transformers/test_onnx.py

from auto_transformers import AutoTransformers

import numpy
import onnx
import onnxruntime

from pathlib import Path
import warnings
import logging

# full_models = AutoTransformers.supported_model_names()
# checked_models = AutoTransformers.supported_model_names(format='onnx')
# models = [x for x in full_models if x not in checked_models]
models = ['distilbert-base-cased']

warnings.filterwarnings('ignore')

log_path = './transformers_onnx.log'

logger = logging.getLogger()

if Path(log_path).exists:
    handler = logging.FileHandler(log_path)
else:
    handler = logging.StreamHandler()
logger.handlers.append(handler)

logging.basicConfig(filename=log_path,
                    filemode='a',
                    format='%(asctime)s, %(msecs)d %(name)s %(levelname)s %(message)s',
                    datefmt='%H:%M:%S',
                    level=logging.info)


logging.info("Test")

for name in models:
    print(f'Model {name}:')
    saved_name = name.replace('/', '-')
    logging.info(f'\nConverting model {name} to {saved_name}:')
    try:
        op = AutoTransformers(model_name=name)
        out1 = op('hello, world.')
        logging.info('OP LOADED.')
    except Exception as e:
        logging.error(f'FAIL TO LOAD OP: {e}')
        continue
    try:
        op.save_model(format='onnx')
        logging.info('ONNX SAVED.')
    except Exception as e:
        logging.info(f'FAIL TO SAVE ONNX: {e}')
        continue
    try:
        try:
            onnx_model = onnx.load(f'saved/onnx/{saved_name}.onnx')
            onnx.checker.check_model(onnx_model)
        except Exception:
            saved_onnx = onnx.load(f'saved/onnx/{saved_name}.onnx', load_external_data=False)
            onnx.checker.check_model(saved_onnx)
        logging.info('ONNX CHECKED.')
    except Exception as e:
        logging.error(f'FAIL TO CHECK ONNX: {e}')
        continue
    try:
        sess = onnxruntime.InferenceSession(f'saved/onnx/{saved_name}.onnx',
                                            providers=onnxruntime.get_available_providers())
        inputs = op.tokenizer('hello, world.', return_tensors='np')
        out2 = sess.run(output_names=["last_hidden_state"], input_feed=dict(inputs))
        logging.info('ONNX WORKED.')
        if numpy.allclose(out1, out2, atol=1e-3):
            logging.info('Check accuracy: OK')
        else:
            logging.info('Check accuracy: atol is larger than 1e-3.')
    except Exception as e:
        logging.error(f'FAIL TO RUN ONNX: {e}')
        continue

print('Finished.')