feat(model): add UniMERNet model configuration and processing files

- Add UnimerMBartConfig and UnimerSwinConfig classes
- Implement UnimerSwinImageProcessor for image preprocessing- Create necessary __init__.py files for module structure

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

+from .unimer_swin import UnimerSwinConfig, UnimerSwinModel, UnimerSwinImageProcessor
+from .unimer_mbart import UnimerMBartConfig, UnimerMBartModel, UnimerMBartForCausalLM
+from .modeling_unimernet import UnimernetModel
+
+__all__ = [
+    "UnimerSwinConfig",
+    "UnimerSwinModel",
+    "UnimerSwinImageProcessor",
+    "UnimerMBartConfig",
+    "UnimerMBartModel",
+    "UnimerMBartForCausalLM",
+    "UnimernetModel",
+]

magic_pdf/model/sub_modules/mfr/unimernet/unimernet_hf/modeling_unimernet.py

+import os
+import re
+import warnings
+from typing import Optional
+
+import torch
+from ftfy import fix_text
+
+from transformers import AutoConfig, AutoModel, AutoModelForCausalLM, AutoTokenizer, PretrainedConfig, PreTrainedModel
+from transformers import VisionEncoderDecoderConfig, VisionEncoderDecoderModel
+from transformers.models.vision_encoder_decoder.modeling_vision_encoder_decoder import logger as base_model_logger
+
+from .unimer_swin import UnimerSwinConfig, UnimerSwinModel, UnimerSwinImageProcessor
+from .unimer_mbart import UnimerMBartConfig, UnimerMBartForCausalLM
+
+AutoConfig.register(UnimerSwinConfig.model_type, UnimerSwinConfig)
+AutoConfig.register(UnimerMBartConfig.model_type, UnimerMBartConfig)
+AutoModel.register(UnimerSwinConfig, UnimerSwinModel)
+AutoModelForCausalLM.register(UnimerMBartConfig, UnimerMBartForCausalLM)
+
+
+# TODO: rewrite tokenizer
+class TokenizerWrapper:
+    def __init__(self, tokenizer):
+        self.tokenizer = tokenizer
+        self.pad_token_id = self.tokenizer.pad_token_id
+        self.bos_token_id = self.tokenizer.bos_token_id
+        self.eos_token_id = self.tokenizer.eos_token_id
+
+    def __len__(self):
+        return len(self.tokenizer)
+
+    def tokenize(self, text, **kwargs):
+        return self.tokenizer(
+            text,
+            return_token_type_ids=False,
+            return_tensors="pt",
+            padding="longest",
+            truncation=True,
+            **kwargs,
+        )
+
+    def token2str(self, tokens) -> list:
+        generated_text = self.tokenizer.batch_decode(tokens, skip_special_tokens=True)
+        generated_text = [fix_text(text) for text in generated_text]
+        return generated_text
+
+    def detokenize(self, tokens):
+        toks = [self.tokenizer.convert_ids_to_tokens(tok) for tok in tokens]
+        for b in range(len(toks)):
+            for i in reversed(range(len(toks[b]))):
+                if toks[b][i] is None:
+                    toks[b][i] = ''
+                toks[b][i] = toks[b][i].replace('Ġ', ' ').strip()
+                if toks[b][i] in ([self.tokenizer.bos_token, self.tokenizer.eos_token, self.tokenizer.pad_token]):
+                    del toks[b][i]
+        return toks
+
+
+def latex_rm_whitespace(s: str):
+    """Remove unnecessary whitespace from LaTeX code.
+    """
+    text_reg = r'(\\(operatorname|mathrm|text|mathbf)\s?\*? {.*?})'
+    letter = r'[a-zA-Z]'
+    noletter = r'[\W_^\d]'
+    names = [x[0].replace(' ', '') for x in re.findall(text_reg, s)]
+    s = re.sub(text_reg, lambda _: str(names.pop(0)), s)
+    news = s
+    while True:
+        s = news
+        news = re.sub(r'(?!\\ )(%s)\s+?(%s)' % (noletter, noletter), r'\1\2', s)
+        news = re.sub(r'(?!\\ )(%s)\s+?(%s)' % (noletter, letter), r'\1\2', news)
+        news = re.sub(r'(%s)\s+?(%s)' % (letter, noletter), r'\1\2', news)
+        if news == s:
+            break
+    return s
+
+
+class UnimernetModel(VisionEncoderDecoderModel):
+    def __init__(
+        self,
+        config: Optional[PretrainedConfig] = None,
+        encoder: Optional[PreTrainedModel] = None,
+        decoder: Optional[PreTrainedModel] = None,
+    ):
+        # VisionEncoderDecoderModel's checking log has bug, disable for temp.
+        base_model_logger.disabled = True
+        try:
+            super().__init__(config, encoder, decoder)
+        finally:
+            base_model_logger.disabled = False
+
+        if not config or not hasattr(config, "_name_or_path"):
+            raise RuntimeError("config._name_or_path is required by UnimernetModel.")
+
+        model_path = config._name_or_path
+        self.transform = UnimerSwinImageProcessor()
+        self.tokenizer = TokenizerWrapper(AutoTokenizer.from_pretrained(model_path))
+        self._post_check()
+    
+    def _post_check(self):
+        tokenizer = self.tokenizer
+
+        if tokenizer.tokenizer.model_max_length != self.config.decoder.max_position_embeddings:
+            warnings.warn(
+                f"decoder.max_position_embeddings={self.config.decoder.max_position_embeddings}," +
+                f" but tokenizer.model_max_length={tokenizer.tokenizer.model_max_length}, will set" +
+                f" tokenizer.model_max_length to {self.config.decoder.max_position_embeddings}.")
+            tokenizer.tokenizer.model_max_length = self.config.decoder.max_position_embeddings
+
+        assert self.config.decoder.vocab_size == len(tokenizer)
+        assert self.config.decoder_start_token_id == tokenizer.bos_token_id
+        assert self.config.pad_token_id == tokenizer.pad_token_id
+
+    @classmethod
+    def from_checkpoint(cls, model_path: str, model_filename: str = "pytorch_model.pth", state_dict_strip_prefix="model.model."):
+        config = VisionEncoderDecoderConfig.from_pretrained(model_path)
+        config._name_or_path = model_path
+        config.encoder = UnimerSwinConfig(**vars(config.encoder))
+        config.decoder = UnimerMBartConfig(**vars(config.decoder))
+
+        encoder = UnimerSwinModel(config.encoder)
+        decoder = UnimerMBartForCausalLM(config.decoder)
+        model = cls(config, encoder, decoder)
+
+        # load model weights
+        model_file_path = os.path.join(model_path, model_filename)
+        checkpoint = torch.load(model_file_path, map_location="cpu", weights_only=True)
+        state_dict = checkpoint["model"] if "model" in checkpoint else checkpoint
+        if not state_dict:
+            raise RuntimeError("state_dict is empty.")
+        if state_dict_strip_prefix:
+            state_dict = {
+                k[len(state_dict_strip_prefix):] if k.startswith(state_dict_strip_prefix) else k: v
+                for k, v in state_dict.items()
+            }
+        missing_keys, unexpected_keys = model.load_state_dict(state_dict, strict=False)
+        if len(unexpected_keys) > 0:
+            warnings.warn("Unexpected key(s) in state_dict: {}.".format(", ".join(f'"{k}"' for k in unexpected_keys)))
+        if len(missing_keys) > 0:
+            raise RuntimeError("Missing key(s) in state_dict: {}.".format(", ".join(f'"{k}"' for k in missing_keys)))
+        return model
+
+    def forward_bak(self, samples):
+        pixel_values, text = samples["image"], samples["text_input"]
+
+        text_inputs = self.tokenizer.tokenize(text).to(pixel_values.device)
+        decoder_input_ids, decoder_attention_mask = text_inputs["input_ids"], text_inputs["attention_mask"]
+
+        num_channels = pixel_values.shape[1]
+        if num_channels == 1:
+            pixel_values = pixel_values.repeat(1, 3, 1, 1)
+
+        labels = decoder_input_ids * 1
+        labels = labels.masked_fill(labels == self.tokenizer.pad_token_id, -100)
+
+        loss = self.model(
+            pixel_values=pixel_values,
+            decoder_input_ids=decoder_input_ids[:, :-1],
+            decoder_attention_mask=decoder_attention_mask[:, :-1],
+            labels=labels[:, 1:],
+        ).loss
+        return {"loss": loss}
+
+    def generate(self, samples, do_sample: bool = False, temperature: float = 0.2, top_p: float = 0.95):
+        pixel_values = samples["image"]
+        num_channels = pixel_values.shape[1]
+        if num_channels == 1:
+            pixel_values = pixel_values.repeat(1, 3, 1, 1)
+        
+        kwargs = {}
+        if do_sample:
+            kwargs["temperature"] = temperature
+            kwargs["top_p"] = top_p
+        
+        outputs = super().generate(
+            pixel_values=pixel_values,
+            max_new_tokens=self.tokenizer.tokenizer.model_max_length, # required
+            decoder_start_token_id=self.tokenizer.tokenizer.bos_token_id,
+            do_sample=do_sample,
+            **kwargs,
+        )
+
+        outputs = outputs[:, 1:].cpu().numpy()
+        pred_tokens = self.tokenizer.detokenize(outputs)
+        pred_str = self.tokenizer.token2str(outputs)
+        fixed_str = [latex_rm_whitespace(s) for s in pred_str]
+        return {"pred_ids": outputs, "pred_tokens": pred_tokens, "pred_str": pred_str, "fixed_str": fixed_str}
+

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

+from .configuration_unimer_mbart import UnimerMBartConfig
+from .modeling_unimer_mbart import UnimerMBartModel, UnimerMBartForCausalLM
+
+__all__ = [
+    "UnimerMBartConfig",
+    "UnimerMBartModel",
+    "UnimerMBartForCausalLM",
+]

magic_pdf/model/sub_modules/mfr/unimernet/unimernet_hf/unimer_mbart/configuration_unimer_mbart.py

+# coding=utf-8
+# Copyright 2021, The Facebook AI Research Team and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""UnimerMBART model configuration"""
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class UnimerMBartConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`MBartModel`]. It is used to instantiate an MBART
+    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 MBART
+    [facebook/mbart-large-cc25](https://huggingface.co/facebook/mbart-large-cc25) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 50265):
+            Vocabulary size of the MBART model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`MBartModel`] or [`TFMBartModel`].
+        d_model (`int`, *optional*, defaults to 1024):
+            Dimensionality of the layers and the pooler layer.
+        qk_squeeze (`int`, *optional*, defaults to 2):
+            Squeeze ratio for query/key's output dimension. See the [UniMERNet paper](https://arxiv.org/abs/2404.15254).
+            Squeeze Attention maps the query and key to a lower-dimensional space without excessive loss of information,
+            thereby accelerating the computation of attention.
+        encoder_layers (`int`, *optional*, defaults to 12):
+            Number of encoder layers.
+        decoder_layers (`int`, *optional*, defaults to 12):
+            Number of decoder layers.
+        encoder_attention_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        decoder_attention_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        decoder_ffn_dim (`int`, *optional*, defaults to 4096):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
+        encoder_ffn_dim (`int`, *optional*, defaults to 4096):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
+        activation_function (`str` or `function`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        dropout (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        activation_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for activations inside the fully connected layer.
+        classifier_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for classifier.
+        max_position_embeddings (`int`, *optional*, defaults to 1024):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        init_std (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        encoder_layerdrop (`float`, *optional*, defaults to 0.0):
+            The LayerDrop probability for the encoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
+            for more details.
+        decoder_layerdrop (`float`, *optional*, defaults to 0.0):
+            The LayerDrop probability for the decoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
+            for more details.
+        scale_embedding (`bool`, *optional*, defaults to `False`):
+            Scale embeddings by diving by sqrt(d_model).
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models)
+        forced_eos_token_id (`int`, *optional*, defaults to 2):
+            The id of the token to force as the last generated token when `max_length` is reached. Usually set to
+            `eos_token_id`.
+
+    Example:
+
+    ```python
+    >>> from transformers import MBartConfig, MBartModel
+
+    >>> # Initializing a MBART facebook/mbart-large-cc25 style configuration
+    >>> configuration = MBartConfig()
+
+    >>> # Initializing a model (with random weights) from the facebook/mbart-large-cc25 style configuration
+    >>> model = MBartModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "unimer-mbart"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    attribute_map = {"num_attention_heads": "encoder_attention_heads", "hidden_size": "d_model"}
+
+    def __init__(
+        self,
+        vocab_size=50265,
+        max_position_embeddings=1024,
+        encoder_layers=12,
+        encoder_ffn_dim=4096,
+        encoder_attention_heads=16,
+        decoder_layers=12,
+        decoder_ffn_dim=4096,
+        decoder_attention_heads=16,
+        encoder_layerdrop=0.0,
+        decoder_layerdrop=0.0,
+        use_cache=True,
+        is_encoder_decoder=True,
+        activation_function="gelu",
+        d_model=1024,
+        qk_squeeze=2,
+        dropout=0.1,
+        attention_dropout=0.0,
+        activation_dropout=0.0,
+        init_std=0.02,
+        classifier_dropout=0.0,
+        scale_embedding=False,
+        pad_token_id=1,
+        bos_token_id=0,
+        eos_token_id=2,
+        forced_eos_token_id=2,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.d_model = d_model
+        self.qk_squeeze = qk_squeeze
+        self.encoder_ffn_dim = encoder_ffn_dim
+        self.encoder_layers = encoder_layers
+        self.encoder_attention_heads = encoder_attention_heads
+        self.decoder_ffn_dim = decoder_ffn_dim
+        self.decoder_layers = decoder_layers
+        self.decoder_attention_heads = decoder_attention_heads
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.activation_dropout = activation_dropout
+        self.activation_function = activation_function
+        self.init_std = init_std
+        self.encoder_layerdrop = encoder_layerdrop
+        self.decoder_layerdrop = decoder_layerdrop
+        self.classifier_dropout = classifier_dropout
+        self.use_cache = use_cache
+        self.num_hidden_layers = encoder_layers
+        self.scale_embedding = scale_embedding  # scale factor will be sqrt(d_model) if True
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            is_encoder_decoder=is_encoder_decoder,
+            forced_eos_token_id=forced_eos_token_id,
+            **kwargs,
+        )

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

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

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

+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Donut Swin Transformer model configuration"""
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class UnimerSwinConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`UnimerSwinModel`]. It is used to instantiate a
+    Donut model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of the Donut
+    [naver-clova-ix/donut-base](https://huggingface.co/naver-clova-ix/donut-base) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        image_size (`int`, *optional*, defaults to 224):
+            The size (resolution) of each image.
+        patch_size (`int`, *optional*, defaults to 4):
+            The size (resolution) of each patch.
+        num_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        embed_dim (`int`, *optional*, defaults to 96):
+            Dimensionality of patch embedding.
+        depths (`list(int)`, *optional*, defaults to `[2, 2, 6, 2]`):
+            Depth of each layer in the Transformer encoder.
+        num_heads (`list(int)`, *optional*, defaults to `[3, 6, 12, 24]`):
+            Number of attention heads in each layer of the Transformer encoder.
+        window_size (`int`, *optional*, defaults to 7):
+            Size of windows.
+        mlp_ratio (`float`, *optional*, defaults to 4.0):
+            Ratio of MLP hidden dimensionality to embedding dimensionality.
+        qkv_bias (`bool`, *optional*, defaults to `True`):
+            Whether or not a learnable bias should be added to the queries, keys and values.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.0):
+            The dropout probability for all fully connected layers in the embeddings and encoder.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        drop_path_rate (`float`, *optional*, defaults to 0.1):
+            Stochastic depth rate.
+        hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder. If string, `"gelu"`, `"relu"`,
+            `"selu"` and `"gelu_new"` are supported.
+        use_absolute_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether or not to add absolute position embeddings to the patch embeddings.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-05):
+            The epsilon used by the layer normalization layers.
+
+    Example:
+
+    ```python
+    >>> from transformers import UnimerSwinConfig, UnimerSwinModel
+
+    >>> # Initializing a Donut naver-clova-ix/donut-base style configuration
+    >>> configuration = UnimerSwinConfig()
+
+    >>> # Randomly initializing a model from the naver-clova-ix/donut-base style configuration
+    >>> model = UnimerSwinModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "unimer-swin"
+
+    attribute_map = {
+        "num_attention_heads": "num_heads",
+        "num_hidden_layers": "num_layers",
+    }
+
+    def __init__(
+        self,
+        image_size=224,
+        patch_size=4,
+        num_channels=3,
+        embed_dim=96,
+        depths=[2, 2, 6, 2],
+        num_heads=[3, 6, 12, 24],
+        window_size=7,
+        mlp_ratio=4.0,
+        qkv_bias=True,
+        hidden_dropout_prob=0.0,
+        attention_probs_dropout_prob=0.0,
+        drop_path_rate=0.1,
+        hidden_act="gelu",
+        use_absolute_embeddings=False,
+        initializer_range=0.02,
+        layer_norm_eps=1e-5,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+        self.embed_dim = embed_dim
+        self.depths = depths
+        self.num_layers = len(depths)
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.mlp_ratio = mlp_ratio
+        self.qkv_bias = qkv_bias
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.drop_path_rate = drop_path_rate
+        self.hidden_act = hidden_act
+        self.use_absolute_embeddings = use_absolute_embeddings
+        self.layer_norm_eps = layer_norm_eps
+        self.initializer_range = initializer_range
+        # we set the hidden_size attribute in order to make Swin work with VisionEncoderDecoderModel
+        # this indicates the channel dimension after the last stage of the model
+        self.hidden_size = int(embed_dim * 2 ** (len(depths) - 1))

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

+from transformers.image_processing_utils import BaseImageProcessor
+from PIL import Image, ImageOps
+import numpy as np
+import cv2
+import albumentations as alb
+from albumentations.pytorch import ToTensorV2
+from torchvision.transforms.functional import resize
+
+
+# TODO: dereference cv2 if possible
+class UnimerSwinImageProcessor(BaseImageProcessor):
+    def __init__(
+            self,
+            image_size = [192, 672],
+        ):
+        self.input_size = [int(_) for _ in image_size]
+        assert len(self.input_size) == 2
+    
+        self.transform = alb.Compose(
+            [
+                alb.ToGray(always_apply=True),
+                alb.Normalize((0.7931, 0.7931, 0.7931), (0.1738, 0.1738, 0.1738)),
+                # alb.Sharpen()
+                ToTensorV2(),
+            ]
+        )
+
+    def __call__(self, item):
+        image = self.prepare_input(item)
+        return self.transform(image=np.array(image))['image'][:1]
+
+    @staticmethod
+    def crop_margin(img: Image.Image) -> Image.Image:
+        data = np.array(img.convert("L"))
+        data = data.astype(np.uint8)
+        max_val = data.max()
+        min_val = data.min()
+        if max_val == min_val:
+            return img
+        data = (data - min_val) / (max_val - min_val) * 255
+        gray = 255 * (data < 200).astype(np.uint8)
+
+        coords = cv2.findNonZero(gray)  # Find all non-zero points (text)
+        a, b, w, h = cv2.boundingRect(coords)  # Find minimum spanning bounding box
+        return img.crop((a, b, w + a, h + b))
+
+    def prepare_input(self, img: Image.Image, random_padding: bool = False):
+        """
+        Convert PIL Image to tensor according to specified input_size after following steps below:
+            - resize
+            - rotate (if align_long_axis is True and image is not aligned longer axis with canvas)
+            - pad
+        """
+        if img is None:
+            return
+        # crop margins
+        try:
+            img = self.crop_margin(img.convert("RGB"))
+        except OSError:
+            # might throw an error for broken files
+            return
+
+        if img.height == 0 or img.width == 0:
+            return
+
+        img = resize(img, min(self.input_size))
+        img.thumbnail((self.input_size[1], self.input_size[0]))
+        delta_width = self.input_size[1] - img.width
+        delta_height = self.input_size[0] - img.height
+        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
+        padding = (
+            pad_width,
+            pad_height,
+            delta_width - pad_width,
+            delta_height - pad_height,
+        )
+        return ImageOps.expand(img, padding)