def test_unique_together(db: Session):
# GIVEN
db.add(SimpleModel(col_1="foo", col_2="bar"))
db.add(SimpleModel(col_1="foo", col_2="bar"))
# THEN 1
with pytest.raises(IntegrityError) as exc:
# WHEN
db.commit()
# THEN 2
assert "Duplicate entry" in str(exc)
def test_valid_value(self, db: Session):
# WHEN
instance = SimpleModel(type=SimpleModel.Type.SUPER)
db.add(instance)
db.commit()
# THEN
assert instance.type == SimpleModel.Type.SUPER
def test_invalid_value(self, db: Session):
# THEN 1
with pytest.raises(StatementError) as exc:
# WHEN
db.add(SimpleModel(type="INVALID_TYPE"))
db.commit()
# THEN 2
assert "is not a valid" in str(exc)
def visualize_entropy_linerp(self):
# We are visualizing only between w_1 and w_2 here
dummy_model = SimpleModel().to(self.config.firelab.device_name)
# acts_ent_linerp_train = compute_activations_entropy_linerp(self.model.w_1, self.model.w_2, dummy_model.nn, self.train_dataloader)
acts_ent_linerp_test = compute_activations_entropy_linerp(self.model.w_1, self.model.w_2, dummy_model.nn, self.val_dataloader)
weights_ent_linerp = compute_weights_entropy_linerp(self.model.w_1, self.model.w_2)
# self.writer.add_figure('activations_entropy_linerp/train', generate_acts_entropy_linerp_plot(acts_ent_linerp_train))
self.writer.add_figure('activations_entropy_linerp/test', generate_acts_entropy_linerp_plot(acts_ent_linerp_test))
self.writer.add_figure('weights_entropy_linerp', generate_weights_entropy_linerp_plot(weights_ent_linerp))
for i, w in enumerate(get_weights_linerp(self.model.w_1, self.model.w_2)):
self.writer.add_histogram('Weights histogram', w.detach().cpu(), i)
def visualize_linerp(self):
dummy_model = SimpleModel().to(self.config.firelab.device_name)
w1to2_linerp_scores_train = linerp(self.model.w_1, self.model.w_2, dummy_model, self.train_dataloader)
w1to2_linerp_scores_test = linerp(self.model.w_1, self.model.w_2, dummy_model, self.val_dataloader)
self.writer.add_figure('w_1 to w_2 linerp/accuracy', generate_linerp_plot(
[s[1] for s in w1to2_linerp_scores_train],
[s[1] for s in w1to2_linerp_scores_test],
))
self.writer.add_figure('w_1 to w_2 linerp/loss', generate_linerp_plot(
[s[0] for s in w1to2_linerp_scores_train],
[s[0] for s in w1to2_linerp_scores_test],
))
def init_torch_model_builder(self):
if self.config.hp.model_name == 'fast_resnet':
self.torch_model_builder = lambda: FastResNet(
n_classes=10, n_input_channels=self.config.hp.get('n_input_channels', 1)).nn
elif self.config.hp.model_name == 'resnet18':
self.torch_model_builder = lambda: ResNet18(
n_classes=10, n_input_channels=self.config.hp.get('n_input_channels', 1)).nn
elif self.config.hp.model_name == "vgg":
self.torch_model_builder = lambda: VGG11(
n_input_channels=self.config.hp.get('n_input_channels', 1),
use_bn=self.config.hp.get('use_bn', True)).model
elif self.config.hp.model_name == "simple":
self.torch_model_builder = lambda: SimpleModel().nn
elif self.config.hp.model_name == "conv":
self.torch_model_builder = lambda: ConvModel(self.config.hp.conv_model_config).nn
else:
raise NotImplementedError(f"Model {self.config.hp.model_name} is not supported")
def test_class_getters_aka_quasi_manager(db: Session):
# GIVEN
with freeze_time("2020-01-02 03:45") as older_timestamp:
timestamp = older_timestamp.time_to_freeze
SimpleModelFactory(type=SimpleModel.Type.NORMAL, created_at=timestamp)
SimpleModelFactory(type=SimpleModel.Type.SUPER, created_at=timestamp)
with freeze_time("2021-02-03 04:56") as newer_timestamp:
timestamp = newer_timestamp.time_to_freeze
obj_to_return = SimpleModelFactory(type=SimpleModel.Type.NORMAL,
created_at=timestamp)
SimpleModelFactory(type=SimpleModel.Type.SUPER, created_at=timestamp)
# WHEN
result = SimpleModel.get_latest_by_type(db, SimpleModel.Type.NORMAL)
# THEN
assert result == obj_to_return