test_model.py

import sys, os

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

# 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

model_name_or_path = './outputs/ckpt_2/'
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 = 16
gen_kwargs = {"max_length": max_length, "num_beams": num_beams}


# encoder data
url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
image = Image.open(requests.get(url, stream=True).raw)
# batch dim is added automatically
encoder_inputs = feature_extractor(images=image, return_tensors="jax")
pixel_values = encoder_inputs.pixel_values
print(f'pixel_values.shape = {pixel_values.shape}')

# decoder data
sentence = 'mon chien est mignon'
# IMPORTANT: For training/evaluation/attention_mask/loss
sentence += ' ' + tokenizer.eos_token
# batch dim is added automatically
decoder_inputs = tokenizer(sentence, return_tensors="jax")
print(decoder_inputs)
print(f'input_ids.shape = {decoder_inputs.input_ids.shape}')

# model data
inputs = dict(decoder_inputs)
inputs['pixel_values'] = pixel_values


logits = flax_vit_gpt2_lm(**inputs)[0]
preds = np.argmax(logits, axis=-1)
print('=' * 60)
print('Flax: Vit-GPT2-LM')
print('predicted token ids:')
print(preds)
print('=' * 60)


# Generation!
batch = {'pixel_values': pixel_values}
generation = flax_vit_gpt2_lm.generate(batch['pixel_values'], **gen_kwargs)
print('generation:')
print(generation)
print('=' * 60)

token_ids = np.array(generation.sequences)[0]
caption = tokenizer.decode(token_ids)
print(f'token_ids: {token_ids}')
print(f'caption: {caption}')
print('=' * 60)