def test_metrics_multilabel(self):
n_classes = 2
classes_labels = ["t1", "t2"]
y_pred = torch.tensor([[[1, 0], [1, 1]], [[0, 1], [0, 0]]],
dtype=torch.float).permute(0, 2, 1)
y_true = torch.tensor([[[1, 1], [1, 0]], [[1, 1], [0, 0]]],
dtype=torch.long).permute(0, 2, 1)
params = {"type": "precision", "average": "micro", "multilabel": True}
precision = Metric.from_params(Params(params),
num_classes=n_classes,
classes_labels=classes_labels)
precision(y_true, y_pred)
val = precision.get_metric_value()
assert val == 3 / 4
params = {
"type": "precision",
"average": "micro",
"batch_average": True,
"multilabel": True,
}
precision = Metric.from_params(Params(params),
num_classes=n_classes,
classes_labels=classes_labels)
precision(y_true, y_pred)
val = precision.get_metric_value()
assert val == (2 / 3 + 1 / 1) / 2
def test_patches(self):
img = np.random.random((64, 64, 3))
mask = np.random.randint(0, 20, (63, 63))
patch_shape = 32
param = {"type": "sample_to_patches", "patch_shape": patch_shape}
sample_to_patches = Transform.from_params(Params(param))
res = sample_to_patches(image=img, mask=mask)
assert all(
[p.shape[:2] == (patch_shape, patch_shape) for p in res["image"]])
assert all(
[p.shape[:2] == (patch_shape, patch_shape) for p in res["mask"]])
patch_shape = (32, 2)
param = {"type": "sample_to_patches", "patch_shape": patch_shape}
sample_to_patches = Transform.from_params(Params(param))
res = sample_to_patches(image=img, mask=mask)
assert all([p.shape[:2] == patch_shape for p in res["image"]])
assert all([p.shape[:2] == patch_shape for p in res["mask"]])
with pytest.raises(ValueError):
sample_to_patches(image=np.ones((2, 32, 32, 3)))
with pytest.raises(IndexError):
sample_to_patches(image=np.ones((2, )))
with pytest.raises(ValueError):
patch_shape = (1, 1)
param = {"type": "sample_to_patches", "patch_shape": patch_shape}
sample_to_patches = Transform.from_params(Params(param))
res = sample_to_patches(image=img, mask=mask)
def test_warmup_lr(self):
named_parameters = [("x", torch.nn.Parameter())]
lr = 1
opt = Optimizer.from_params(Params({
"type": "sgd",
"lr": lr
}),
model_params=named_parameters)
warmup_start = 0.1
warmup_end = lr
warmup_duration = 10
gamma_exp = 0.95
sched_params = {
"type": "warmup",
"scheduler": {
"type": "exponential",
"gamma": gamma_exp
},
"warmup_start_value": warmup_start,
"warmup_end_value": warmup_end,
"warmup_duration": warmup_duration,
}
sched = Scheduler.from_params(Params(sched_params), optimizer=opt)
assert np.allclose(sched.get_last_lr()[0], warmup_start)
for _ in range(warmup_duration):
opt.step()
sched.step()
assert np.allclose(sched.get_last_lr()[0], warmup_end)
for i in range(50):
opt.step()
sched.step()
assert np.allclose(sched.get_last_lr()[0],
warmup_end * gamma_exp**(i + 1))
def test_metrics_multiclass(self):
n_classes = 3
classes_labels = ["b", "t1", "t2"]
logits = (torch.tensor([
[[[10, 1, 1], [10, 1, 1], [1, 1, 10], [1, 1, 10]]],
[[[1, 10, 1], [1, 1, 10], [1, 1, 10], [1, 10, 1]]],
]).to(torch.float).permute(0, 3, 1, 2))
y_true = torch.tensor([[[0, 1, 2, 2]], [[1, 2, 2, 0]]])
params = {"type": "precision", "average": "micro"}
precision = Metric.from_params(Params(params),
num_classes=n_classes,
classes_labels=classes_labels)
precision.get_metric_value()
precision(y_true, logits)
val = precision.get_metric_value()
assert val == 5 / 6
params = {
"type": "precision",
"average": "micro",
"batch_average": True,
}
precision = Metric.from_params(Params(params),
num_classes=n_classes,
classes_labels=classes_labels)
precision(y_true, logits)
val = precision.get_metric_value()
assert val == (2 / 2 + 3 / 4) / 2
def test_lr_concat(self):
named_parameters = [("x", torch.nn.Parameter())]
lr = 1
opt = Optimizer.from_params(Params({
"type": "sgd",
"lr": lr
}),
model_params=named_parameters)
gamma_exp = 0.95
gamma_plateau = 0.01
patience = 2
sched_params = {
"type":
"concat",
"schedulers": [
{
"type": "exponential",
"gamma": gamma_exp
},
{
"type": "reduce_on_plateau",
"mode": "max",
"factor": gamma_plateau,
"patience": patience,
"min_lr": 1e-3,
},
],
"durations": [
10,
],
}
sched = Scheduler.from_params(Params(sched_params), optimizer=opt)
assert np.allclose(sched.get_last_lr()[0], lr)
for i in range(10):
opt.step()
sched.step(0.1)
assert np.allclose(sched.get_last_lr()[0], lr * gamma_exp**(i + 1))
start_lr = lr * gamma_exp**10
sched.step(0.5)
assert np.allclose(sched.get_last_lr()[0], start_lr)
sched.step(0.4)
assert np.allclose(sched.get_last_lr()[0], start_lr)
sched.step(0.3)
assert np.allclose(sched.get_last_lr()[0], start_lr)
sched.step(0.3)
assert np.allclose(sched.get_last_lr()[0], start_lr * gamma_plateau)
# test min_lr
sched.step(0.3)
assert np.allclose(sched.get_last_lr()[0], start_lr * gamma_plateau)
sched.step(0.4)
assert np.allclose(sched.get_last_lr()[0], start_lr * gamma_plateau)
sched.step(0.3)
assert np.allclose(sched.get_last_lr()[0],
max(1e-3, start_lr * gamma_plateau**2))
def test_blur(self):
blur = Transform.from_params(Params({"type": "blur", "p": 1}))
img = np.random.random((128, 128, 3))
res = blur(image=img)
assert img.mean() != res["image"].mean()
blur = Transform.from_params(Params({"type": "blur", "p": 0}))
img = np.random.random((128, 128, 3))
res = blur(image=img)
assert img.mean() == res["image"].mean()
def test_read_jsonnet(self):
with pytest.raises(RuntimeError):
Params.from_file(self.FIXTURES_ROOT / "configs" /
"resnet50_unet_bad.jsonnet")
params = Params.from_file(self.FIXTURES_ROOT / "configs" /
"resnet50_unet.jsonnet")
assert len(params) == 3
assert params["encoder"]["type"] == "resnet50"
assert params["decoder"]["type"] == "unet"
assert params["decoder"]["decoder_channels"] == [512, 256, 128, 64, 32]
def test_normalizations_with_resnet(self):
params = Params({
"type": "resnet50",
"normalization": {
"type": "identity"
}
})
encoder = Encoder.from_params(params)
assert isinstance(encoder.layer4[0].bn1, Identity)
x = torch.zeros((2, 3, 4, 4)).normal_()
encoder.forward(x)
params = Params({
"type": "resnet50",
"normalization": {
"type": "batch_norm_2d"
}
})
encoder = Encoder.from_params(params)
assert isinstance(encoder.layer4[0].bn1, BatchNorm2d)
encoder.forward(x)
params = Params({
"type": "resnet50",
"normalization": {
"type": "batch_renorm_2d"
}
})
encoder = Encoder.from_params(params)
assert isinstance(encoder.layer4[0].bn1, BatchRenorm2d)
encoder.forward(x)
params = Params({
"type": "resnet50",
"normalization": {
"type": "group_norm"
}
})
with pytest.raises(ConfigurationError):
Encoder.from_params(params)
params = Params({
"type": "resnet50",
"normalization": {
"type": "group_norm",
"num_groups": 8
},
})
encoder = Encoder.from_params(params)
assert isinstance(encoder.layer4[0].bn1, GroupNorm)
encoder.forward(x)
def test_compose(self):
param = {
"type":
"compose",
"transforms": [
"blur",
{
"type": "horizontal_flip",
"always_apply": True
},
{
"type": "random_crop",
"height": 32,
"width": 32
},
],
"additional_targets": {
"image2": "image",
"mask1": "mask",
"label": "mask"
},
}
compose = Transform.from_params(Params(param))
img = np.random.random((128, 128, 3))
mask = np.random.randint(0, 20, (128, 128))
res = compose(image2=img, image=img, mask1=mask, label=mask)
assert np.allclose(res["image"], res["image2"])
assert np.allclose(res["mask1"], res["label"])
assert res["image"].shape[:2] == (32, 32)
assert res["mask1"].shape[:2] == (32, 32)
def test_union_str_list_str(self):
class A(FromParams):
def __init__(self, x: List[Tuple[Union[str, List[str]],
Dict[str, Any]]]):
self.x = x
a = A.from_params(Params({"x": [["test", {}]]}))
def test_make_param_groups(self):
model = Model.from_params(
Params({
"encoder": {
"type": "resnet50"
},
"decoder": {
"type": "unet",
"decoder_channels": [512, 256, 128, 64, 32],
},
"num_classes": 4,
}))
named_parameters = [x for x in model.named_parameters()]
alpha_decoder = 0.01
alpha_logits = 0.1
alpha_encoder = 0.0001
params = {
"param_groups": [
{
"regexes": "encoder",
"params": {
"alpha": alpha_encoder
}
},
{
"regexes":
["decoder.logits.*.weight", "decoder.logits.*.bias"],
"params": {
"alpha": alpha_logits
},
},
],
"alpha":
alpha_decoder,
}
reg = Regularizer.from_params(Params(params),
model_params=named_parameters)
assert reg.param_groups[0]["alpha"] == alpha_encoder
assert len(reg.param_groups[0]["params"]) == 161
assert reg.param_groups[2]["alpha"] == alpha_decoder
assert len(reg.param_groups[2]["params"]) == 20
assert reg.param_groups[1]["alpha"] == alpha_logits
assert len(reg.param_groups[1]["params"]) == 2
def test_assert_empty(self):
config_dict = {"test": "hello"}
params = Params(config_dict)
with pytest.raises(ConfigurationError):
params.assert_empty("dummy")
assert params.pop("test") == "hello"
params.assert_empty("dummy")
def test_trainer(self):
blocks = 2
num_channels = 32
params = Params({
"color_labels": {
"type":
"txt",
"label_text_file":
self.FIXTURES_ROOT / "dataset" / "multiclass" / "classes.txt",
},
"train_dataset": {
"type": "image_csv",
"csv_filename":
self.FIXTURES_ROOT / "dataset" / "multiclass" / "train.csv",
"base_dir": self.FIXTURES_ROOT / "dataset" / "multiclass",
},
"model": {
"encoder": "resnet50",
"decoder": {
"decoder_channels": [512, 256, 128, 64, 32]
}
# "loss": {"type": "dice"},
},
"metrics": [
"iou",
("iou_class", {
"type": "iou",
"average": None
}),
"precision",
],
"val_dataset": {
"type": "image_csv",
"csv_filename":
self.FIXTURES_ROOT / "dataset" / "multiclass" / "test.csv",
"base_dir": self.FIXTURES_ROOT / "dataset" / "multiclass",
},
"lr_scheduler": {
"type": "exponential",
"gamma": 0.95
},
"early_stopping": {
"patience": 20
},
"model_out_dir":
str(self.TEMPORARY_DIR / "model"),
"num_epochs":
2,
"evaluate_every_epoch":
1,
})
trainer = Trainer.from_params(params)
trainer.train()
def test_build_optimizers(self):
model = Model.from_params(
Params(
{
"encoder": {},
"decoder": {"decoder_channels": [512, 256, 128, 64, 32],},
"num_classes": 4,
}
)
)
named_parameters = [x for x in model.named_parameters()]
optimizers = Optimizer.get_available()
for optimizer in optimizers:
lr = 10
params = {"type": optimizer, "lr": lr}
opt = Optimizer.from_params(Params(params), model_params=named_parameters)
assert opt.state_dict()["param_groups"][0]["lr"] == lr
def test_build_regularizers(self):
model = Model.from_params(
Params({
"encoder": {},
"decoder": {
"decoder_channels": [512, 256, 128, 64, 32],
},
"num_classes": 4,
}))
named_parameters = [x for x in model.named_parameters()]
regularizers = Regularizer.get_available()
for regularizer in regularizers:
alpha = 10
params = {"type": regularizer, "alpha": alpha}
reg = Regularizer.from_params(Params(params),
model_params=named_parameters)
assert reg.param_groups[0]["alpha"] == alpha
def test_basic_from_params(self):
config_dict = {}
transform = Transform.from_params(Params(config_dict))
assert transform.apply(1) == 7
config_dict["x"] = 4
with pytest.raises(ConfigurationError):
Transform.from_params(Params(config_dict))
config_dict["type"] = "mult_by_x"
transform = Transform.from_params(Params(config_dict))
assert transform.apply(4) == 16
config_dict["type"] = "mult_by_x_add_y"
config_dict["x"] = 4
with pytest.raises(ConfigurationError):
Transform.from_params(Params(config_dict))
config_dict["type"] = "mult_by_x_add_y"
config_dict["x"] = 4
config_dict["y"] = 2
transform = Transform.from_params(Params(config_dict))
assert transform.apply(4) == 18
config_dict["type"] = "mult_by_x_add_y"
config_dict["x"] = 4
config_dict["y"] = "test"
with pytest.raises(TypeError):
Transform.from_params(Params(config_dict))
def test_encoders_from_params(self):
resnet_encoders = ["resnet50", "resnet34", "resnet18"]
available_encoders = Encoder.get_available()
for encoder in resnet_encoders:
assert encoder in available_encoders
params = Params({"type": "resnet50", "blocks": 3})
encoder = Encoder.from_params(params)
assert encoder.blocks == 3
assert len(encoder.output_dims) == 5
assert encoder.layer4[-1].conv1.in_channels == 2048
x = torch.zeros((2, 3, 128, 128)).normal_()
encoder.forward(x)
params = Params({"type": "resnet34", "blocks": 2})
encoder = Encoder.from_params(params)
assert encoder.blocks == 2
assert len(encoder.output_dims) == 4
assert len(encoder.layer4) == 3
assert encoder.layer4[-1].conv1.in_channels == 512
encoder.forward(x)
params = Params({"type": "resnet18"})
encoder = Encoder.from_params(params)
assert len(encoder.layer4) == 2
assert encoder.layer4[-1].conv1.in_channels == 512
encoder.forward(x)
params = Params({
"type": "resnet50",
"blocks": 3,
"pretrained": False,
"replace_stride_with_dilation": [False, True, True],
"normalization": {
"type": "identity"
},
})
encoder = Encoder.from_params(params)
assert isinstance(encoder.layer4[0].bn1, torch.nn.Identity)
assert encoder.layer4[1].conv2.dilation == (4, 4)
encoder.forward(x)
def test_make_param_groups(self):
model = Model.from_params(
Params(
{
"encoder": {"type": "resnet50"},
"decoder": {
"type": "unet",
"decoder_channels": [512, 256, 128, 64, 32],
},
"num_classes": 4,
}
)
)
named_parameters = [x for x in model.named_parameters()]
lr_decoder = 0.01
lr_logits = 0.1
lr_encoder = 0.0001
params = {
"param_groups": {
"encoder": {"params": {"lr": lr_encoder}},
"decoder_logits": {
"regexes": ["decoder.logits.*.weight", "decoder.logits.*.bias"],
"params": {"lr": lr_logits},
},
},
"lr": lr_decoder,
}
opt = Optimizer.from_params(Params(params), model_params=named_parameters)
assert opt.state_dict()["param_groups"][0]["lr"] == lr_encoder
assert len(opt.state_dict()["param_groups"][0]["params"]) == 161
assert opt.state_dict()["param_groups"][2]["lr"] == lr_decoder
assert len(opt.state_dict()["param_groups"][2]["params"]) == 20
assert opt.state_dict()["param_groups"][1]["lr"] == lr_logits
assert len(opt.state_dict()["param_groups"][1]["params"]) == 2
def test_encoders_from_params(self):
mobilenet_encoders = ["mobilenetv2"]
available_encoders = Encoder.get_available()
for encoder in mobilenet_encoders:
assert encoder in available_encoders
params = Params({"type": "mobilenetv2", "blocks": 3})
encoder = Encoder.from_params(params)
assert encoder.blocks == 3
assert len(encoder.output_dims) == 5
assert encoder.features[-1][0].out_channels == 1280
x = torch.zeros((2, 3, 256, 256)).normal_()
encoder.forward(x)
params = Params({
"type": "mobilenetv2",
"blocks": 2,
"pretrained": False
})
encoder = Encoder.from_params(params)
assert encoder.blocks == 2
assert len(encoder.output_dims) == 4
assert encoder.features[-1][0].out_channels == 1280
encoder.forward(x)
def test_simple_nesting(self):
config_dict = {
"type": "from_single_transform",
"data_path": 10,
"transform": {
"type": "mult_by_x",
"x": 4
},
}
dataset = Dataset.from_params(Params(config_dict))
assert dataset.data_path == 10
assert len(dataset.transforms) == 1
assert dataset.transforms[0].x == 4
def test_reduce_lr_on_plateau(self):
named_parameters = [("x", torch.nn.Parameter())]
lr = 1
opt = Optimizer.from_params(Params({
"type": "sgd",
"lr": lr
}),
model_params=named_parameters)
gamma = 0.01
patience = 2
sched_params = {
"type": "reduce_on_plateau",
"mode": "max",
"factor": gamma,
"patience": patience,
"min_lr": 1e-3,
}
sched = Scheduler.from_params(Params(sched_params), optimizer=opt)
assert np.allclose(sched.get_last_lr()[0], lr)
sched.step(0.5)
assert np.allclose(sched.get_last_lr()[0], lr)
sched.step(0.4)
assert np.allclose(sched.get_last_lr()[0], lr)
sched.step(0.3)
assert np.allclose(sched.get_last_lr()[0], lr)
sched.step(0.3)
assert np.allclose(sched.get_last_lr()[0], lr * gamma)
# test min_lr
sched.step(0.3)
assert np.allclose(sched.get_last_lr()[0], lr * gamma)
sched.step(0.4)
assert np.allclose(sched.get_last_lr()[0], lr * gamma)
sched.step(0.3)
assert np.allclose(sched.get_last_lr()[0], max(1e-3, lr * gamma**2))
def test_encoders_from_params(self):
blocks = 2
num_channels = 32
n_classes = 2
params = Params({
"encoder": {
"type": "resnet50",
"blocks": blocks,
"pretrained": False,
"replace_stride_with_dilation": [False, True, True],
"normalization": {
"type": "identity"
},
},
"decoder": {
"type": "pan",
"decoder_channels_size": num_channels,
"normalization": {
"type": "batch_renorm_2d"
},
"activation": {
"type": "leaky_relu",
"inplace": True
},
"gau_activation": {
"type": "swish"
},
"upscale_mode": "nearest",
},
"loss": {
"type": "dice"
},
"num_classes": n_classes,
})
x = torch.zeros((2, 3, 128, 128)).normal_()
y = (torch.zeros(
(2, n_classes, 128, 128)).normal_() > 0.5).to(torch.float)
model = Model.from_params(params)
res = model.forward(x, y)
assert res["loss"] > 0
assert res["logits"].shape[1] == n_classes
assert isinstance(model.encoder, ResNetEncoder)
assert isinstance(model.decoder, PanDecoder)
def test_lazy(self):
test_string = "this is a test"
extra_string = "extra string"
class ConstructedObject(FromParams):
def __init__(self, string: str, extra: str):
self.string = string
self.extra = extra
class Testing(FromParams):
def __init__(self, lazy_object: Lazy[ConstructedObject]):
first_time = lazy_object.construct(extra=extra_string)
second_time = lazy_object.construct(extra=extra_string)
assert first_time.string == test_string
assert first_time.extra == extra_string
assert second_time.string == test_string
assert second_time.extra == extra_string
Testing.from_params(Params({"lazy_object": {"string": test_string}}))
def test_list_with_strings(self):
params = {
"type":
"from_transforms_list",
"data_path":
"./",
"transforms": [
"mult_by_2_add_5",
{
"type": "mult_by_x",
"x": 5
},
"mult_by_2",
],
}
dataset = Dataset.from_params(Params(params))
assert len(dataset.transforms) == 3
assert isinstance(dataset.transforms[0], MultiplyByXAddY)
assert dataset.transforms[0].x == 2 and dataset.transforms[0].y == 5
assert isinstance(dataset.transforms[1], MultiplyByX)
assert dataset.transforms[1].x == 5
assert isinstance(dataset.transforms[2], MultiplyByX)
assert dataset.transforms[2].x == 2
def test_none_default_args(self):
params = {"type": "hardtanh", "min_val": "a"}
with pytest.raises(TypeError):
Activation.from_params(Params(params))
def test_complex_nesting(self):
base_config_dict = {
"data_path": "./",
"transforms": {
"t1": {
"type": "mult_by_x",
"x": 4
},
"t2": {
"type": "mult_by_2_add_5"
},
},
}
config_dict = deepcopy(base_config_dict)
dataset = Dataset.from_params(Params(config_dict))
assert dataset.data_path == "./"
assert len(dataset.transforms) == 2
assert dataset.transforms["t2"].y == 5
# We now test with a list of transforms
# Default dataset expects to be mapping
with pytest.raises(TypeError):
config_dict = deepcopy(base_config_dict)
config_dict["transforms"] = list(
config_dict["transforms"].values())
Dataset.from_params(Params(config_dict))
config_dict = deepcopy(base_config_dict)
config_dict["transforms"] = list(config_dict["transforms"].values())
config_dict["type"] = "from_transforms_list"
dataset = Dataset.from_params(Params(config_dict))
assert dataset.data_path == "./"
assert len(dataset.transforms) == 2
assert dataset.transforms[1].y == 5
with pytest.raises(TypeError):
config_dict = deepcopy(base_config_dict)
config_dict["type"] = "from_transforms_list"
Dataset.from_params(Params(config_dict))
# With tuples
# Default dataset expects to be mapping
with pytest.raises(TypeError):
config_dict = deepcopy(base_config_dict)
config_dict["transforms"] = tuple(
config_dict["transforms"].values())
Dataset.from_params(Params(config_dict))
config_dict = deepcopy(base_config_dict)
config_dict["transforms"] = tuple(config_dict["transforms"].values())
config_dict["type"] = "from_transforms_tuple"
dataset = Dataset.from_params(Params(config_dict))
assert dataset.data_path == "./"
assert len(dataset.transforms) == 2
assert dataset.transforms[1].y == 5
with pytest.raises(TypeError):
config_dict = deepcopy(base_config_dict)
config_dict["type"] = "from_transforms_tuple"
Dataset.from_params(Params(config_dict))
# With set
# Default dataset expects to be mapping
with pytest.raises(TypeError):
config_dict = deepcopy(base_config_dict)
config_dict["transforms"] = set(config_dict["transforms"].values())
Dataset.from_params(Params(config_dict))
config_dict = deepcopy(base_config_dict)
config_dict["type"] = "from_set_param"
config_dict["transform"] = list(config_dict["transforms"].values())[0]
del config_dict["transforms"]
config_dict["set_arg"] = [{"x": 1}, {"x": 2}]
dataset = Dataset.from_params(Params(config_dict))
assert dataset.data_path == "./"
assert len(dataset.transforms) == 1
assert dataset.transforms[0].x == 4
assert len(dataset.set_arg) == 2
with pytest.raises(TypeError):
config_dict = deepcopy(base_config_dict)
config_dict["type"] = "from_set_param"
config_dict["transform"] = list(
config_dict["transforms"].values())[0]
del config_dict["transforms"]
config_dict["set_arg"] = {"k1": {"x": 1}, "k2": {"x": 2}}
Dataset.from_params(Params(config_dict))
def test_write_read_from_file(self):
config_dict = {"test": "hello"}
params = Params(config_dict)
assert params.as_dict() == config_dict
write_path = self.TEMPORARY_DIR / "dummy_config.json"
params.to_file(str(write_path))
params2 = Params.from_file(str(write_path))
assert params.as_dict() == params2.as_dict()
assert params.pop("test") == "hello"
assert params.pop("test2", "none") == "none"
with pytest.raises(ConfigurationError):
params.pop("test")
def test_encoders_from_params(self):
available_decoders = Decoder.get_available()
assert "pan" in available_decoders
channels = [3, 6, 12, 24, 48]
num_channels = 16
xs = [
torch.zeros(4, c, 100 // (i + 1), 100 // (i + 1)).normal_()
for i, c in enumerate(channels)
]
n_classes = 8
params = Params(
{
"type": "pan",
"encoder_channels": channels,
"decoder_channels_size": num_channels,
"num_classes": n_classes,
}
)
decoder = Decoder.from_params(params)
assert decoder.fpa.pooling_branch[1][0].in_channels == channels[-1]
assert decoder.fpa.pooling_branch[1][0].out_channels == num_channels
assert decoder.gau3.process_high[1][0].in_channels == num_channels
assert decoder.gau3.process_high[1][0].out_channels == num_channels
assert decoder.gau3.process_low[0].in_channels == channels[-4]
assert decoder.gau3.process_low[0].out_channels == num_channels
assert decoder.logits[0].in_channels == num_channels
assert decoder.logits[0].out_channels == n_classes
decoder.forward(*xs)
params = Params(
{
"type": "pan",
"encoder_channels": channels,
"decoder_channels_size": num_channels,
}
)
with pytest.raises(ConfigurationError):
Decoder.from_params(params)
params = Params(
{
"type": "pan",
"encoder_channels": channels,
"decoder_channels_size": num_channels,
"num_classes": n_classes,
"normalization": {"type": "batch_renorm_2d"},
"activation": {"type": "leaky_relu", "inplace": True},
"gau_activation": {"type": "swish"},
"upscale_mode": "nearest",
}
)
decoder = Decoder.from_params(params)
assert isinstance(decoder.gau3.process_low[2], BatchRenorm2d)
assert isinstance(decoder.gau3.process_low[1], torch.nn.LeakyReLU)
assert decoder.gau3.process_high[1][1]._get_name() == "Swish"
decoder.forward(*xs)
def test_steps_scheduler(self):
named_parameters = [("x", torch.nn.Parameter())]
lr = 10
opt = Optimizer.from_params(Params({
"type": "sgd",
"lr": lr
}),
model_params=named_parameters)
step_size = 10
gamma = 0.1
sched_params = {"type": "step", "step_size": step_size, "gamma": gamma}
sched = Scheduler.from_params(Params(sched_params), optimizer=opt)
assert sched.get_last_lr()[0] == lr
for _ in range(step_size - 1):
opt.step()
sched.step()
assert sched.get_last_lr()[0] == lr
opt.step()
sched.step()
assert np.allclose(sched.get_last_lr()[0], lr * gamma)
lr = 10
opt = Optimizer.from_params(Params({
"type": "sgd",
"lr": lr
}),
model_params=named_parameters)
gamma = 0.1
sched_params = {
"type": "multi_step",
"milestones": [5, 20],
"gamma": gamma
}
sched = Scheduler.from_params(Params(sched_params), optimizer=opt)
assert sched.get_last_lr()[0] == lr
for _ in range(4):
opt.step()
sched.step()
assert sched.get_last_lr()[0] == lr
opt.step()
sched.step()
assert np.allclose(sched.get_last_lr()[0], lr * gamma)
for _ in range(14):
sched.step()
opt.step()
assert np.allclose(sched.get_last_lr()[0], lr * gamma)
opt.step()
sched.step()
assert np.allclose(sched.get_last_lr()[0], lr * gamma**2)
lr = 10
opt = Optimizer.from_params(Params({
"type": "sgd",
"lr": lr
}),
model_params=named_parameters)
gamma = 0.1
sched_params = {"type": "exponential", "gamma": gamma}
sched = Scheduler.from_params(Params(sched_params), optimizer=opt)
assert sched.get_last_lr()[0] == lr
for i in range(50):
opt.step()
sched.step()
assert np.allclose(sched.get_last_lr()[0], lr * gamma**(i + 1))
def test_pop_nested_param(self):
config_dict = {"model": {"type": "test", "other_param": 1}}
params = Params(config_dict)
assert isinstance(params.pop("model"), Params)