def get_split(
dataset_df,
force_split=False,
split_probabilities=(0.7, 0.1, 0.2),
stratify=None,
random_seed=default_random_seed,
):
if SPLIT in dataset_df and not force_split:
split = dataset_df[SPLIT]
else:
set_random_seed(random_seed)
if stratify is None or stratify not in dataset_df:
split = np.random.choice(
3,
len(dataset_df),
p=split_probabilities,
).astype(np.int8)
else:
split = np.zeros(len(dataset_df))
for val in dataset_df[stratify].unique():
idx_list = (
dataset_df.index[dataset_df[stratify] == val].tolist()
)
val_list = np.random.choice(
3,
len(idx_list),
p=split_probabilities,
).astype(np.int8)
split[idx_list] = val_list
return split
def test_fc_stack(
first_layer_input_size: Optional[int],
layers: Optional[List],
num_layers: Optional[int],
):
set_random_seed(RANDOM_SEED)
fc_stack = FCStack(first_layer_input_size=first_layer_input_size,
layers=layers,
num_layers=num_layers).to(DEVICE)
input_tensor = torch.randn(BATCH_SIZE,
first_layer_input_size,
device=DEVICE)
output_tensor = fc_stack(input_tensor)
assert output_tensor.shape[1:] == fc_stack.output_shape
def test_fc_layer(
input_size: int,
output_size: int,
activation: str,
dropout: float,
):
set_random_seed(RANDOM_SEED) # make repeatable
fc_layer = FCLayer(input_size=input_size,
output_size=output_size,
activation=activation,
dropout=dropout).to(DEVICE)
input_tensor = torch.randn(BATCH_SIZE, input_size, device=DEVICE)
output_tensor = fc_layer(input_tensor)
assert output_tensor.shape[1:] == fc_layer.output_shape