generate.py

import sys, os, datasets, json

current_path = os.path.dirname(os.path.abspath(__file__))
sys.path.append(current_path)

# jax
import jax

# Main model -  ViTGPT2LM
from vit_gpt2.modeling_flax_vit_gpt2_lm import FlaxViTGPT2LMForConditionalGeneration

# Vit - as encoder
from transformers import ViTFeatureExtractor
from PIL import Image
import requests
import numpy as np

# GPT2 / GPT2LM - as decoder
from transformers import ViTFeatureExtractor, GPT2Tokenizer

ckpt_no = 5
model_name_or_path = f'./outputs/ckpt_{ckpt_no}/'
flax_vit_gpt2_lm = FlaxViTGPT2LMForConditionalGeneration.from_pretrained(model_name_or_path)

vit_model_name = 'google/vit-base-patch16-224-in21k'
feature_extractor = ViTFeatureExtractor.from_pretrained(vit_model_name)

gpt2_model_name = 'asi/gpt-fr-cased-small'
tokenizer = GPT2Tokenizer.from_pretrained(gpt2_model_name)

max_length = 32
num_beams = 8
gen_kwargs = {"max_length": max_length, "num_beams": num_beams}


@jax.jit
def predict_fn(pixel_values):

    return flax_vit_gpt2_lm.generate(pixel_values, **gen_kwargs)

def predict(image):

    # batch dim is added automatically
    encoder_inputs = feature_extractor(images=image, return_tensors="jax")
    pixel_values = encoder_inputs.pixel_values

    # generation
    generation = predict_fn(pixel_values)

    token_ids = np.array(generation.sequences)[0]
    caption = tokenizer.decode(token_ids)

    return caption, token_ids


if __name__ == '__main__':

    from datetime import datetime

    split = 'val'
    image_id = 322141
    p = f'/home/33611/caption/{split}2014/COCO_{split}2014_{str(image_id).zfill(12)}.jpg'
    image = Image.open(p)
    caption, token_ids = predict(image)
    image.close()

    print(f'token_ids: {token_ids}')
    print(f'caption: {caption}')

    ds = datasets.load_dataset('./coco_dataset_script.py', data_dir='/home/33611/caption/')
    ds = ds['validation']
    ds = ds.select(range(100))

    predictions = []
    for ex in ds:

        p = ex['image_file']
        image = Image.open(p)
        s = datetime.now()
        caption, token_ids = predict(image)
        caption = caption.replace('<s>', '').replace('</s>', '').replace('<pad>', '').strip()
        image.close()
        e = datetime.now()
        e = (e - s).total_seconds()
        print(f'    timing: {e}')
        print(f'   caption: {ex["fr"]}')
        print(f'prediction: {caption}')
        print('-' * 20)
        ex['pred'] = caption
        predictions.append(ex)

    with open(f'ckpt_{ckpt_no}_preds.json', 'w', encoding='UTF-8') as fp:
        json.dump(predictions, fp, ensure_ascii=False, indent=4)