def test_mask_conv_1d_out_lens(data, conv1d: torch.nn.Conv1d) -> None:
"""Ensures out_lens returns correct output sequence lengths."""
# PyTorch conv1d requires input size to be greater than effective kernel
# size. Force this to be True.
effective_ks = conv1d.dilation[0] * (conv1d.kernel_size[0] - 1) + 1
padding = data.draw(st.integers(0, effective_ks + 100))
seq_lens = data.draw(
st.lists(
elements=st.integers(max(0, effective_ks - padding), 256),
min_size=0,
max_size=32,
))
exp = []
for seq_len in seq_lens:
out = conv1d(torch.empty([1, conv1d.in_channels, seq_len + padding]))
exp.append(out.size(2))
act = list(
out_lens(
seq_lens=torch.tensor(seq_lens),
kernel_size=conv1d.kernel_size[0],
stride=conv1d.stride[0],
dilation=conv1d.dilation[0],
padding=padding,
))
assert exp == act
def test_mask_conv_output_size_and_seq_lens(
mask_conv1d_input: Tuple[MaskConv1d, Tuple[torch.Tensor, torch.Tensor]]
) -> None:
"""Ensures the MaskConv1d output size and seq_len values are correct."""
mask_conv1d, inputs = mask_conv1d_input
batch, channels, seq_len = inputs[0].size()
out_tensor, out_seq_lens = mask_conv1d(inputs)
out_batch, out_channels, out_seq_len = out_tensor.size()
assert out_batch == len(out_seq_lens) == batch
assert out_channels == mask_conv1d.out_channels
# test out_seq_len dimension of tensor output and set padding value ready
# for testing out_seq_lens
if mask_conv1d._padding_mode == PaddingMode.NONE:
padding = (0, 0)
# out_lens function returns correct expected length if the tests pass
exp_len = out_lens(
seq_lens=torch.tensor(seq_len),
kernel_size=mask_conv1d.kernel_size[0],
stride=mask_conv1d.stride[0],
dilation=mask_conv1d.dilation[0],
padding=0,
).item()
assert out_seq_len == exp_len
elif mask_conv1d._padding_mode == PaddingMode.SAME:
padding = pad_same(
length=seq_len,
kernel_size=mask_conv1d.kernel_size[0],
stride=mask_conv1d.stride[0],
dilation=mask_conv1d.dilation[0],
)
# by definition of SAME padding
assert out_seq_len == math.ceil(float(seq_len) / mask_conv1d.stride[0])
else:
raise ValueError(f"unknown PaddingMode {mask_conv1d._padding_mode}")
# test out_seq_lens
exp_out_seq_lens = out_lens(
seq_lens=inputs[1],
kernel_size=mask_conv1d.kernel_size[0],
stride=mask_conv1d.stride[0],
dilation=mask_conv1d.dilation[0],
padding=sum(padding),
)
assert torch.all(out_seq_lens == exp_out_seq_lens)
def _build_cnn(conv_blocks, input_features: int, input_channels: int):
act_dims = 4 # batch, channels, features, seq_len
layers = []
for conv_block in conv_blocks:
convnd_str = conv_block.WhichOneof("convnd")
if convnd_str == "conv1d":
if act_dims == 4:
layers.append(Conv2dTo1d())
act_dims = 3
input_channels *= input_features
input_features = 1
conv_cfg = conv_block.conv1d
if conv_cfg.padding_mode == conv_layer_pb2.PADDING_MODE.NONE:
padding_mode = PaddingMode.NONE
elif conv_cfg.padding_mode == conv_layer_pb2.PADDING_MODE.SAME:
padding_mode = PaddingMode.SAME
layers.append(
MaskConv1d(
in_channels=input_channels,
out_channels=conv_cfg.output_channels,
kernel_size=conv_cfg.kernel_time,
stride=conv_cfg.stride_time,
padding_mode=padding_mode,
bias=conv_cfg.bias,
))
input_channels = conv_cfg.output_channels
elif convnd_str == "conv2d":
if act_dims == 3:
layers.append(Conv1dTo2d())
act_dims = 4
input_features = input_channels
input_channels = 1
conv_cfg = conv_block.conv2d
if conv_cfg.padding_mode == conv_layer_pb2.PADDING_MODE.NONE:
padding_mode = PaddingMode.NONE
input_features = out_lens(
torch.tensor([input_features]),
kernel_size=conv_cfg.kernel_feature,
stride=conv_cfg.stride_feature,
dilation=1,
padding=0,
).item()
elif conv_cfg.padding_mode == conv_layer_pb2.PADDING_MODE.SAME:
padding_mode = PaddingMode.SAME
input_features = math.ceil(input_features /
conv_cfg.stride_feature)
layers.append(
MaskConv2d(
in_channels=input_channels,
out_channels=conv_cfg.output_channels,
kernel_size=[
conv_cfg.kernel_feature,
conv_cfg.kernel_time,
],
stride=[conv_cfg.stride_feature, conv_cfg.stride_time],
padding_mode=padding_mode,
bias=conv_cfg.bias,
))
input_channels = conv_cfg.output_channels
layers.append(
SeqLenWrapper(build_activation(conv_block.activation),
torch.nn.Identity()))
if act_dims == 3:
layers.append(Conv1dTo2d())
input_features = input_channels
input_channels = 1
return torch.nn.Sequential(*layers), input_features * input_channels
