python-doctr 0.11.0__py3-none-any.whl → 1.0.0__py3-none-any.whl

doctr/io/image/tensorflow.py

@@ -1,101 +0,0 @@
-# Copyright (C) 2021-2025, Mindee.
-
-# This program is licensed under the Apache License 2.0.
-# See LICENSE or go to <https://opensource.org/licenses/Apache-2.0> for full license details.
-
-
-import numpy as np
-import tensorflow as tf
-from PIL import Image
-from tensorflow.keras.utils import img_to_array
-
-from doctr.utils.common_types import AbstractPath
-
-__all__ = ["tensor_from_pil", "read_img_as_tensor", "decode_img_as_tensor", "tensor_from_numpy", "get_img_shape"]
-
-
-def tensor_from_pil(pil_img: Image.Image, dtype: tf.dtypes.DType = tf.float32) -> tf.Tensor:
-    """Convert a PIL Image to a TensorFlow tensor
-
-    Args:
-        pil_img: a PIL image
-        dtype: the output tensor data type
-
-    Returns:
-        decoded image as tensor
-    """
-    npy_img = img_to_array(pil_img)
-
-    return tensor_from_numpy(npy_img, dtype)
-
-
-def read_img_as_tensor(img_path: AbstractPath, dtype: tf.dtypes.DType = tf.float32) -> tf.Tensor:
-    """Read an image file as a TensorFlow tensor
-
-    Args:
-        img_path: location of the image file
-        dtype: the desired data type of the output tensor. If it is float-related, values will be divided by 255.
-
-    Returns:
-        decoded image as a tensor
-    """
-    if dtype not in (tf.uint8, tf.float16, tf.float32):
-        raise ValueError("insupported value for dtype")
-
-    img = tf.io.read_file(img_path)
-    img = tf.image.decode_jpeg(img, channels=3)
-
-    if dtype != tf.uint8:
-        img = tf.image.convert_image_dtype(img, dtype=dtype)
-        img = tf.clip_by_value(img, 0, 1)
-
-    return img
-
-
-def decode_img_as_tensor(img_content: bytes, dtype: tf.dtypes.DType = tf.float32) -> tf.Tensor:
-    """Read a byte stream as a TensorFlow tensor
-
-    Args:
-        img_content: bytes of a decoded image
-        dtype: the desired data type of the output tensor. If it is float-related, values will be divided by 255.
-
-    Returns:
-        decoded image as a tensor
-    """
-    if dtype not in (tf.uint8, tf.float16, tf.float32):
-        raise ValueError("insupported value for dtype")
-
-    img = tf.io.decode_image(img_content, channels=3)
-
-    if dtype != tf.uint8:
-        img = tf.image.convert_image_dtype(img, dtype=dtype)
-        img = tf.clip_by_value(img, 0, 1)
-
-    return img
-
-
-def tensor_from_numpy(npy_img: np.ndarray, dtype: tf.dtypes.DType = tf.float32) -> tf.Tensor:
-    """Read an image file as a TensorFlow tensor
-
-    Args:
-        npy_img: image encoded as a numpy array of shape (H, W, C) in np.uint8
-        dtype: the desired data type of the output tensor. If it is float-related, values will be divided by 255.
-
-    Returns:
-        same image as a tensor of shape (H, W, C)
-    """
-    if dtype not in (tf.uint8, tf.float16, tf.float32):
-        raise ValueError("insupported value for dtype")
-
-    if dtype == tf.uint8:
-        img = tf.convert_to_tensor(npy_img, dtype=dtype)
-    else:
-        img = tf.image.convert_image_dtype(npy_img, dtype=dtype)
-        img = tf.clip_by_value(img, 0, 1)
-
-    return img
-
-
-def get_img_shape(img: tf.Tensor) -> tuple[int, int]:
-    """Get the shape of an image"""
-    return img.shape[:2]

doctr/models/classification/magc_resnet/tensorflow.py

@@ -1,196 +0,0 @@
-# Copyright (C) 2021-2025, Mindee.
-
-# This program is licensed under the Apache License 2.0.
-# See LICENSE or go to <https://opensource.org/licenses/Apache-2.0> for full license details.
-
-import math
-from copy import deepcopy
-from functools import partial
-from typing import Any
-
-import tensorflow as tf
-from tensorflow.keras import activations, layers
-from tensorflow.keras.models import Sequential
-
-from doctr.datasets import VOCABS
-
-from ...utils import _build_model, load_pretrained_params
-from ..resnet.tensorflow import ResNet
-
-__all__ = ["magc_resnet31"]
-
-
-default_cfgs: dict[str, dict[str, Any]] = {
-    "magc_resnet31": {
-        "mean": (0.694, 0.695, 0.693),
-        "std": (0.299, 0.296, 0.301),
-        "input_shape": (32, 32, 3),
-        "classes": list(VOCABS["french"]),
-        "url": "https://doctr-static.mindee.com/models?id=v0.9.0/magc_resnet31-16aa7d71.weights.h5&src=0",
-    },
-}
-
-
-class MAGC(layers.Layer):
-    """Implements the Multi-Aspect Global Context Attention, as described in
-    <https://arxiv.org/pdf/1910.02562.pdf>`_.
-
-    Args:
-        inplanes: input channels
-        headers: number of headers to split channels
-        attn_scale: if True, re-scale attention to counteract the variance distibutions
-        ratio: bottleneck ratio
-        **kwargs
-    """
-
-    def __init__(
-        self,
-        inplanes: int,
-        headers: int = 8,
-        attn_scale: bool = False,
-        ratio: float = 0.0625,  # bottleneck ratio of 1/16 as described in paper
-        **kwargs,
-    ) -> None:
-        super().__init__(**kwargs)
-
-        self.headers = headers  # h
-        self.inplanes = inplanes  # C
-        self.attn_scale = attn_scale
-        self.ratio = ratio
-        self.planes = int(inplanes * ratio)
-
-        self.single_header_inplanes = int(inplanes / headers)  # C / h
-
-        self.conv_mask = layers.Conv2D(filters=1, kernel_size=1, kernel_initializer=tf.initializers.he_normal())
-
-        self.transform = Sequential(
-            [
-                layers.Conv2D(filters=self.planes, kernel_size=1, kernel_initializer=tf.initializers.he_normal()),
-                layers.LayerNormalization([1, 2, 3]),
-                layers.ReLU(),
-                layers.Conv2D(filters=self.inplanes, kernel_size=1, kernel_initializer=tf.initializers.he_normal()),
-            ],
-            name="transform",
-        )
-
-    def context_modeling(self, inputs: tf.Tensor) -> tf.Tensor:
-        b, h, w, c = (tf.shape(inputs)[i] for i in range(4))
-
-        # B, H, W, C -->> B*h, H, W, C/h
-        x = tf.reshape(inputs, shape=(b, h, w, self.headers, self.single_header_inplanes))
-        x = tf.transpose(x, perm=(0, 3, 1, 2, 4))
-        x = tf.reshape(x, shape=(b * self.headers, h, w, self.single_header_inplanes))
-
-        # Compute shorcut
-        shortcut = x
-        # B*h, 1, H*W, C/h
-        shortcut = tf.reshape(shortcut, shape=(b * self.headers, 1, h * w, self.single_header_inplanes))
-        # B*h, 1, C/h, H*W
-        shortcut = tf.transpose(shortcut, perm=[0, 1, 3, 2])
-
-        # Compute context mask
-        # B*h, H, W, 1
-        context_mask = self.conv_mask(x)
-        # B*h, 1, H*W, 1
-        context_mask = tf.reshape(context_mask, shape=(b * self.headers, 1, h * w, 1))
-        # scale variance
-        if self.attn_scale and self.headers > 1:
-            context_mask = context_mask / math.sqrt(self.single_header_inplanes)
-        # B*h, 1, H*W, 1
-        context_mask = activations.softmax(context_mask, axis=2)
-
-        # Compute context
-        # B*h, 1, C/h, 1
-        context = tf.matmul(shortcut, context_mask)
-        context = tf.reshape(context, shape=(b, 1, c, 1))
-        # B, 1, 1, C
-        context = tf.transpose(context, perm=(0, 1, 3, 2))
-        # Set shape to resolve shape when calling this module in the Sequential MAGCResnet
-        batch, chan = inputs.get_shape().as_list()[0], inputs.get_shape().as_list()[-1]
-        context.set_shape([batch, 1, 1, chan])
-        return context
-
-    def call(self, inputs: tf.Tensor, **kwargs) -> tf.Tensor:
-        # Context modeling: B, H, W, C  ->  B, 1, 1, C
-        context = self.context_modeling(inputs)
-        # Transform: B, 1, 1, C  ->  B, 1, 1, C
-        transformed = self.transform(context, **kwargs)
-        return inputs + transformed
-
-
-def _magc_resnet(
-    arch: str,
-    pretrained: bool,
-    num_blocks: list[int],
-    output_channels: list[int],
-    stage_downsample: list[bool],
-    stage_conv: list[bool],
-    stage_pooling: list[tuple[int, int] | None],
-    origin_stem: bool = True,
-    **kwargs: Any,
-) -> ResNet:
-    kwargs["num_classes"] = kwargs.get("num_classes", len(default_cfgs[arch]["classes"]))
-    kwargs["input_shape"] = kwargs.get("input_shape", default_cfgs[arch]["input_shape"])
-    kwargs["classes"] = kwargs.get("classes", default_cfgs[arch]["classes"])
-
-    _cfg = deepcopy(default_cfgs[arch])
-    _cfg["num_classes"] = kwargs["num_classes"]
-    _cfg["classes"] = kwargs["classes"]
-    _cfg["input_shape"] = kwargs["input_shape"]
-    kwargs.pop("classes")
-
-    # Build the model
-    model = ResNet(
-        num_blocks,
-        output_channels,
-        stage_downsample,
-        stage_conv,
-        stage_pooling,
-        origin_stem,
-        attn_module=partial(MAGC, headers=8, attn_scale=True),
-        cfg=_cfg,
-        **kwargs,
-    )
-    _build_model(model)
-
-    # Load pretrained parameters
-    if pretrained:
-        # The number of classes is not the same as the number of classes in the pretrained model =>
-        # skip the mismatching layers for fine tuning
-        load_pretrained_params(
-            model, default_cfgs[arch]["url"], skip_mismatch=kwargs["num_classes"] != len(default_cfgs[arch]["classes"])
-        )
-
-    return model
-
-
-def magc_resnet31(pretrained: bool = False, **kwargs: Any) -> ResNet:
-    """Resnet31 architecture with Multi-Aspect Global Context Attention as described in
-    `"MASTER: Multi-Aspect Non-local Network for Scene Text Recognition",
-    <https://arxiv.org/pdf/1910.02562.pdf>`_.
-
-    >>> import tensorflow as tf
-    >>> from doctr.models import magc_resnet31
-    >>> model = magc_resnet31(pretrained=False)
-    >>> input_tensor = tf.random.uniform(shape=[1, 224, 224, 3], maxval=1, dtype=tf.float32)
-    >>> out = model(input_tensor)
-
-    Args:
-        pretrained: boolean, True if model is pretrained
-        **kwargs: keyword arguments of the ResNet architecture
-
-    Returns:
-        A feature extractor model
-    """
-    return _magc_resnet(
-        "magc_resnet31",
-        pretrained,
-        [1, 2, 5, 3],
-        [256, 256, 512, 512],
-        [False] * 4,
-        [True] * 4,
-        [(2, 2), (2, 1), None, None],
-        False,
-        stem_channels=128,
-        **kwargs,
-    )