def test_linear_state_based_metrics():
m = LinearModel()
w = SimpleWrapper(m)
assert w != m
# History should reflect the change in metrics over for a window
# specified by the epoch range but continuing from where the model
# state was as of the last call to fit().
assert w.eval({}) == (0., 0.)
h = w.fit({'initial_epoch': 0, 'epochs': 3})
compare_dicts(
h, {
'epochs': [e for e in range(0, 3)],
'acc': [1, 2, 3],
'loss': [-1, -2, -3],
'val_acc': [1.5, 2.5, 3.5],
'val_loss': [-1.5, -2.5, -3.5]
})
# Run for a few more epochs
h = w.fit({'initial_epoch': 3, 'epochs': 5})
compare_dicts(
h, {
'epochs': [e for e in range(3, 5)],
'acc': [4, 5],
'loss': [-4, -5],
'val_acc': [4.5, 5.5],
'val_loss': [-4.5, -5.5]
})
# And now run for a few more, but reset the epoch range -- the key here
# is that the metrics should continue relative to the previous state,
# not the epoch range
h = w.fit({'initial_epoch': 1, 'epochs': 3})
compare_dicts(
h, {
'epochs': [e for e in range(1, 4)],
'acc': [6, 7, 8],
'loss': [-6, -7, -8],
'val_acc': [6.5, 7.5, 8.5],
'val_loss': [-6.5, -7.5, -8.5]
})
def test_quadratic_model():
m = QuadraticModel()
w = SimpleWrapper(m)
h = w.fit({'initial_epoch': 0, 'epochs': 5})
compare_dicts(
h, {
'epochs': [e for e in range(0, 3)],
'acc': [4, 1, 0, 1, 4],
'loss': [-4, -1, 0, -1, -4],
'val_acc': [4.5, 1.5, 0.5, 1.5, 4.5],
'val_loss': [-4.5, -1.5, -0.5, -1.5, -4.5]
})
def test_fit(keras_model):
good_results = {
'acc': [-76, -96, -28],
'val_loss': [4.75, 5.75, 6.75],
'loss': [4.5, 5.5, 6.5],
'val_acc': [4.25, 5.25, 6.25],
'epochs': [4, 5, 6]
}
m = SimpleWrapper(keras_model)
obs = m.fit(fit_args={'initial_epoch': 4, 'epochs': 7})
print('result:\n{}'.format(obs))
print('should be:\n{}'.format(good_results))
compare_dicts(obs, good_results)
def test_linear_copy():
from copy import copy
m = LinearModel()
w = SimpleWrapper(m)
w_copy = copy(w)
assert w != w_copy
assert w._model != w_copy._model
assert w.eval({}) == w_copy.eval({})
# Move original model ahead 3 epochs
h = w.fit({'initial_epoch': 0, 'epochs': 3})
compare_dicts(
h, {
'epochs': [e for e in range(0, 3)],
'acc': [1, 2, 3],
'loss': [-1, -2, -3],
'val_acc': [1.5, 2.5, 3.5],
'val_loss': [-1.5, -2.5, -3.5]
})
assert w.eval({}) == (-3, 3)
assert w.eval({}) != w_copy.eval({})
# Move copy ahead 2 epochs
h = w_copy.fit({'initial_epoch': 1, 'epochs': 3})
compare_dicts(
h, {
'epochs': [e for e in range(1, 3)],
'acc': [1, 2],
'loss': [-1, -2],
'val_acc': [1.5, 2.5],
'val_loss': [-1.5, -2.5]
})
assert w_copy.eval({}) == (-2, 2)
assert w.eval({}) != w_copy.eval({})
# Move original ahead 1 epochs
h = w.fit({'initial_epoch': 3, 'epochs': 4})
compare_dicts(
h, {
'epochs': [e for e in range(3, 4)],
'acc': [4],
'loss': [-4],
'val_acc': [4.5],
'val_loss': [-4.5]
})
assert w.eval({}) == (-4, 4)
assert w.eval({}) != w_copy.eval({})
# Move copy ahead 2 epochs
h = w_copy.fit({'initial_epoch': 3, 'epochs': 5})
compare_dicts(
h, {
'epochs': [e for e in range(3, 5)],
'acc': [3, 4],
'loss': [-3, -4],
'val_acc': [3.5, 4.5],
'val_loss': [-3.5, -4.5]
})
assert w_copy.eval({}) == (-4, 4)
assert w.eval({}) == w_copy.eval({})