config
+    Args:
+        params(list): a dict list, used to create some operators
+    """
+    assert isinstance(op_param_list, list), ('operator config should be a list')
+    ops = []
+    for operator in op_param_list:
+        assert isinstance(operator,
+                          dict) and len(operator) == 1, "yaml format error"
+        op_name = list(operator)[0]
+        param = {} if operator[op_name] is None else operator[op_name]
+        if global_config is not None:
+            param.update(global_config)
+        op = eval(op_name)(**param)
+        ops.append(op)
+    return ops
\ No newline at end of file

magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/data/imaug/operators.py

+"""
+# Copyright (c) 2020 PaddlePaddle Authors. 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.
+"""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+from __future__ import unicode_literals
+
+import sys
+import six
+import cv2
+import numpy as np
+
+
+class DecodeImage(object):
+    """ decode image """
+
+    def __init__(self, img_mode='RGB', channel_first=False, **kwargs):
+        self.img_mode = img_mode
+        self.channel_first = channel_first
+
+    def __call__(self, data):
+        img = data['image']
+        if six.PY2:
+            assert type(img) is str and len(
+                img) > 0, "invalid input 'img' in DecodeImage"
+        else:
+            assert type(img) is bytes and len(
+                img) > 0, "invalid input 'img' in DecodeImage"
+        img = np.frombuffer(img, dtype='uint8')
+        img = cv2.imdecode(img, 1)
+        if img is None:
+            return None
+        if self.img_mode == 'GRAY':
+            img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)
+        elif self.img_mode == 'RGB':
+            assert img.shape[2] == 3, 'invalid shape of image[%s]' % (img.shape)
+            img = img[:, :, ::-1]
+
+        if self.channel_first:
+            img = img.transpose((2, 0, 1))
+
+        data['image'] = img
+        return data
+
+
+class NRTRDecodeImage(object):
+    """ decode image """
+
+    def __init__(self, img_mode='RGB', channel_first=False, **kwargs):
+        self.img_mode = img_mode
+        self.channel_first = channel_first
+
+    def __call__(self, data):
+        img = data['image']
+        if six.PY2:
+            assert type(img) is str and len(
+                img) > 0, "invalid input 'img' in DecodeImage"
+        else:
+            assert type(img) is bytes and len(
+                img) > 0, "invalid input 'img' in DecodeImage"
+        img = np.frombuffer(img, dtype='uint8')
+
+        img = cv2.imdecode(img, 1)
+
+        if img is None:
+            return None
+        if self.img_mode == 'GRAY':
+            img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)
+        elif self.img_mode == 'RGB':
+            assert img.shape[2] == 3, 'invalid shape of image[%s]' % (img.shape)
+            img = img[:, :, ::-1]
+        img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
+        if self.channel_first:
+            img = img.transpose((2, 0, 1))
+        data['image'] = img
+        return data
+
+
+class NormalizeImage(object):
+    """ normalize image such as substract mean, divide std
+    """
+
+    def __init__(self, scale=None, mean=None, std=None, order='chw', **kwargs):
+        if isinstance(scale, str):
+            scale = eval(scale)
+        self.scale = np.float32(scale if scale is not None else 1.0 / 255.0)
+        mean = mean if mean is not None else [0.485, 0.456, 0.406]
+        std = std if std is not None else [0.229, 0.224, 0.225]
+
+        shape = (3, 1, 1) if order == 'chw' else (1, 1, 3)
+        self.mean = np.array(mean).reshape(shape).astype('float32')
+        self.std = np.array(std).reshape(shape).astype('float32')
+
+    def __call__(self, data):
+        img = data['image']
+        from PIL import Image
+        if isinstance(img, Image.Image):
+            img = np.array(img)
+        assert isinstance(img,
+                          np.ndarray), "invalid input 'img' in NormalizeImage"
+        data['image'] = (
+            img.astype('float32') * self.scale - self.mean) / self.std
+        return data
+
+
+class ToCHWImage(object):
+    """ convert hwc image to chw image
+    """
+
+    def __init__(self, **kwargs):
+        pass
+
+    def __call__(self, data):
+        img = data['image']
+        from PIL import Image
+        if isinstance(img, Image.Image):
+            img = np.array(img)
+        data['image'] = img.transpose((2, 0, 1))
+        return data
+
+
+class Fasttext(object):
+    def __init__(self, path="None", **kwargs):
+        import fasttext
+        self.fast_model = fasttext.load_model(path)
+
+    def __call__(self, data):
+        label = data['label']
+        fast_label = self.fast_model[label]
+        data['fast_label'] = fast_label
+        return data
+
+
+class KeepKeys(object):
+    def __init__(self, keep_keys, **kwargs):
+        self.keep_keys = keep_keys
+
+    def __call__(self, data):
+        data_list = []
+        for key in self.keep_keys:
+            data_list.append(data[key])
+        return data_list
+
+
+class Resize(object):
+    def __init__(self, size=(640, 640), **kwargs):
+        self.size = size
+
+    def resize_image(self, img):
+        resize_h, resize_w = self.size
+        ori_h, ori_w = img.shape[:2]  # (h, w, c)
+        ratio_h = float(resize_h) / ori_h
+        ratio_w = float(resize_w) / ori_w
+        img = cv2.resize(img, (int(resize_w), int(resize_h)))
+        return img, [ratio_h, ratio_w]
+
+    def __call__(self, data):
+        img = data['image']
+        text_polys = data['polys']
+
+        img_resize, [ratio_h, ratio_w] = self.resize_image(img)
+        new_boxes = []
+        for box in text_polys:
+            new_box = []
+            for cord in box:
+                new_box.append([cord[0] * ratio_w, cord[1] * ratio_h])
+            new_boxes.append(new_box)
+        data['image'] = img_resize
+        data['polys'] = np.array(new_boxes, dtype=np.float32)
+        return data
+
+
+class DetResizeForTest(object):
+    def __init__(self, **kwargs):
+        super(DetResizeForTest, self).__init__()
+        self.resize_type = 0
+        if 'image_shape' in kwargs:
+            self.image_shape = kwargs['image_shape']
+            self.resize_type = 1
+        elif 'limit_side_len' in kwargs:
+            self.limit_side_len = kwargs['limit_side_len']
+            self.limit_type = kwargs.get('limit_type', 'min')
+        elif 'resize_long' in kwargs:
+            self.resize_type = 2
+            self.resize_long = kwargs.get('resize_long', 960)
+        else:
+            self.limit_side_len = 736
+            self.limit_type = 'min'
+
+    def __call__(self, data):
+        img = data['image']
+        src_h, src_w, _ = img.shape
+
+        if self.resize_type == 0:
+            # img, shape = self.resize_image_type0(img)
+            img, [ratio_h, ratio_w] = self.resize_image_type0(img)
+        elif self.resize_type == 2:
+            img, [ratio_h, ratio_w] = self.resize_image_type2(img)
+        else:
+            # img, shape = self.resize_image_type1(img)
+            img, [ratio_h, ratio_w] = self.resize_image_type1(img)
+        data['image'] = img
+        data['shape'] = np.array([src_h, src_w, ratio_h, ratio_w])
+        return data
+
+    def resize_image_type1(self, img):
+        resize_h, resize_w = self.image_shape
+        ori_h, ori_w = img.shape[:2]  # (h, w, c)
+        ratio_h = float(resize_h) / ori_h
+        ratio_w = float(resize_w) / ori_w
+        img = cv2.resize(img, (int(resize_w), int(resize_h)))
+        # return img, np.array([ori_h, ori_w])
+        return img, [ratio_h, ratio_w]
+
+    def resize_image_type0(self, img):
+        """
+        resize image to a size multiple of 32 which is required by the network
+        args:
+            img(array): array with shape [h, w, c]
+        return(tuple):
+            img, (ratio_h, ratio_w)
+        """
+        limit_side_len = self.limit_side_len
+        h, w, c = img.shape
+
+        # limit the max side
+        if self.limit_type == 'max':
+            if max(h, w) > limit_side_len:
+                if h > w:
+                    ratio = float(limit_side_len) / h
+                else:
+                    ratio = float(limit_side_len) / w
+            else:
+                ratio = 1.
+        elif self.limit_type == 'min':
+            if min(h, w) < limit_side_len:
+                if h < w:
+                    ratio = float(limit_side_len) / h
+                else:
+                    ratio = float(limit_side_len) / w
+            else:
+                ratio = 1.
+        elif self.limit_type == 'resize_long':
+            ratio = float(limit_side_len) / max(h, w)
+        else:
+            raise Exception('not support limit type, image ')
+        resize_h = int(h * ratio)
+        resize_w = int(w * ratio)
+
+        resize_h = max(int(round(resize_h / 32) * 32), 32)
+        resize_w = max(int(round(resize_w / 32) * 32), 32)
+
+        try:
+            if int(resize_w) <= 0 or int(resize_h) <= 0:
+                return None, (None, None)
+            img = cv2.resize(img, (int(resize_w), int(resize_h)))
+        except:
+            print(img.shape, resize_w, resize_h)
+            sys.exit(0)
+        ratio_h = resize_h / float(h)
+        ratio_w = resize_w / float(w)
+        return img, [ratio_h, ratio_w]
+
+    def resize_image_type2(self, img):
+        h, w, _ = img.shape
+
+        resize_w = w
+        resize_h = h
+
+        if resize_h > resize_w:
+            ratio = float(self.resize_long) / resize_h
+        else:
+            ratio = float(self.resize_long) / resize_w
+
+        resize_h = int(resize_h * ratio)
+        resize_w = int(resize_w * ratio)
+
+        max_stride = 128
+        resize_h = (resize_h + max_stride - 1) // max_stride * max_stride
+        resize_w = (resize_w + max_stride - 1) // max_stride * max_stride
+        img = cv2.resize(img, (int(resize_w), int(resize_h)))
+        ratio_h = resize_h / float(h)
+        ratio_w = resize_w / float(w)
+
+        return img, [ratio_h, ratio_w]
+
+
+class E2EResizeForTest(object):
+    def __init__(self, **kwargs):
+        super(E2EResizeForTest, self).__init__()
+        self.max_side_len = kwargs['max_side_len']
+        self.valid_set = kwargs['valid_set']
+
+    def __call__(self, data):
+        img = data['image']
+        src_h, src_w, _ = img.shape
+        if self.valid_set == 'totaltext':
+            im_resized, [ratio_h, ratio_w] = self.resize_image_for_totaltext(
+                img, max_side_len=self.max_side_len)
+        else:
+            im_resized, (ratio_h, ratio_w) = self.resize_image(
+                img, max_side_len=self.max_side_len)
+        data['image'] = im_resized
+        data['shape'] = np.array([src_h, src_w, ratio_h, ratio_w])
+        return data
+
+    def resize_image_for_totaltext(self, im, max_side_len=512):
+
+        h, w, _ = im.shape
+        resize_w = w
+        resize_h = h
+        ratio = 1.25
+        if h * ratio > max_side_len:
+            ratio = float(max_side_len) / resize_h
+        resize_h = int(resize_h * ratio)
+        resize_w = int(resize_w * ratio)
+
+        max_stride = 128
+        resize_h = (resize_h + max_stride - 1) // max_stride * max_stride
+        resize_w = (resize_w + max_stride - 1) // max_stride * max_stride
+        im = cv2.resize(im, (int(resize_w), int(resize_h)))
+        ratio_h = resize_h / float(h)
+        ratio_w = resize_w / float(w)
+        return im, (ratio_h, ratio_w)
+
+    def resize_image(self, im, max_side_len=512):
+        """
+        resize image to a size multiple of max_stride which is required by the network
+        :param im: the resized image
+        :param max_side_len: limit of max image size to avoid out of memory in gpu
+        :return: the resized image and the resize ratio
+        """
+        h, w, _ = im.shape
+
+        resize_w = w
+        resize_h = h
+
+        # Fix the longer side
+        if resize_h > resize_w:
+            ratio = float(max_side_len) / resize_h
+        else:
+            ratio = float(max_side_len) / resize_w
+
+        resize_h = int(resize_h * ratio)
+        resize_w = int(resize_w * ratio)
+
+        max_stride = 128
+        resize_h = (resize_h + max_stride - 1) // max_stride * max_stride
+        resize_w = (resize_w + max_stride - 1) // max_stride * max_stride
+        im = cv2.resize(im, (int(resize_w), int(resize_h)))
+        ratio_h = resize_h / float(h)
+        ratio_w = resize_w / float(w)
+
+        return im, (ratio_h, ratio_w)
+
+
+class KieResize(object):
+    def __init__(self, **kwargs):
+        super(KieResize, self).__init__()
+        self.max_side, self.min_side = kwargs['img_scale'][0], kwargs[
+            'img_scale'][1]
+
+    def __call__(self, data):
+        img = data['image']
+        points = data['points']
+        src_h, src_w, _ = img.shape
+        im_resized, scale_factor, [ratio_h, ratio_w
+                                   ], [new_h, new_w] = self.resize_image(img)
+        resize_points = self.resize_boxes(img, points, scale_factor)
+        data['ori_image'] = img
+        data['ori_boxes'] = points
+        data['points'] = resize_points
+        data['image'] = im_resized
+        data['shape'] = np.array([new_h, new_w])
+        return data
+
+    def resize_image(self, img):
+        norm_img = np.zeros([1024, 1024, 3], dtype='float32')
+        scale = [512, 1024]
+        h, w = img.shape[:2]
+        max_long_edge = max(scale)
+        max_short_edge = min(scale)
+        scale_factor = min(max_long_edge / max(h, w),
+                           max_short_edge / min(h, w))
+        resize_w, resize_h = int(w * float(scale_factor) + 0.5), int(h * float(
+            scale_factor) + 0.5)
+        max_stride = 32
+        resize_h = (resize_h + max_stride - 1) // max_stride * max_stride
+        resize_w = (resize_w + max_stride - 1) // max_stride * max_stride
+        im = cv2.resize(img, (resize_w, resize_h))
+        new_h, new_w = im.shape[:2]
+        w_scale = new_w / w
+        h_scale = new_h / h
+        scale_factor = np.array(
+            [w_scale, h_scale, w_scale, h_scale], dtype=np.float32)
+        norm_img[:new_h, :new_w, :] = im
+        return norm_img, scale_factor, [h_scale, w_scale], [new_h, new_w]
+
+    def resize_boxes(self, im, points, scale_factor):
+        points = points * scale_factor
+        img_shape = im.shape[:2]
+        points[:, 0::2] = np.clip(points[:, 0::2], 0, img_shape[1])
+        points[:, 1::2] = np.clip(points[:, 1::2], 0, img_shape[0])
+        return points

magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/modeling/architectures/__init__.py

+# Copyright (c) 2020 PaddlePaddle Authors. 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.
+
+import copy
+
+__all__ = ["build_model"]
+
+
+def build_model(config, **kwargs):
+    from .base_model import BaseModel
+
+    config = copy.deepcopy(config)
+    module_class = BaseModel(config, **kwargs)
+    return module_class

magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/modeling/architectures/base_model.py

+from torch import nn
+
+from ..backbones import build_backbone
+from ..heads import build_head
+from ..necks import build_neck
+
+
+class BaseModel(nn.Module):
+    def __init__(self, config, **kwargs):
+        """
+        the module for OCR.
+        args:
+            config (dict): the super parameters for module.
+        """
+        super(BaseModel, self).__init__()
+
+        in_channels = config.get("in_channels", 3)
+        model_type = config["model_type"]
+        # build backbone, backbone is need for del, rec and cls
+        if "Backbone" not in config or config["Backbone"] is None:
+            self.use_backbone = False
+        else:
+            self.use_backbone = True
+            config["Backbone"]["in_channels"] = in_channels
+            self.backbone = build_backbone(config["Backbone"], model_type)
+            in_channels = self.backbone.out_channels
+
+        # build neck
+        # for rec, neck can be cnn,rnn or reshape(None)
+        # for det, neck can be FPN, BIFPN and so on.
+        # for cls, neck should be none
+        if "Neck" not in config or config["Neck"] is None:
+            self.use_neck = False
+        else:
+            self.use_neck = True
+            config["Neck"]["in_channels"] = in_channels
+            self.neck = build_neck(config["Neck"])
+            in_channels = self.neck.out_channels
+
+        # # build head, head is need for det, rec and cls
+        if "Head" not in config or config["Head"] is None:
+            self.use_head = False
+        else:
+            self.use_head = True
+            config["Head"]["in_channels"] = in_channels
+            self.head = build_head(config["Head"], **kwargs)
+
+        self.return_all_feats = config.get("return_all_feats", False)
+
+        self._initialize_weights()
+
+    def _initialize_weights(self):
+        # weight initialization
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode="fan_out")
+                if m.bias is not None:
+                    nn.init.zeros_(m.bias)
+            elif isinstance(m, nn.BatchNorm2d):
+                nn.init.ones_(m.weight)
+                nn.init.zeros_(m.bias)
+            elif isinstance(m, nn.Linear):
+                nn.init.normal_(m.weight, 0, 0.01)
+                if m.bias is not None:
+                    nn.init.zeros_(m.bias)
+            elif isinstance(m, nn.ConvTranspose2d):
+                nn.init.kaiming_normal_(m.weight, mode="fan_out")
+                if m.bias is not None:
+                    nn.init.zeros_(m.bias)
+
+    def forward(self, x):
+        y = dict()
+        if self.use_backbone:
+            x = self.backbone(x)
+        if isinstance(x, dict):
+            y.update(x)
+        else:
+            y["backbone_out"] = x
+        final_name = "backbone_out"
+        if self.use_neck:
+            x = self.neck(x)
+            if isinstance(x, dict):
+                y.update(x)
+            else:
+                y["neck_out"] = x
+            final_name = "neck_out"
+        if self.use_head:
+            x = self.head(x)
+        # for multi head, save ctc neck out for udml
+        if isinstance(x, dict) and "ctc_nect" in x.keys():
+            y["neck_out"] = x["ctc_neck"]
+            y["head_out"] = x
+        elif isinstance(x, dict):
+            y.update(x)
+        else:
+            y["head_out"] = x
+        if self.return_all_feats:
+            if self.training:
+                return y
+            elif isinstance(x, dict):
+                return x
+            else:
+                return {final_name: x}
+        else:
+            return x

magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/modeling/backbones/__init__.py

+# Copyright (c) 2020 PaddlePaddle Authors. 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.
+
+__all__ = ["build_backbone"]
+
+
+def build_backbone(config, model_type):
+    if model_type == "det":
+        from .det_mobilenet_v3 import MobileNetV3
+        from .rec_hgnet import PPHGNet_small
+        from .rec_lcnetv3 import PPLCNetV3
+
+        support_dict = [
+            "MobileNetV3",
+            "ResNet",
+            "ResNet_vd",
+            "ResNet_SAST",
+            "PPLCNetV3",
+            "PPHGNet_small",
+        ]
+    elif model_type == "rec" or model_type == "cls":
+        from .rec_hgnet import PPHGNet_small
+        from .rec_lcnetv3 import PPLCNetV3
+        from .rec_mobilenet_v3 import MobileNetV3
+        from .rec_svtrnet import SVTRNet
+        from .rec_mv1_enhance import MobileNetV1Enhance
+
+        support_dict = [
+            "MobileNetV1Enhance",
+            "MobileNetV3",
+            "ResNet",
+            "ResNetFPN",
+            "MTB",
+            "ResNet31",
+            "SVTRNet",
+            "ViTSTR",
+            "DenseNet",
+            "PPLCNetV3",
+            "PPHGNet_small",
+        ]
+    else:
+        raise NotImplementedError
+
+    module_name = config.pop("name")
+    assert module_name in support_dict, Exception(
+        "when model typs is {}, backbone only support {}".format(
+            model_type, support_dict
+        )
+    )
+    module_class = eval(module_name)(**config)
+    return module_class

magic_pdf/model/sub_modules/ocr/paddleocr2pytorch/pytorchocr/modeling/backbones/det_mobilenet_v3.py

+from torch import nn
+
+from ..common import Activation
+
+
+def make_divisible(v, divisor=8, min_value=None):
+    if min_value is None:
+        min_value = divisor
+    new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
+    if new_v < 0.9 * v:
+        new_v += divisor
+    return new_v
+
+
+class ConvBNLayer(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        kernel_size,
+        stride,
+        padding,
+        groups=1,
+        if_act=True,
+        act=None,
+        name=None,
+    ):
+        super(ConvBNLayer, self).__init__()
+        self.if_act = if_act
+        self.conv = nn.Conv2d(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            groups=groups,
+            bias=False,
+        )
+
+        self.bn = nn.BatchNorm2d(
+            out_channels,
+        )
+        if self.if_act:
+            self.act = Activation(act_type=act, inplace=True)
+
+    def forward(self, x):
+        x = self.conv(x)
+        x = self.bn(x)
+        if self.if_act:
+            x = self.act(x)
+        return x
+
+
+class SEModule(nn.Module):
+    def __init__(self, in_channels, reduction=4, name=""):
+        super(SEModule, self).__init__()
+        self.avg_pool = nn.AdaptiveAvgPool2d(1)
+        self.conv1 = nn.Conv2d(
+            in_channels=in_channels,
+            out_channels=in_channels // reduction,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=True,
+        )
+        self.relu1 = Activation(act_type="relu", inplace=True)
+        self.conv2 = nn.Conv2d(
+            in_channels=in_channels // reduction,
+            out_channels=in_channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=True,
+        )
+        self.hard_sigmoid = Activation(act_type="hard_sigmoid", inplace=True)
+
+    def forward(self, inputs):
+        outputs = self.avg_pool(inputs)
+        outputs = self.conv1(outputs)
+        outputs = self.relu1(outputs)
+        outputs = self.conv2(outputs)
+        outputs = self.hard_sigmoid(outputs)
+        outputs = inputs * outputs
+        return outputs
+
+
+class ResidualUnit(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        mid_channels,
+        out_channels,
+        kernel_size,
+        stride,
+        use_se,
+        act=None,
+        name="",
+    ):
+        super(ResidualUnit, self).__init__()
+        self.if_shortcut = stride == 1 and in_channels == out_channels
+        self.if_se = use_se
+
+        self.expand_conv = ConvBNLayer(
+            in_channels=in_channels,
+            out_channels=mid_channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            if_act=True,
+            act=act,
+            name=name + "_expand",
+        )
+        self.bottleneck_conv = ConvBNLayer(
+            in_channels=mid_channels,
+            out_channels=mid_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=int((kernel_size - 1) // 2),
+            groups=mid_channels,
+            if_act=True,
+            act=act,
+            name=name + "_depthwise",
+        )
+        if self.if_se:
+            self.mid_se = SEModule(mid_channels, name=name + "_se")
+        self.linear_conv = ConvBNLayer(
+            in_channels=mid_channels,
+            out_channels=out_channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            if_act=False,
+            act=None,
+            name=name + "_linear",
+        )
+
+    def forward(self, inputs):
+        x = self.expand_conv(inputs)
+        x = self.bottleneck_conv(x)
+        if self.if_se:
+            x = self.mid_se(x)
+        x = self.linear_conv(x)
+        if self.if_shortcut:
+            x = inputs + x
+        return x
+
+
+class MobileNetV3(nn.Module):
+    def __init__(
+        self, in_channels=3, model_name="large", scale=0.5, disable_se=False, **kwargs
+    ):
+        """
+        the MobilenetV3 backbone network for detection module.
+        Args:
+            params(dict): the super parameters for build network
+        """
+        super(MobileNetV3, self).__init__()
+
+        self.disable_se = disable_se
+
+        if model_name == "large":
+            cfg = [
+                # k, exp, c,  se,     nl,  s,
+                [3, 16, 16, False, "relu", 1],
+                [3, 64, 24, False, "relu", 2],
+                [3, 72, 24, False, "relu", 1],
+                [5, 72, 40, True, "relu", 2],
+                [5, 120, 40, True, "relu", 1],
+                [5, 120, 40, True, "relu", 1],
+                [3, 240, 80, False, "hard_swish", 2],
+                [3, 200, 80, False, "hard_swish", 1],
+                [3, 184, 80, False, "hard_swish", 1],
+                [3, 184, 80, False, "hard_swish", 1],
+                [3, 480, 112, True, "hard_swish", 1],
+                [3, 672, 112, True, "hard_swish", 1],
+                [5, 672, 160, True, "hard_swish", 2],
+                [5, 960, 160, True, "hard_swish", 1],
+                [5, 960, 160, True, "hard_swish", 1],
+            ]
+            cls_ch_squeeze = 960
+        elif model_name == "small":
+            cfg = [
+                # k, exp, c,  se,     nl,  s,
+                [3, 16, 16, True, "relu", 2],
+                [3, 72, 24, False, "relu", 2],
+                [3, 88, 24, False, "relu", 1],
+                [5, 96, 40, True, "hard_swish", 2],
+                [5, 240, 40, True, "hard_swish", 1],
+                [5, 240, 40, True, "hard_swish", 1],
+                [5, 120, 48, True, "hard_swish", 1],
+                [5, 144, 48, True, "hard_swish", 1],
+                [5, 288, 96, True, "hard_swish", 2],
+                [5, 576, 96, True, "hard_swish", 1],
+                [5, 576, 96, True, "hard_swish", 1],
+            ]
+            cls_ch_squeeze = 576
+        else:
+            raise NotImplementedError(
+                "mode[" + model_name + "_model] is not implemented!"
+            )
+
+        supported_scale = [0.35, 0.5, 0.75, 1.0, 1.25]
+        assert (
+            scale in supported_scale
+        ), "supported scale are {} but input scale is {}".format(supported_scale, scale)
+        inplanes = 16
+        # conv1
+        self.conv = ConvBNLayer(
+            in_channels=in_channels,
+            out_channels=make_divisible(inplanes * scale),
+            kernel_size=3,
+            stride=2,
+            padding=1,
+            groups=1,
+            if_act=True,
+            act="hard_swish",
+            name="conv1",
+        )
+
+        self.stages = nn.ModuleList()
+        self.out_channels = []
+        block_list = []
+        i = 0
+        inplanes = make_divisible(inplanes * scale)
+        for k, exp, c, se, nl, s in cfg:
+            se = se and not self.disable_se
+            if s == 2 and i > 2:
+                self.out_channels.append(inplanes)
+                self.stages.append(nn.Sequential(*block_list))
+                block_list = []
+            block_list.append(
+                ResidualUnit(
+                    in_channels=inplanes,
+                    mid_channels=make_divisible(scale * exp),
+                    out_channels=make_divisible(scale * c),
+                    kernel_size=k,
+                    stride=s,
+                    use_se=se,
+                    act=nl,
+                    name="conv" + str(i + 2),
+                )
+            )
+            inplanes = make_divisible(scale * c)
+            i += 1
+        block_list.append(
+            ConvBNLayer(
+                in_channels=inplanes,
+                out_channels=make_divisible(scale * cls_ch_squeeze),
+                kernel_size=1,
+                stride=1,
+                padding=0,
+                groups=1,
+                if_act=True,
+                act="hard_swish",
+                name="conv_last",
+            )
+        )
+        self.stages.append(nn.Sequential(*block_list))
+        self.out_channels.append(make_divisible(scale * cls_ch_squeeze))
+        # for i, stage in enumerate(self.stages):
+        #     self.add_sublayer(sublayer=stage, name="stage{}".format(i))
+
+    def forward(self, x):
+        x = self.conv(x)
+        out_list = []
+        for stage in self.stages:
+            x = stage(x)
+            out_list.append(x)
+        return out_list