def test_conv_and_pool_kwargs(self):
inputs = torch.ones([128, 64, 300])
hparams = {
# Conv layers
"num_conv_layers": 2,
"out_channels": 128,
"kernel_size": [[3, 4, 5], 4],
"other_conv_kwargs": [{"padding": 0}, {"padding": 0}],
# Pooling layers
"pooling": "AvgPool1d",
"pool_size": 2,
"pool_stride": 1,
"other_pool_kwargs": {},
# Dense layers
"num_dense_layers": 3,
"out_features": [128, 128, 10],
"dense_activation": "ReLU",
"other_dense_kwargs": None,
# Dropout
"dropout_conv": [0, 1, 2],
"dropout_dense": 2
}
network = Conv1DNetwork(in_channels=inputs.shape[1],
in_features=inputs.shape[2],
hparams=hparams)
# dropout-merge-dropout-(Sequential(Conv, ReLU))-avgpool-dropout-
# flatten-(Sequential(Linear,ReLU))-(Sequential(Linear,ReLU))-dropout
# -linear
self.assertEqual(len(network.layers), 1 + 1 + 1 + 3 + 4 + 1)
self.assertIsInstance(
network.layer_by_name("MergeLayer"), MergeLayer)
for layer in network.layers[1].layers:
self.assertIsInstance(layer, nn.Sequential)
def test_sequence_length(self):
batch_size = 4
channels = 64
max_time = 300
sequence_lengths = torch.LongTensor(batch_size).random_(1, max_time)
inputs = torch.ones([batch_size, channels, max_time])
hparams = {
# Conv layers
"num_conv_layers": 2,
"out_channels": 128,
"kernel_size": [[3, 4, 5], 4],
"other_conv_kwargs": {
"padding": 0
},
# Pooling layers
"pooling": "AvgPool1d",
"pool_size": 2,
"pool_stride": 1,
# Dense layers
"num_dense_layers": 0,
"dense_activation": "ReLU",
"other_dense_kwargs": None,
# Dropout
"dropout_conv": [0, 1, 2],
"dropout_dense": 2
}
network = Conv1DNetwork(in_channels=inputs.shape[1],
in_features=inputs.shape[2],
hparams=hparams)
outputs = network(input=inputs, sequence_length=sequence_lengths)
self.assertEqual(
outputs.shape,
torch.Size([batch_size, network.hparams["out_channels"], 884]))
# channels last
inputs = torch.ones([batch_size, max_time, channels])
network = Conv1DNetwork(in_channels=inputs.shape[2],
in_features=inputs.shape[1],
hparams=hparams)
outputs = network(input=inputs,
sequence_length=sequence_lengths,
data_format="channels_last")
self.assertEqual(
outputs.shape,
torch.Size([batch_size, 884, network.hparams["out_channels"]]))
def default_hparams() -> Dict[str, Any]:
r"""Returns a dictionary of hyperparameters with default values.
The same as :meth:`~texar.torch.modules.Conv1DNetwork.default_hparams`
of :class:`~texar.torch.modules.Conv1DNetwork`, except that the default
name is ``"conv_encoder"``.
"""
hparams = Conv1DNetwork.default_hparams()
hparams['name'] = 'conv_encoder'
return hparams
def test_feedforward(self):
"""Tests feed forward.
"""
inputs_1 = torch.ones([128, 32, 300])
network_1 = Conv1DNetwork(in_channels=inputs_1.shape[1],
in_features=inputs_1.shape[2])
# dense layers are not constructed yet
self.assertEqual(len(network_1.layers), 4)
self.assertIsInstance(
network_1.layer_by_name("MergeLayer"), MergeLayer)
for layer in network_1.layers[0].layers:
self.assertIsInstance(layer, nn.Sequential)
outputs_1 = network_1(inputs_1)
self.assertEqual(outputs_1.shape, torch.Size([128, 256]))
self.assertEqual(outputs_1.size(-1), network_1.output_size)
inputs_2 = torch.ones([128, 64, 300])
hparams = {
# Conv layers
"num_conv_layers": 2,
"out_channels": 128,
"kernel_size": [[3, 4, 5], 4],
"other_conv_kwargs": {"padding": 0},
# Pooling layers
"pooling": "AvgPool1d",
"pool_size": 2,
"pool_stride": 1,
# Dense layers
"num_dense_layers": 3,
"out_features": [128, 128, 10],
"dense_activation": "ReLU",
"other_dense_kwargs": None,
# Dropout
"dropout_conv": [0, 1, 2],
"dropout_dense": 2
}
network_2 = Conv1DNetwork(in_channels=inputs_2.shape[1],
in_features=inputs_2.shape[2],
hparams=hparams)
# dropout-merge-dropout-(Sequential(Conv, ReLU))-avgpool-dropout-
# flatten-(Sequential(Linear,ReLU))-(Sequential(Linear,ReLU))-dropout
# -linear
self.assertEqual(len(network_2.layers), 1 + 1 + 1 + 3 + 4 + 1)
self.assertIsInstance(
network_2.layer_by_name("MergeLayer"), MergeLayer)
for layer in network_2.layers[1].layers:
self.assertIsInstance(layer, nn.Sequential)
outputs_2 = network_2(inputs_2)
self.assertEqual(outputs_2.shape, torch.Size([128, 10]))
self.assertEqual(outputs_2.size(-1), network_2.output_size)
# test whether concatenation happens along channel dim when feature dim
# has been reduced
hparams = {
# Conv layers
"num_conv_layers": 1,
"out_channels": 128,
"kernel_size": [3, 4, 5],
"other_conv_kwargs": {"padding": 0},
# Pooling layers
"pooling": "AvgPool1d",
"pool_size": None,
"pool_stride": 1,
# Dense layers
"num_dense_layers": 0,
"out_features": [],
"dense_activation": "ReLU",
"other_dense_kwargs": None,
# Dropout
"dropout_conv": [],
"dropout_dense": []
}
inputs_3 = torch.ones([128, 64, 300])
network_3 = Conv1DNetwork(in_channels=inputs_3.shape[1],
in_features=inputs_3.shape[2],
hparams=hparams)
outputs_3 = network_3(inputs_3)
num_of_kernels = len(hparams["kernel_size"])
out_channels = hparams["out_channels"]
self.assertEqual(outputs_3.shape,
torch.Size([128, num_of_kernels * out_channels]))
self.assertEqual(outputs_3.size(-1), network_3.output_size)
# test for channels last tensors
inputs_3 = inputs_3.permute(0, 2, 1)
outputs_3 = network_3(inputs_3, data_format="channels_last")
self.assertEqual(outputs_3.shape,
torch.Size([128, num_of_kernels * out_channels]))
self.assertEqual(outputs_3.size(-1), network_3.output_size)