def test_run_clm(self):
stream_handler = logging.StreamHandler(sys.stdout)
logger.addHandler(stream_handler)
tmp_dir = self.get_auto_remove_tmp_dir()
testargs = f"""
run_clm.py
--model_name_or_path distilgpt2
--train_file ./tests/fixtures/sample_text.txt
--validation_file ./tests/fixtures/sample_text.txt
--do_train
--do_eval
--block_size 128
--per_device_train_batch_size 5
--per_device_eval_batch_size 5
--num_train_epochs 2
--output_dir {tmp_dir}
--overwrite_output_dir
""".split()
if torch.cuda.device_count() > 1:
# Skipping because there are not enough batches to train the model + would need a drop_last to work.
return
if torch_device != "cuda":
testargs.append("--no_cuda")
with patch.object(sys, "argv", testargs):
run_clm.main()
result = get_results(tmp_dir)
self.assertLess(result["perplexity"], 100)
def test_run_clm_config_overrides(self):
# test that config_overrides works, despite the misleading dumps of default un-updated
# config via tokenizer
tmp_dir = self.get_auto_remove_tmp_dir()
testargs = f"""
run_clm.py
--model_type gpt2
--tokenizer_name gpt2
--train_file ./tests/fixtures/sample_text.txt
--output_dir {tmp_dir}
--config_overrides n_embd=10,n_head=2
""".split()
if torch_device != "cuda":
testargs.append("--no_cuda")
logger = run_clm.logger
with patch.object(sys, "argv", testargs):
with CaptureLogger(logger) as cl:
run_clm.main()
self.assertIn('"n_embd": 10', cl.out)
self.assertIn('"n_head": 2', cl.out)
