markupdm / processing_markupdm.py
ktrk115's picture
Update processing_markupdm.py
0012b8d verified
"""Processor class for MarkupDM."""
import math
import re
import shutil
import subprocess
import tempfile
from pathlib import Path
import numpy as np
import torch
from .fonts import FontManager
from PIL import Image, ImageDraw
from transformers import (
ImageProcessingMixin,
PreTrainedModel,
PreTrainedTokenizerBase,
ProcessorMixin,
)
from transformers.utils import logging
logger = logging.get_logger(__name__)
MAXIMUM_DECODE_IMAGE_SIZE = 4096
IMG_FORMAT = "{:03d}.png"
FONT_FORMAT = "{:03d}.ttf"
class MarkupDMProcessor(ProcessorMixin): # type: ignore
attributes = ["tokenizer", "image_processor"]
# The superclass checks if the tokenizer is a subclass of `PreTrainedTokenizerBase`
tokenizer_class = "AutoTokenizer"
tokenizer: PreTrainedTokenizerBase
# and the image_processor is a subclass of `ImageProcessingMixin`.
image_processor_class = "AutoImageProcessor"
image_processor: ImageProcessingMixin
def __init__(
self,
tokenizer: PreTrainedTokenizerBase,
image_processor: ImageProcessingMixin,
):
super().__init__(tokenizer, image_processor)
# Extend the tokenizer if it has not been extended yet.
if "<begin_of_image>" not in tokenizer.additional_special_tokens:
self.extend_base_tokenizer(self.tokenizer)
# Regular expressions
boi = "<begin_of_image>"
img_sep = "<image_sep>"
self.re_img_size = re.compile(rf"{boi}(\d+){img_sep}(\d+){img_sep}")
self.re_svg_width = re.compile(r'<svg[^>]*\bwidth="(\d+)"[^>]*>')
self.re_svg_height = re.compile(r'<svg[^>]*\bheight="(\d+)"[^>]*>')
# Font manager
self.font_manager = None
def extend_base_tokenizer(self, tokenizer: PreTrainedTokenizerBase) -> None:
logger.info("Extending tokenizer...")
tokenizer.clean_up_tokenization_spaces = False
# Add special tokens
additional_special_tokens = [
"<begin_of_image>",
"<end_of_image>",
"<image_sep>",
"<image_token>",
]
logger.info(f"Add special tokens: {additional_special_tokens}")
tokenizer.add_special_tokens(
{"additional_special_tokens": additional_special_tokens},
replace_additional_special_tokens=False,
)
def __call__(
self,
svg: str | None = None,
images: list[Image.Image] | None = None,
filenames: list[str] | None = None,
vision_model: PreTrainedModel | None = None,
) -> dict:
# Process images
if not isinstance(images, list):
images = [images] # type: ignore
if len(images) > 0 and images[0] is not None:
output = self.preprocess_images(images)
output = self.encode_images(output, vision_model)
else:
output = {"width": [], "height": [], "image_ids": []}
# Process the entire example
output.update({"svg": svg, "filenames": filenames})
output = self.tokenize_example(output)
return output
def preprocess_images(self, images: list[Image.Image]) -> dict:
assert images is not None, "Images must be provided."
output: dict = {"image": [], "width": [], "height": []}
for image in images:
processed = self.image_processor(image)
for key, value in processed.items():
output[key].append(value)
# Stack tensors
output["image"] = torch.stack(output["image"])
return output
def encode_images(self, example: dict, vision_model: PreTrainedModel) -> dict:
if "images" in example and "width" not in example:
example = self.preprocess_images(example["images"])
assert vision_model is not None, "Vision model must be provided."
image = example.pop("image")
image = image.to(dtype=vision_model.dtype, device=vision_model.device)
with torch.inference_mode():
_, _, (_, _, image_ids) = vision_model.model.encode(image)
example["image_ids"] = list(image_ids.view(image.size(0), -1).cpu())
return example
def tokenize_example(self, example: dict) -> dict:
# Validate the input example
for key in ["svg", "filenames", "width", "height", "image_ids"]:
msg = f"Missing key: {key}."
if key in ["width", "height", "image_ids"]:
msg += " Images must be encoded first using `encode_images`."
assert example.get(key, None) is not None, msg
tokenizer = self.tokenizer
bos_id = tokenizer.bos_token_id
eos_id = tokenizer.eos_token_id
bos_id = bos_id if bos_id is not None else eos_id
boi_id = tokenizer.convert_tokens_to_ids("<begin_of_image>")
eoi_id = tokenizer.convert_tokens_to_ids("<end_of_image>")
img_sep_id = tokenizer.convert_tokens_to_ids("<image_sep>")
# Tokenize images and build a mapping from image filenames to tokens
name2token = {}
for filename, image_ids, width, height in zip(
example["filenames"],
example["image_ids"],
example["width"],
example["height"],
):
_image_ids = (image_ids + len(tokenizer)).tolist()
W_tokens = tokenizer.encode(str(width))
H_tokens = tokenizer.encode(str(height))
# Image tokens
image_tokens = [
boi_id,
*W_tokens,
img_sep_id,
*H_tokens,
img_sep_id,
*_image_ids,
eoi_id,
]
name2token[filename] = image_tokens
# Tokenize SVG
# TODO: remove bos_id as it seems to be not necessary in modern practice
tokens = [bos_id]
svg = example["svg"]
while svg:
# Find the start position of the next image filename
start, end = len(svg), len(svg)
for name in name2token.keys():
_start = svg.find(name)
if -1 < _start and _start < start:
start = _start
end = start + len(name)
# Tokenize the text before the image filename
tokens += tokenizer.encode(svg[:start])
# Append the tokenized image
if start < end:
tokens += name2token[svg[start:end]]
# Update the remaining text
svg = svg[end:]
tokens.append(eos_id)
# Format output data
input_ids = torch.tensor(tokens)
image_mask = input_ids >= len(tokenizer)
# Compute image position ids
image_pos_ids = torch.zeros_like(input_ids)
if len(example["image_ids"]) > 0:
length = example["image_ids"][0].size(0)
num_images = sum(image_mask) // length
image_pos_ids[image_mask] = torch.arange(length).repeat(num_images)
return {
"input_ids": input_ids,
"image_mask": image_mask,
"image_pos_ids": image_pos_ids,
}
def decode(
self,
tokens: torch.Tensor | np.ndarray,
vision_model: PreTrainedModel | None = None,
) -> dict:
tokenizer = self.tokenizer
bos = tokenizer.bos_token
eos = tokenizer.eos_token
bos = bos if bos is not None else eos
# Validate the input tokens
msg = "Should be reverted from FIM format before decoding."
for fim_type in ["prefix", "middle", "suffix"]:
token_id = tokenizer.convert_tokens_to_ids(f"<fim_{fim_type}>")
if token_id is None:
token_id = tokenizer.convert_tokens_to_ids(f"<|fim_{fim_type}|>")
assert token_id is not None, f"{fim_type} token not found"
assert token_id not in tokens, msg
tokens = torch.asarray(tokens).detach().cpu()
assert tokens.ndim == 1, "Tokens must be 1D."
boi_id = tokenizer.convert_tokens_to_ids("<begin_of_image>")
eoi_id = tokenizer.convert_tokens_to_ids("<end_of_image>")
# Decode tokens
svg = ""
images: list = []
filenames: list = []
while len(tokens) > 0:
# Find the start position of the next image filename
boi_idx = torch.where(tokens == boi_id)[0]
eoi_idx = torch.where(tokens == eoi_id)[0]
if boi_idx.size(0) > 0:
start = int(boi_idx[0].item())
end = int(eoi_idx[0].item()) + 1 if eoi_idx.size(0) > 0 else len(tokens)
assert start < end, "Invalid image tokens."
else:
start, end = len(tokens), len(tokens)
# Decode the tokens before the image tokens
svg += tokenizer.decode(tokens[:start])
# Decode the image tokens
if start < end:
# Extract image size
image_tokens = tokens[start:end]
image_text = tokenizer.decode(image_tokens)
matched = self.re_img_size.match(image_text)
if matched is not None:
width, height = map(int, matched.groups())
else:
width = self.image_processor.size
height = self.image_processor.size
# Decode tokens to PIL image
image_mask = image_tokens >= len(tokenizer)
image_ids = image_tokens[image_mask] - len(tokenizer)
image = self.decode_image(vision_model, image_ids, width, height)
filename = IMG_FORMAT.format(len(images))
svg += filename
images.append(image)
filenames.append(filename)
# Update the remaining tokens
tokens = tokens[end:]
# Remove consecutive <bos> and <eos>
svg = re.sub(rf"({re.escape(bos)})+", bos, svg)
svg = re.sub(rf"({re.escape(eos)})+", eos, svg)
# Extract the text between <bos> and <eos>
i_bos = svg.find(bos)
svg = svg[i_bos + len(bos) :] if i_bos > -1 else svg
i_eos = svg.find(eos, i_bos + 1)
svg = svg[:i_eos] if i_eos > -1 else svg
return {"svg": svg, "images": images, "filenames": filenames}
def decode_image(
self,
vision_model: PreTrainedModel | None = None,
image_ids: torch.Tensor | np.ndarray | None = None,
width: int | None = None,
height: int | None = None,
dummy_color: tuple[int, int, int, int] = (200,) * 4,
pad_value: int = 0,
) -> Image.Image:
# Prepare image size
width = width or self.image_processor.size
height = height or self.image_processor.size
width, height = self.compute_safe_image_size(width, height)
if vision_model is None and image_ids is None:
# Return a dummy image
return Image.new("RGBA", (width, height), dummy_color)
# Compute required length
assert vision_model is not None, "Vision model must be provided."
scale_factor = 2 ** (vision_model.model.encoder.num_resolutions - 1)
latent_size = self.image_processor.size // scale_factor
required_length = latent_size**2
# Pad image ids if necessary
image_ids = torch.asarray(image_ids, device=vision_model.device)
code_length = image_ids.shape[0] # type: ignore
if code_length < required_length:
pad_size = required_length - code_length
pad = torch.full((pad_size,), pad_value).to(image_ids)
image_ids = torch.cat([image_ids, pad])
# Decode image
with torch.inference_mode():
codebook_entry = vision_model.model.quantize.get_codebook_entry(
image_ids, (1, latent_size, latent_size, -1)
)
recon = vision_model.model.decode(codebook_entry)[0].float()
# Postprocess image
img = self.image_processor.postprocess(
recon, self.image_processor.size, self.image_processor.size
)
# Mask the padded area
if code_length < required_length:
img = self.mask_padded_area(img, code_length, scale_factor)
# Resize the image to the original size
img = img.resize((width, height), resample=self.image_processor.resample)
return img # type: ignore
def compute_safe_image_size(self, width: int, height: int) -> tuple[int, int]:
long_edge = max(width, height)
if MAXIMUM_DECODE_IMAGE_SIZE < long_edge:
scale = MAXIMUM_DECODE_IMAGE_SIZE / long_edge
width = min(max(int(width * scale), 1), MAXIMUM_DECODE_IMAGE_SIZE)
height = min(max(int(height * scale), 1), MAXIMUM_DECODE_IMAGE_SIZE)
return width, height
def mask_padded_area(
self,
img: Image.Image,
code_length: int,
scale_factor: int,
fill: tuple[int, int, int, int] = (200, 200, 200, 255),
) -> Image.Image:
draw = ImageDraw.Draw(img, mode="RGBA")
width, height = img.size
zw = math.ceil(width / scale_factor)
cw = code_length % zw
ch = code_length // zw
draw.polygon(
[
(cw * scale_factor, ch * scale_factor),
(width, ch * scale_factor),
(width, height),
(0, height),
(0, (ch + 1) * scale_factor),
(cw * scale_factor, (ch + 1) * scale_factor),
],
fill=fill,
)
return img
def set_font_manager(self, fonts_path: str | None = None) -> None:
self.font_manager = FontManager(fonts_path)
def render_preprocess(self, example: dict, out_dir: str | Path) -> None:
msg = "Font manager is not set. Call `set_font_manager` first."
assert self.font_manager is not None, msg
out_dir = Path(out_dir)
out_dir.mkdir(parents=True, exist_ok=True)
svg = example["svg"]
# Costruct style tag
found = set()
style_text = "text{dominant-baseline:text-before-edge}"
for i, text_str in enumerate(re.findall("<text[^>]*>", svg)):
matched = re.search('font-family="([^"]*)"', text_str)
if matched is None:
logger.warning(f"Font family not found in {text_str}")
continue
# Parse font attributes
font_family = matched.group(1)
is_bold = 'font-weight="bold"' in text_str
is_italic = 'font-style="italic"' in text_str
font_weight = "bold" if is_bold else "regular"
if is_italic:
font_style = "bolditalic" if is_bold else "italic"
else:
font_style = font_weight
key = (font_family, font_weight, font_style)
if key in found:
continue
font_bytes = self.font_manager.lookup(
font_family=font_family,
font_weight=font_weight,
font_style=font_style,
)
# @font-face
font_path = FONT_FORMAT.format(i)
font_face = "@font-face{"
font_face += f"font-family:'{font_family}';"
font_face += f"font-weight:{font_weight};"
font_face += f"font-style:{font_style};"
font_face += f"src:url('{font_path}');"
font_face += "}"
style_text += font_face
# Save font
Path(f"{out_dir}/{font_path}").write_bytes(font_bytes)
found.add(key)
# Insert style tag
matched = re.search("<svg[^>]*>", svg)
assert matched is not None, "SVG tag not found"
i = matched.span()[1]
style = f"<style>{style_text}</style>"
example["svg"] = svg[:i] + style + svg[i:]
def render(self, example: dict, save_dir: str | Path | None = None) -> Image.Image:
with tempfile.TemporaryDirectory() as tmp_dir:
self.render_preprocess(example, tmp_dir)
# Parse the SVG size
matched = self.re_svg_width.search(example["svg"])
assert matched is not None, "Width not found in SVG."
width = int(matched.group(1))
matched = self.re_svg_height.search(example["svg"])
assert matched is not None, "Height not found in SVG."
height = int(matched.group(1))
# Convert SVG to HTML
html = '<!DOCTYPE html><html><body style="margin: 0px">'
html += f"{example['svg']}</body></html>"
# Save HTML
Path(f"{tmp_dir}/index.html").write_text(html, encoding="utf-8")
# Save images
for img, filename in zip(example["images"], example["filenames"]):
Path(f"{tmp_dir}/{filename}").parent.mkdir(parents=True, exist_ok=True)
img.save(f"{tmp_dir}/{filename}")
# Take screenshot
command = [
"google-chrome",
"--headless",
"--disable-web-security",
"--allow-running-insecure-content",
"--no-sandbox",
"--disable-infobars",
"--hide-scrollbars",
"--disable-dev-shm-usage",
"--no-zygote",
f"--window-size={width},{height}",
f"--screenshot={tmp_dir}/screenshot.png",
f"{tmp_dir}/index.html",
]
subprocess.run(command, check=True, stderr=subprocess.DEVNULL)
# Load the screenshot as PIL image
out = Image.open(f"{tmp_dir}/screenshot.png")
size = (width, height)
out = out.resize(size, resample=Image.Resampling.LANCZOS) # type: ignore
# Copy the result if save_dir is specified
if save_dir is not None:
shutil.copytree(tmp_dir, save_dir, dirs_exist_ok=True)
return out