def test_early_stopping(tmpdir):
model_params = {
'type': 'toy:DebugToyModel',
'model_dir': str(tmpdir / 'test_early_stopping/'),
'learning_rate': 0.01,
'num_classes': 2,
'num_steps_per_epoch': 1,
'hidden_neurons': 512
}
# Initialize the model
model = DebugToyModel(**model_params)
# Read training data
train_feeder = MemorySamplesFeeder(
np.asarray(np.arange(0, 10).reshape(10, 1), dtype=np.float32),
np.array([int(0 if i < 5 else 1) for i in range(10)],
dtype=np.int64,
ndmin=1),
batch_size=10,
feed_as_dict=False)
# Read validation data
validation_feeder = MemorySamplesFeeder(
np.asarray(np.arange(10, 20).reshape(10, 1), dtype=np.float32),
np.array([int(0 if i < 5 else 1) for i in range(10)],
dtype=np.int64,
ndmin=1),
batch_size=10,
feed_as_dict=False)
# Initialize the Trainer
trainer = Trainer()
train_params = {
'epochs': 12,
'save_summary_steps': 1,
'save_checkpoint_steps': 1,
'early_stopping_metric': 'loss',
'early_stopping_steps_without_decrease': 3,
'early_stopping_min_steps': 1,
'early_stopping_hook_run_every_steps': 1
}
# Start training process
trainer.train(model=model,
train_feeder=train_feeder,
eval_feeder=validation_feeder,
**train_params)
# Grab global step from model.estimator object
global_step = model.estimator().get_variable_value('global_step')
assert global_step < train_params['epochs']
def test_dict_input_with_2d(self):
x = {
'feature1': [[1, 11], [2, 12], [3, 13], [4, 14]]
}
y = {
'label': [4, 5, 6, 7]
}
feeder = MemorySamplesFeeder(x, y, shuffle=False, batch_size=2, feed_as_dict=True)
input_fn = feeder.get_input_fn(1)
results = read_iterator(input_fn().make_one_shot_iterator())
assert len(results) == 2
assert np.all(results[0][0]['feature1'] == np.array([[1, 11], [2, 12]]))
assert np.all(results[0][1]['label'] == np.array([4, 5]))
def test_pandas_input_with_dict(self):
x = pd.DataFrame(np.arange(0, 20).reshape(-1, 2),
columns=['feature_1', 'feature_2'])
y = pd.DataFrame(np.arange(100, 110))
feeder = MemorySamplesFeeder(x, y, shuffle=False, batch_size=2, feed_as_dict=True)
input_fn = feeder.get_input_fn(1)
results = read_iterator(input_fn().make_one_shot_iterator())
assert len(results) == 5
assert np.all(results[0][0]['feature_1'] == np.array([0, 2]))
assert np.all(results[0][0]['feature_2'] == np.array([1, 3]))
assert np.all(results[0][1][0] == np.array([100, 101]))
def test_repr(self):
x = np.arange(0, 100)
y = np.arange(1000, 1100)
res = repr(MemorySamplesFeeder(x=x, y=y, batch_size=16, shuffle=False))
assert res == "MemorySamplesFeeder(x=(100, 1), y=(100, 1), batch_size=16, " \
"shuffle=False, as_dict=True)"
def test_python_single_list(self):
x = list(range(20))
y = list(range(100, 120))
feeder = MemorySamplesFeeder(
x=x,
y=y,
shuffle=False, batch_size=2,
feed_as_dict=False)
input_fn = feeder.get_input_fn(1)
results = read_iterator(input_fn().make_one_shot_iterator())
expected = list(zip(batch_numpy_array(np.arange(20), 2),
batch_numpy_array(np.arange(100, 120), 2)))
assert equal_list_of_arrays(results, expected)
def test_np_with_batch_2(self):
x = np.arange(0, 10)
y = np.arange(0, 1).repeat(10)
feeder = MemorySamplesFeeder(
x=x,
y=y,
batch_size=2,
feed_as_dict=False,
shuffle=False
)
input_fn = feeder.get_input_fn(1)
results = read_iterator(input_fn().make_one_shot_iterator())
assert len(results) == 5
assert equal_list_of_arrays(results, list(zip(batch_numpy_array(x, 2), batch_numpy_array(y, 2))))
assert feeder.total_steps(1) == 5
def test_wrong_x_y_shape(self):
x = np.arange(0, 101)
y = np.arange(1000, 1100)
with pytest.raises(ValueError):
MemorySamplesFeeder(x=x,
y=y)
def test_np_with_batch_1(self):
x = np.arange(0, 10)
y = np.arange(0, 1).repeat(10)
feeder = MemorySamplesFeeder(
x=x,
y=y,
batch_size=1,
feed_as_dict=False,
shuffle=False
)
input_fn = feeder.get_input_fn(1)
results = read_iterator(input_fn().make_one_shot_iterator())
assert len(results) == 10
assert results == list(zip(x, y))
assert feeder.total_steps(1) == 10
def test_wrong_x_y_dict_shape(self):
x = {
'f1': np.arange(0, 12),
'f2': np.arange(20, 22),
}
y = np.arange(1000, 1100)
with pytest.raises(ValueError):
MemorySamplesFeeder(x=x,
y=y)
def test_shuffle_with_epochs(self):
x = np.arange(0, 100)
y = np.arange(1000, 1100)
feeder = MemorySamplesFeeder(
x=x,
y=y,
batch_size=10,
shuffle=True,
feed_as_dict=False
)
tf.set_random_seed(1)
input_fn = feeder.get_input_fn(5)
results = read_iterator(input_fn().make_one_shot_iterator())
assert len(results) == 50
for (batch_x, batch_y) in results:
assert np.all(batch_x == batch_y - 1000)
assert feeder.total_steps(5) == 50
def test_piecewise_learning_rate(tmpdir):
model_params = {
'type': 'toy:DebugToyModel',
'model_dir': str(tmpdir / 'test_piecewise_learning_rate/'),
'num_batches_per_epoch': 1,
'piecewise_learning_rate_schedule': "0.1:3:0.01:6:0.001",
'piecewise_lr_boundaries': [2, 5],
'piecewise_lr_values': [0.1, 0.01, 0.001],
'num_classes': 2,
'num_steps_per_epoch': 1,
'hidden_neurons': 10
}
train_params = {
'epochs': 10,
'save_summary_steps': 1,
'save_checkpoint_steps': 1
}
assert model_params['num_batches_per_epoch'] == 1
assert train_params['epochs'] == 10
# Create the ground truth learning rate values
ground_trouth_learning_rate_values = [0.1, 0.1, 0.1, 0.01, 0.01, 0.01, 0.001, 0.001, 0.001, 0.001]
# Initialize the model
model = DebugToyModel(**model_params)
# Read training data Be careful for batch_size == Number of examples, because it affects num_steps_per_epoch
train_feeder = MemorySamplesFeeder(np.asarray(np.arange(0, 10).reshape(10, 1), dtype=np.float32),
np.array([int(0 if i < 5 else 1)
for i in range(10)], dtype=np.int64, ndmin=1),
batch_size=10, feed_as_dict=False)
# Initialize the Trainer
trainer = Trainer()
# Start training process
trainer.train(model=model, train_feeder=train_feeder, eval_feeder=None, **train_params)
learning_rate_values = list()
# Parse events out tensorboard file
events_out_file = os.path.join(model_params['model_dir'],
[str(f) for f in os.listdir(model_params['model_dir'])
if f.startswith('events.out.tfevents')][0])
for e in tf.train.summary_iterator(events_out_file):
for v in e.summary.value:
if v.tag == 'learning_rate':
learning_rate_values.append(float('{:.7f}' .format(v.simple_value)))
# Testing condition
assert ground_trouth_learning_rate_values == learning_rate_values
def test_np_with_batch_1_as_dict(self):
x = np.arange(0, 10)
y = np.arange(0, 1).repeat(10)
feeder = MemorySamplesFeeder(
x=x,
y=y,
batch_size=1,
feed_as_dict=True,
shuffle=False
)
input_fn = feeder.get_input_fn(1)
results = read_iterator(input_fn().make_one_shot_iterator())
expected = [
({0: xi}, {0: yi})
for xi, yi in zip(x.reshape(-1, 1), y.reshape(-1, 1))
]
assert len(results) == 10
assert results == expected
assert feeder.total_steps(1) == 10
def test_np_with_epochs(self):
x = np.arange(0, 100)
y = np.arange(1000, 1100)
feeder = MemorySamplesFeeder(
x=x,
y=y,
batch_size=10,
feed_as_dict=False,
shuffle=False
)
input_fn = feeder.get_input_fn(5)
results = read_iterator(input_fn().make_one_shot_iterator())
assert len(results) == 50
assert equal_list_of_arrays(
results,
list(zip(batch_numpy_array(x, 10),
batch_numpy_array(y, 10))
) * 5
)
assert feeder.total_steps(5) == 50
def test_total_steps(self):
x = np.arange(0, 100)
y = np.arange(1000, 1100)
feeder = MemorySamplesFeeder(
x=x,
y=y,
batch_size=10,
shuffle=True,
feed_as_dict=False
)
assert feeder.total_steps(1) == 10
assert feeder.total_steps(2) == 20
feeder = MemorySamplesFeeder(
x=x,
y=y,
batch_size=7,
shuffle=True,
feed_as_dict=False
)
assert feeder.total_steps(1) == 15
assert feeder.total_steps(2) == 29
feeder = MemorySamplesFeeder(
x=x,
y=y,
batch_size=10,
shuffle=True,
feed_as_dict=True
)
assert feeder.total_steps(1) == 10
assert feeder.total_steps(2) == 20
