nv_embed / benchmark.py
Tonic's picture
Initial Commit
2e1d4b5
raw
history blame
7.34 kB
import argparse
import os.path
import random
import time
from functools import partial
import evaluate
from tabulate import tabulate
from tqdm import tqdm
from texify.inference import batch_inference
from texify.model.model import load_model
from texify.model.processor import load_processor
from PIL import Image
from texify.settings import settings
import json
import base64
import io
from rapidfuzz.distance import Levenshtein
def normalize_text(text):
# Replace fences
text = text.replace("$", "")
text = text.replace("\[", "")
text = text.replace("\]", "")
text = text.replace("\(", "")
text = text.replace("\)", "")
text = text.strip()
return text
def score_text(predictions, references):
bleu = evaluate.load("bleu")
bleu_results = bleu.compute(predictions=predictions, references=references)
meteor = evaluate.load('meteor')
meteor_results = meteor.compute(predictions=predictions, references=references)
lev_dist = []
for p, r in zip(predictions, references):
lev_dist.append(Levenshtein.normalized_distance(p, r))
return {
'bleu': bleu_results["bleu"],
'meteor': meteor_results['meteor'],
'edit': sum(lev_dist) / len(lev_dist)
}
def image_to_pil(image):
decoded = base64.b64decode(image)
return Image.open(io.BytesIO(decoded))
def load_images(source_data):
images = [sd["image"] for sd in source_data]
images = [image_to_pil(image) for image in images]
return images
def inference_texify(source_data, model, processor):
images = load_images(source_data)
write_data = []
for i in tqdm(range(0, len(images), settings.BATCH_SIZE), desc="Texify inference"):
batch = images[i:i+settings.BATCH_SIZE]
text = batch_inference(batch, model, processor)
for j, t in enumerate(text):
eq_idx = i + j
write_data.append({"text": t, "equation": source_data[eq_idx]["equation"]})
return write_data
def inference_pix2tex(source_data):
from pix2tex.cli import LatexOCR
model = LatexOCR()
images = load_images(source_data)
write_data = []
for i in tqdm(range(len(images)), desc="Pix2tex inference"):
try:
text = model(images[i])
except ValueError:
# Happens when resize fails
text = ""
write_data.append({"text": text, "equation": source_data[i]["equation"]})
return write_data
def image_to_bmp(image):
img_out = io.BytesIO()
image.save(img_out, format="BMP")
return img_out
def inference_nougat(source_data, batch_size=1):
import torch
from nougat.postprocessing import markdown_compatible
from nougat.utils.checkpoint import get_checkpoint
from nougat.utils.dataset import ImageDataset
from nougat.utils.device import move_to_device
from nougat import NougatModel
# Load images, then convert to bmp format for nougat
images = load_images(source_data)
images = [image_to_bmp(image) for image in images]
predictions = []
ckpt = get_checkpoint(None, model_tag="0.1.0-small")
model = NougatModel.from_pretrained(ckpt)
if settings.TORCH_DEVICE_MODEL != "cpu":
move_to_device(model, bf16=settings.CUDA, cuda=settings.CUDA)
model.eval()
dataset = ImageDataset(
images,
partial(model.encoder.prepare_input, random_padding=False),
)
# Batch sizes higher than 1 explode memory usage on CPU/MPS
dataloader = torch.utils.data.DataLoader(
dataset,
batch_size=batch_size,
pin_memory=True,
shuffle=False,
)
for idx, sample in tqdm(enumerate(dataloader), desc="Nougat inference", total=len(dataloader)):
model.config.max_length = settings.MAX_TOKENS
model_output = model.inference(image_tensors=sample, early_stopping=False)
output = [markdown_compatible(o) for o in model_output["predictions"]]
predictions.extend(output)
return predictions
def main():
parser = argparse.ArgumentParser(description="Benchmark the performance of texify.")
parser.add_argument("--data_path", type=str, help="Path to JSON file with source images/equations", default=os.path.join(settings.DATA_DIR, "bench_data.json"))
parser.add_argument("--result_path", type=str, help="Path to JSON file to save results to.", default=os.path.join(settings.DATA_DIR, "bench_results.json"))
parser.add_argument("--max", type=int, help="Maximum number of images to benchmark.", default=None)
parser.add_argument("--pix2tex", action="store_true", help="Run pix2tex scoring", default=False)
parser.add_argument("--nougat", action="store_true", help="Run nougat scoring", default=False)
args = parser.parse_args()
source_path = os.path.abspath(args.data_path)
result_path = os.path.abspath(args.result_path)
os.makedirs(os.path.dirname(result_path), exist_ok=True)
model = load_model()
processor = load_processor()
with open(source_path, "r") as f:
source_data = json.load(f)
if args.max:
random.seed(1)
source_data = random.sample(source_data, args.max)
start = time.time()
predictions = inference_texify(source_data, model, processor)
times = {"texify": time.time() - start}
text = [normalize_text(p["text"]) for p in predictions]
references = [normalize_text(p["equation"]) for p in predictions]
scores = score_text(text, references)
write_data = {
"texify": {
"scores": scores,
"text": [{"prediction": p, "reference": r} for p, r in zip(text, references)]
}
}
if args.pix2tex:
start = time.time()
predictions = inference_pix2tex(source_data)
times["pix2tex"] = time.time() - start
p_text = [normalize_text(p["text"]) for p in predictions]
p_scores = score_text(p_text, references)
write_data["pix2tex"] = {
"scores": p_scores,
"text": [{"prediction": p, "reference": r} for p, r in zip(p_text, references)]
}
if args.nougat:
start = time.time()
predictions = inference_nougat(source_data)
times["nougat"] = time.time() - start
n_text = [normalize_text(p) for p in predictions]
n_scores = score_text(n_text, references)
write_data["nougat"] = {
"scores": n_scores,
"text": [{"prediction": p, "reference": r} for p, r in zip(n_text, references)]
}
score_table = []
score_headers = ["bleu", "meteor", "edit"]
score_dirs = ["⬆", "⬆", "⬇", "⬇"]
for method in write_data.keys():
score_table.append([method, *[write_data[method]["scores"][h] for h in score_headers], times[method]])
score_headers.append("time taken (s)")
score_headers = [f"{h} {d}" for h, d in zip(score_headers, score_dirs)]
print()
print(tabulate(score_table, headers=["Method", *score_headers]))
print()
print("Higher is better for BLEU and METEOR, lower is better for edit distance and time taken.")
print("Note that pix2tex is unbatched (I couldn't find a batch inference method in the docs), so time taken is higher than it should be.")
with open(result_path, "w") as f:
json.dump(write_data, f, indent=4)
if __name__ == "__main__":
main()