def test_freeze_pipeline():
exp = Experiment('experiments/sample-exp', read_only=True)
exp.config['trainer'].update(dict(steps=50, check_point=25))
# enable these
trainable = {
'include':
['src_embed', 'tgt_embed', 'generator', 'encoder:0', 'decoder:0,1']
}
exp.config['optim']['trainable'] = trainable
pipe = Pipeline(exp)
pipe.run(run_tests=False)
def test_finetune_pipeline_transformer():
codec_lib = 'nlcodec'
tmp_dir = tempfile.mkdtemp()
print(f"Testing finetune transformer: {tmp_dir}")
config = load_conf('experiments/sample-exp/conf.yml')
prep = config['prep']
prep.update(
dict(codec_lib=codec_lib,
char_coverage=0.9995,
finetune_src=prep['train_src'],
finetune_tgt=prep['train_tgt']))
exp = Experiment(tmp_dir, config=config, read_only=False)
exp.config['trainer'].update(
dict(steps=50,
check_point=25,
finetune_steps=100,
batch_size=400,
split_ratio=0.1,
dynamic_epoch=True))
Pipeline(exp).run()
assert exp.train_file.exists() or exp.train_db.exists()
assert exp.finetune_file.exists()
# TODO: add more assertions
print(f"Cleaning up {tmp_dir}")
shutil.rmtree(tmp_dir, ignore_errors=True)
def test_robertamt_2layer_init():
tmp_dir = tempfile.mkdtemp()
config = load_conf('experiments/pretrained/robertamt-xlmr-2layer.yml')
model_id = config['model_args']['model_id']
print(f"Testing {model_id} --> {tmp_dir}")
assert 'pretrainmatch' == config['prep'].get('codec_lib')
exp = Experiment(tmp_dir, config=config, read_only=False)
exp.config['trainer'].update(dict(steps=4, check_point=1))
Pipeline(exp).run(run_tests=False)
sanity_check_experiment(exp)
print(f"Cleaning up {tmp_dir}")
shutil.rmtree(tmp_dir, ignore_errors=True)
def test_parent_child_pipeline():
parent_dir = tempfile.mkdtemp()
# parent_dir = 'tmp-xyz-parent'
print(f"Making parent at {parent_dir}")
exp = Experiment(parent_dir,
config='experiments/transformer.test.yml',
read_only=False)
exp.config['trainer'].update(dict(steps=50, check_point=25))
Pipeline(exp).run(run_tests=False)
sanity_check_experiment(exp)
assert not exp.parent_model_state.exists()
child_config = load_conf('experiments/transformer.test.yml')
child_config.update({
'parent': {
'experiment': str(parent_dir),
'vocab': {
'shared': 'shared'
},
'model': {
'ensemble': 2
}
}
})
child_dir = tempfile.mkdtemp()
# child_dir = 'tmp-xyz-child'
print(f"Making child at {child_dir}")
exp = Experiment(child_dir, config=child_config, read_only=False)
exp.config['trainer'].update(dict(steps=50, check_point=25))
Pipeline(exp).run(run_tests=False)
sanity_check_experiment(exp)
assert exp.parent_model_state.exists()
for dir in [parent_dir, child_dir]:
print(f"Cleaning up {dir}")
shutil.rmtree(dir, ignore_errors=True)
def test_pipeline_transformer():
for codec_lib in ['sentpiece', 'nlcodec']:
tmp_dir = tempfile.mkdtemp()
config = load_conf('experiments/transformer.test.yml')
print(f"Testing {codec_lib} --> {tmp_dir}")
config['prep'].update({
'codec_lib': codec_lib,
'char_coverage': 0.9995
})
exp = Experiment(tmp_dir, config=config, read_only=False)
exp.config['trainer'].update(dict(steps=50, check_point=25))
exp.config['prep']['num_samples'] = 0
Pipeline(exp).run(run_tests=False)
sanity_check_experiment(exp)
print(f"Cleaning up {tmp_dir}")
shutil.rmtree(tmp_dir, ignore_errors=True)
def test_spark_prep():
tmp_dir = tempfile.mkdtemp()
try:
print(f"Testing dataprep on pyspark: {tmp_dir}")
config = load_conf('experiments/spark-bigdataprep.yml')
exp = Experiment(tmp_dir, config=config, read_only=False)
exp.config['trainer'].update(
dict(steps=50, check_point=25, batch_size=400))
Pipeline(exp).run()
assert exp._prepared_flag.exists()
assert exp._trained_flag.exists()
assert exp.train_file.exists() or exp.train_db.exists()
sanity_check_experiment(exp)
finally:
print(f"Cleaning up {tmp_dir}")
shutil.rmtree(tmp_dir, ignore_errors=True)
def test_prepared_pipeline():
exp = Experiment('experiments/sample-exp', read_only=True)
exp.config['trainer'].update(dict(steps=50, check_point=25))
pipe = Pipeline(exp)
pipe.run(run_tests=False)