[feat] Add texteller training script

examples/train_texteller/utils/__init__.py

@@ -0,0 +1,17 @@
+from .functional import (
+    collate_fn,
+    filter_fn,
+    tokenize_fn,
+)
+from .transforms import (
+    img_train_transform,
+    img_inf_transform,
+)
+
+__all__ = [
+    "collate_fn",
+    "filter_fn",
+    "tokenize_fn",
+    "img_train_transform",
+    "img_inf_transform",
+]

examples/train_texteller/utils/augraphy_pipe.py

@@ -0,0 +1,165 @@
+"""
+Custom augraphy pipeline for training
+
+This file implements a custom augraphy data augmentation pipeline. We found that using augraphy's
+default pipeline can cause significant degradation to formula images, potentially losing semantic
+information. Therefore, we carefully selected several common augmentation effects,
+adjusting their parameters and combination methods to preserve the original semantic information
+of the images as much as possible.
+"""
+
+from augraphy import (
+    InkColorSwap,
+    LinesDegradation,
+    OneOf,
+    Dithering,
+    InkBleed,
+    InkShifter,
+    NoiseTexturize,
+    BrightnessTexturize,
+    ColorShift,
+    DirtyDrum,
+    LightingGradient,
+    Brightness,
+    Gamma,
+    SubtleNoise,
+    Jpeg,
+    AugraphyPipeline,
+)
+import random
+
+
+def get_custom_augraphy():
+    pre_phase = []
+
+    ink_phase = [
+        InkColorSwap(
+            ink_swap_color="random",
+            ink_swap_sequence_number_range=(5, 10),
+            ink_swap_min_width_range=(2, 3),
+            ink_swap_max_width_range=(100, 120),
+            ink_swap_min_height_range=(2, 3),
+            ink_swap_max_height_range=(100, 120),
+            ink_swap_min_area_range=(10, 20),
+            ink_swap_max_area_range=(400, 500),
+            p=0.2,
+        ),
+        LinesDegradation(
+            line_roi=(0.0, 0.0, 1.0, 1.0),
+            line_gradient_range=(32, 255),
+            line_gradient_direction=(0, 2),
+            line_split_probability=(0.2, 0.4),
+            line_replacement_value=(250, 255),
+            line_min_length=(30, 40),
+            line_long_to_short_ratio=(5, 7),
+            line_replacement_probability=(0.4, 0.5),
+            line_replacement_thickness=(1, 3),
+            p=0.2,
+        ),
+        #  ============================
+        OneOf(
+            [
+                Dithering(
+                    dither="floyd-steinberg",
+                    order=(3, 5),
+                ),
+                InkBleed(
+                    intensity_range=(0.1, 0.2),
+                    kernel_size=random.choice([(7, 7), (5, 5), (3, 3)]),
+                    severity=(0.4, 0.6),
+                ),
+            ],
+            p=0.2,
+        ),
+        #  ============================
+        #  ============================
+        InkShifter(
+            text_shift_scale_range=(18, 27),
+            text_shift_factor_range=(1, 4),
+            text_fade_range=(0, 2),
+            blur_kernel_size=(5, 5),
+            blur_sigma=0,
+            noise_type="perlin",
+            p=0.2,
+        ),
+        #  ============================
+    ]
+
+    paper_phase = [
+        NoiseTexturize(
+            sigma_range=(3, 10),
+            turbulence_range=(2, 5),
+            texture_width_range=(300, 500),
+            texture_height_range=(300, 500),
+            p=0.2,
+        ),
+        BrightnessTexturize(texturize_range=(0.9, 0.99), deviation=0.03, p=0.2),
+    ]
+
+    post_phase = [
+        ColorShift(
+            color_shift_offset_x_range=(3, 5),
+            color_shift_offset_y_range=(3, 5),
+            color_shift_iterations=(2, 3),
+            color_shift_brightness_range=(0.9, 1.1),
+            color_shift_gaussian_kernel_range=(3, 3),
+            p=0.2,
+        ),
+        DirtyDrum(
+            line_width_range=(1, 6),
+            line_concentration=random.uniform(0.05, 0.15),
+            direction=random.randint(0, 2),
+            noise_intensity=random.uniform(0.6, 0.95),
+            noise_value=(64, 224),
+            ksize=random.choice([(3, 3), (5, 5), (7, 7)]),
+            sigmaX=0,
+            p=0.2,
+        ),
+        # =====================================
+        OneOf(
+            [
+                LightingGradient(
+                    light_position=None,
+                    direction=None,
+                    max_brightness=255,
+                    min_brightness=0,
+                    mode="gaussian",
+                    linear_decay_rate=None,
+                    transparency=None,
+                ),
+                Brightness(
+                    brightness_range=(0.9, 1.1),
+                    min_brightness=0,
+                    min_brightness_value=(120, 150),
+                ),
+                Gamma(
+                    gamma_range=(0.9, 1.1),
+                ),
+            ],
+            p=0.2,
+        ),
+        # =====================================
+        # =====================================
+        OneOf(
+            [
+                SubtleNoise(
+                    subtle_range=random.randint(5, 10),
+                ),
+                Jpeg(
+                    quality_range=(70, 95),
+                ),
+            ],
+            p=0.2,
+        ),
+        # =====================================
+    ]
+
+    pipeline = AugraphyPipeline(
+        ink_phase=ink_phase,
+        paper_phase=paper_phase,
+        post_phase=post_phase,
+        pre_phase=pre_phase,
+        log=False,
+    )
+
+    return pipeline

examples/train_texteller/utils/functional.py

@@ -0,0 +1,47 @@
+from typing import Any
+
+import torch
+from transformers import DataCollatorForLanguageModeling
+
+from texteller.constants import MAX_TOKEN_SIZE, MIN_HEIGHT, MIN_WIDTH
+
+
+def _left_move(x: torch.Tensor, pad_val):
+    assert len(x.shape) == 2, "x should be 2-dimensional"
+    lefted_x = torch.ones_like(x)
+    lefted_x[:, :-1] = x[:, 1:]
+    lefted_x[:, -1] = pad_val
+    return lefted_x
+
+
+def tokenize_fn(samples: dict[str, list[Any]], tokenizer=None) -> dict[str, list[Any]]:
+    assert tokenizer is not None, "tokenizer should not be None"
+    tokenized_formula = tokenizer(samples["latex_formula"], return_special_tokens_mask=True)
+    tokenized_formula["pixel_values"] = samples["image"]
+    return tokenized_formula
+
+
+def collate_fn(samples: list[dict[str, Any]], tokenizer=None) -> dict[str, list[Any]]:
+    assert tokenizer is not None, "tokenizer should not be None"
+    pixel_values = [dic.pop("pixel_values") for dic in samples]
+
+    clm_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
+
+    batch = clm_collator(samples)
+    batch["pixel_values"] = pixel_values
+    batch["decoder_input_ids"] = batch.pop("input_ids")
+    batch["decoder_attention_mask"] = batch.pop("attention_mask")
+
+    batch["labels"] = _left_move(batch["labels"], -100)
+
+    # convert list of Image to a tensor with (B, C, H, W)
+    batch["pixel_values"] = torch.stack(batch["pixel_values"], dim=0)
+    return batch
+
+
+def filter_fn(sample, tokenizer=None) -> bool:
+    return (
+        sample["image"].height > MIN_HEIGHT
+        and sample["image"].width > MIN_WIDTH
+        and len(tokenizer(sample["latex_formula"])["input_ids"]) < MAX_TOKEN_SIZE - 10
+    )

examples/train_texteller/utils/transforms.py

@@ -0,0 +1,154 @@
+import torch
+import random
+import numpy as np
+import cv2
+
+from torchvision.transforms import v2
+from typing import Any
+from PIL import Image
+from collections import Counter
+
+from texteller.constants import (
+    IMG_CHANNELS,
+    MAX_RESIZE_RATIO,
+    MIN_RESIZE_RATIO,
+)
+from texteller.utils import transform as inference_transform
+from .augraphy_pipe import get_custom_augraphy
+
+augraphy_pipeline = get_custom_augraphy()
+
+
+def trim_white_border(image: np.ndarray):
+    if len(image.shape) != 3 or image.shape[2] != 3:
+        raise ValueError("Image is not in RGB format or channel is not in third dimension")
+
+    if image.dtype != np.uint8:
+        raise ValueError(f"Image should stored in uint8")
+
+    corners = [tuple(image[0, 0]), tuple(image[0, -1]), tuple(image[-1, 0]), tuple(image[-1, -1])]
+    bg_color = Counter(corners).most_common(1)[0][0]
+    bg_color_np = np.array(bg_color, dtype=np.uint8)
+
+    h, w = image.shape[:2]
+    bg = np.full((h, w, 3), bg_color_np, dtype=np.uint8)
+
+    diff = cv2.absdiff(image, bg)
+    mask = cv2.cvtColor(diff, cv2.COLOR_BGR2GRAY)
+
+    threshold = 15
+    _, diff = cv2.threshold(mask, threshold, 255, cv2.THRESH_BINARY)
+
+    x, y, w, h = cv2.boundingRect(diff)
+
+    trimmed_image = image[y : y + h, x : x + w]
+
+    return trimmed_image
+
+
+def add_white_border(image: np.ndarray, max_size: int) -> np.ndarray:
+    randi = [random.randint(0, max_size) for _ in range(4)]
+    pad_height_size = randi[1] + randi[3]
+    pad_width_size = randi[0] + randi[2]
+    if pad_height_size + image.shape[0] < 30:
+        compensate_height = int((30 - (pad_height_size + image.shape[0])) * 0.5) + 1
+        randi[1] += compensate_height
+        randi[3] += compensate_height
+    if pad_width_size + image.shape[1] < 30:
+        compensate_width = int((30 - (pad_width_size + image.shape[1])) * 0.5) + 1
+        randi[0] += compensate_width
+        randi[2] += compensate_width
+    return v2.functional.pad(
+        torch.from_numpy(image).permute(2, 0, 1),
+        padding=randi,
+        padding_mode="constant",
+        fill=(255, 255, 255),
+    )
+
+
+def padding(images: list[torch.Tensor], required_size: int) -> list[torch.Tensor]:
+    images = [
+        v2.functional.pad(
+            img, padding=[0, 0, required_size - img.shape[2], required_size - img.shape[1]]
+        )
+        for img in images
+    ]
+    return images
+
+
+def random_resize(images: list[np.ndarray], minr: float, maxr: float) -> list[np.ndarray]:
+    if len(images[0].shape) != 3 or images[0].shape[2] != 3:
+        raise ValueError("Image is not in RGB format or channel is not in third dimension")
+
+    ratios = [random.uniform(minr, maxr) for _ in range(len(images))]
+    return [
+        # Anti-aliasing
+        cv2.resize(
+            img, (int(img.shape[1] * r), int(img.shape[0] * r)), interpolation=cv2.INTER_LANCZOS4
+        )
+        for img, r in zip(images, ratios)
+    ]
+
+
+def rotate(image: np.ndarray, min_angle: int, max_angle: int) -> np.ndarray:
+    # Get the center of the image to define the point of rotation
+    image_center = tuple(np.array(image.shape[1::-1]) / 2)
+
+    # Generate a random angle within the specified range
+    angle = random.randint(min_angle, max_angle)
+
+    # Get the rotation matrix for rotating the image around its center
+    rotation_mat = cv2.getRotationMatrix2D(image_center, angle, 1.0)
+
+    # Determine the size of the rotated image
+    cos = np.abs(rotation_mat[0, 0])
+    sin = np.abs(rotation_mat[0, 1])
+    new_width = int((image.shape[0] * sin) + (image.shape[1] * cos))
+    new_height = int((image.shape[0] * cos) + (image.shape[1] * sin))
+
+    # Adjust the rotation matrix to take into account translation
+    rotation_mat[0, 2] += (new_width / 2) - image_center[0]
+    rotation_mat[1, 2] += (new_height / 2) - image_center[1]
+
+    # Rotate the image with the specified border color (white in this case)
+    rotated_image = cv2.warpAffine(
+        image, rotation_mat, (new_width, new_height), borderValue=(255, 255, 255)
+    )
+
+    return rotated_image
+
+
+def ocr_aug(image: np.ndarray) -> np.ndarray:
+    if random.random() < 0.2:
+        image = rotate(image, -5, 5)
+    image = add_white_border(image, max_size=25).permute(1, 2, 0).numpy()
+    image = augraphy_pipeline(image)
+    return image
+
+
+def train_transform(images: list[Image.Image]) -> list[torch.Tensor]:
+    assert IMG_CHANNELS == 1, "Only support grayscale images for now"
+
+    images = [np.array(img.convert("RGB")) for img in images]
+    # random resize first
+    images = random_resize(images, MIN_RESIZE_RATIO, MAX_RESIZE_RATIO)
+    images = [trim_white_border(image) for image in images]
+
+    # OCR augmentation
+    images = [ocr_aug(image) for image in images]
+
+    # general transform pipeline
+    images = inference_transform(images)
+    return images
+
+
+def img_train_transform(samples: dict[str, list[Any]]) -> dict[str, list[Any]]:
+    processed_img = train_transform(samples["pixel_values"])
+    samples["pixel_values"] = processed_img
+    return samples
+
+
+def img_inf_transform(samples: dict[str, list[Any]]) -> dict[str, list[Any]]:
+    processed_img = inference_transform(samples["pixel_values"])
+    samples["pixel_values"] = processed_img
+    return samples