def test_local_config_missing_attrs():
with pytest.raises(AttributeError):
TensorFlowConfig()
with pytest.raises(AttributeError):
TensorFlowConfig(checkpoint_dir="foo")
with pytest.raises(AttributeError):
TensorFlowConfig(input_data_path="foo")
def test_local_config_save_model_params():
test_data_dir = Path(__file__).parent
target = test_data_dir / uuid.uuid4().hex
test_data_file = test_data_dir / "data" / "smol.txt"
lc = TensorFlowConfig(checkpoint_dir=target.as_posix(),
input_data_path=test_data_file.as_posix())
check = lc.save_model_params()
assert json.loads(open(check).read())
shutil.rmtree(target)
def test_train_batch_sp_tok(train_df, tmp_path):
config = TensorFlowConfig(
epochs=5,
field_delimiter=",",
checkpoint_dir=tmp_path,
input_data_path=PATH_HOLDER,
learning_rate=.01
)
tokenizer = SentencePieceTokenizerTrainer(
vocab_size=10000,
config=config
)
batcher = DataFrameBatch(
df=train_df,
config=config,
tokenizer=tokenizer
)
batcher.create_training_data()
batcher.train_all_batches()
batcher.generate_all_batch_lines(num_lines=_tok_gen_count, max_invalid=5000)
syn_df = batcher.batches_to_df()
assert syn_df.shape[0] == _tok_gen_count
# Generate with a RecordFactory
factory = batcher.create_record_factory(num_lines=_tok_gen_count, max_invalid=5000)
syn_df = factory.generate_all(output="df")
assert syn_df.shape[0] == _tok_gen_count
assert list(syn_df.columns) == list(train_df.columns)
assert factory.summary["valid_count"] == _tok_gen_count
def test_epoch_callback(train_df, tmp_path):
def epoch_callback(s: EpochState):
with open(tmp_path / 'callback_dump.txt', 'a') as f:
f.write(f'{s.epoch},{s.accuracy},{s.loss},{s.batch}\n')
config = TensorFlowConfig(epochs=5,
field_delimiter=",",
checkpoint_dir=tmp_path,
input_data_path=PATH_HOLDER,
learning_rate=.01,
epoch_callback=epoch_callback)
tokenizer = SentencePieceTokenizerTrainer(vocab_size=10000, config=config)
batcher = DataFrameBatch(batch_size=4,
df=train_df,
config=config,
tokenizer=tokenizer)
batcher.create_training_data()
batcher.train_all_batches()
with open(tmp_path / 'callback_dump.txt', 'r') as f:
lines = f.readlines()
assert len(lines) == 20
for i, line in enumerate(lines):
fields = line.strip().split(',')
assert len(fields) == 4
assert int(fields[0]) == i % 5
assert int(fields[3]) == i // 5
float(fields[1])
float(fields[2])
os.remove(tmp_path / 'callback_dump.txt')
def test_bad_epoch_callback(tmp_path):
with pytest.raises(ValueError) as err:
config = TensorFlowConfig(epochs=1,
field_delimiter=",",
checkpoint_dir=tmp_path,
input_data_path=PATH_HOLDER,
epoch_callback=1)
assert "must be a callable" in str(err)
def test_local_config(mkdir):
target = uuid.uuid4().hex
test_data_dir = Path(__file__).parent
test_data_file = test_data_dir / "data" / "smol.txt"
lc = TensorFlowConfig(checkpoint_dir=target, input_data_path=test_data_file.as_posix())
mkdir.assert_called
assert lc.epochs == 100
assert lc.input_data_path == test_data_file.as_posix()
assert lc.training_data_path == Path(target, "training_data.txt").as_posix()
def test_local_config_settings(mkdir):
lc = TensorFlowConfig(checkpoint_dir="foo", input_data_path="bar")
check = lc.as_dict()
assert check == {
"max_lines": 0,
"epochs": 100,
"epoch_callback": None,
"early_stopping": True,
"early_stopping_patience": 5,
"validation_split": True,
"best_model_metric": METRIC_VAL_LOSS,
"batch_size": 64,
"buffer_size": 10000,
"seq_length": 100,
"embedding_dim": 256,
"rnn_units": 256,
"dropout_rate": 0.2,
"rnn_initializer": "glorot_uniform",
"vocab_size": 20000,
"character_coverage": 1.0,
"pretrain_sentence_count": 1000000,
"dp": False,
"learning_rate": 0.01,
"dp_noise_multiplier": 0.1,
"dp_l2_norm_clip": 3.0,
"dp_microbatches": 64,
"gen_temp": 1.0,
"gen_chars": 0,
"gen_lines": 1000,
"max_line_len": 2048,
"save_all_checkpoints": False,
"save_best_model": True,
"checkpoint_dir": "foo",
"field_delimiter": None,
"field_delimiter_token": "<d>",
"overwrite": False,
"input_data_path": "bar",
"predict_batch_size": 64,
"reset_states": True,
"training_data_path": "foo/training_data.txt",
"model_type": "TensorFlowConfig"
}
def test_train_small_df(train_df, tmp_path):
small_df = train_df.sample(n=50)
config = TensorFlowConfig(epochs=5,
field_delimiter=",",
checkpoint_dir=tmp_path,
input_data_path=PATH_HOLDER)
batcher = DataFrameBatch(df=small_df, config=config)
batcher.create_training_data()
with pytest.raises(RuntimeError) as excinfo:
batcher.train_all_batches()
assert "Model training failed" in str(excinfo.value)
def tf_config():
target = test_data_dir / "ckpoint"
input_data = test_data_dir / "data" / "smol.txt"
if not target.exists():
target.mkdir()
config = TensorFlowConfig(
checkpoint_dir=target.as_posix(),
input_data_path=input_data.as_posix(),
field_delimiter=",",
predict_batch_size=1,
overwrite=True,
)
yield config
shutil.rmtree(target)
def test_train_batch_sp_tok(train_df, tmp_path):
config = TensorFlowConfig(epochs=5,
field_delimiter=",",
checkpoint_dir=tmp_path,
input_data_path=PATH_HOLDER,
learning_rate=.01)
tokenizer = SentencePieceTokenizerTrainer(vocab_size=10000, config=config)
batcher = DataFrameBatch(df=train_df, config=config, tokenizer=tokenizer)
batcher.create_training_data()
batcher.train_all_batches()
batcher.generate_all_batch_lines(num_lines=100, max_invalid=5000)
syn_df = batcher.batches_to_df()
assert syn_df.shape[0] == 100
def test_train_batch_sp(train_df, tmp_path):
config = TensorFlowConfig(epochs=1,
field_delimiter=",",
checkpoint_dir=tmp_path,
input_data_path=PATH_HOLDER)
batcher = DataFrameBatch(df=train_df, config=config)
batcher.create_training_data()
batcher.train_all_batches()
model_params = json.loads(
open(tmp_path / "batch_0" / const.MODEL_PARAMS).read())
assert model_params[const.MODEL_TYPE] == TensorFlowConfig.__name__
tok_params = json.loads(
open(tmp_path / "batch_0" /
BaseTokenizerTrainer.settings_fname).read())
assert tok_params[
"tokenizer_type"] == SentencePieceTokenizerTrainer.__name__
def test_train_batch_char_tok(train_df, tmp_path):
config = TensorFlowConfig(epochs=5,
field_delimiter=",",
checkpoint_dir=tmp_path,
input_data_path=PATH_HOLDER,
learning_rate=.01)
batcher = DataFrameBatch(df=train_df,
config=config,
tokenizer=CharTokenizerTrainer(config=config))
batcher.create_training_data()
batcher.train_all_batches()
tok_params = json.loads(
open(tmp_path / "batch_0" /
BaseTokenizerTrainer.settings_fname).read())
assert tok_params["tokenizer_type"] == CharTokenizerTrainer.__name__
batcher.generate_all_batch_lines(num_lines=100, max_invalid=5000)
syn_df = batcher.batches_to_df()
assert syn_df.shape[0] == 100
def test_no_delim_bad_start_string(tmpdir):
config = TensorFlowConfig(checkpoint_dir=tmpdir, input_data_path=tmpdir)
with pytest.raises(GenerationError):
Settings(config=config, start_string=123, tokenizer=mock_tokenizer)
def test_local_config_no_validation_split():
lc = TensorFlowConfig(checkpoint_dir="foo",
input_data_path="bar",
validation_split=False)
check = lc.as_dict()
assert check['best_model_metric'] == METRIC_LOSS
