def test_trainer_prediction_1(data):
model = deepcopy(data[0])
trainer = Trainer(model=model,
optimizer=Adam(lr=0.1),
loss_func=MSELoss(),
epochs=200)
trainer.extend(TensorConverter())
trainer.fit(*data[1], *data[1])
trainer = Trainer(model=model).extend(TensorConverter())
y_p = trainer.predict(data[1][0])
assert np.any(np.not_equal(y_p, data[1][1].numpy()))
assert np.allclose(y_p, data[1][1].numpy(), rtol=0, atol=0.2)
y_p, y_t = trainer.predict(*data[1])
assert np.any(np.not_equal(y_p, y_t))
assert np.allclose(y_p, y_t, rtol=0, atol=0.2)
val_set = DataLoader(TensorDataset(*data[1]), batch_size=50)
y_p, y_t = trainer.predict(dataset=val_set)
assert np.any(np.not_equal(y_p, y_t))
assert np.allclose(y_p, y_t, rtol=0, atol=0.2)
with pytest.raises(
RuntimeError,
match='parameters <x_in> and <dataset> are mutually exclusive'):
trainer.predict(*data[1], dataset='not none')
def test_trainer_prediction_2():
model = _Net(n_feature=2, n_hidden=10, n_output=2)
n_data = np.ones((100, 2))
x0 = np.random.normal(2 * n_data, 1)
y0 = np.zeros(100)
x1 = np.random.normal(-2 * n_data, 1)
y1 = np.ones(100)
x = np.vstack((x0, x1))
y = np.concatenate((y0, y1))
s = np.arange(x.shape[0])
np.random.shuffle(s)
x, y = x[s], y[s]
trainer = Trainer(model=model, optimizer=Adam(lr=0.1), loss_func=CrossEntropyLoss(), epochs=200)
trainer.extend(TensorConverter(x_dtype=torch.float32, y_dtype=torch.long, argmax=True))
trainer.fit(x, y)
y_p, y_t = trainer.predict(x, y)
assert y_p.shape == (200,)
assert np.all(y_p == y_t)
# trainer.reset()
val_set = DataLoader(ArrayDataset(x, y, dtypes=(torch.float, torch.long)), batch_size=20)
trainer.extend(TensorConverter(x_dtype=torch.float32, y_dtype=torch.long, auto_reshape=False))
y_p, y_t = trainer.predict(dataset=val_set)
assert y_p.shape == (200, 2)
y_p = np.argmax(y_p, 1)
assert np.all(y_p == y_t)
def test_trainer_fit_3(data):
model = deepcopy(data[0])
trainer = Trainer(model=model,
optimizer=Adam(),
loss_func=MSELoss(),
epochs=5)
trainer.fit(*data[1])
assert len(trainer.checkpoints.keys()) == 0
trainer.reset()
assert trainer.total_iterations == 0
assert trainer.total_epochs == 0
assert len(trainer.get_checkpoint()) == 0
trainer.fit(*data[1], checkpoint=True)
assert len(trainer.get_checkpoint()) == 5
assert isinstance(trainer.get_checkpoint(2), trainer.checkpoint_tuple)
assert isinstance(trainer.get_checkpoint('cp_2'), trainer.checkpoint_tuple)
with pytest.raises(TypeError, match='parameter <cp> must be str or int'):
trainer.get_checkpoint([])
trainer.reset(to=3, remove_checkpoints=False)
assert len(trainer.get_checkpoint()) == 5
assert isinstance(trainer.get_checkpoint(2), trainer.checkpoint_tuple)
assert isinstance(trainer.get_checkpoint('cp_2'), trainer.checkpoint_tuple)
trainer.reset(to='cp_3')
assert trainer.total_iterations == 0
assert trainer.total_epochs == 0
assert len(trainer.get_checkpoint()) == 0
with pytest.raises(
TypeError,
match='parameter <to> must be torch.nnModule, int, or str'):
trainer.reset(to=[])
# todo: need a real testing
trainer.fit(*data[1], checkpoint=True)
trainer.predict(*data[1], checkpoint=3)
trainer.reset()
trainer.fit(*data[1], checkpoint=lambda i: (True, f'new:{i}'))
assert len(trainer.get_checkpoint()) == 5
assert trainer.get_checkpoint() == list([f'new:{i + 1}' for i in range(5)])