This is an unofficial PyTorch implementation by Ignacio Oguiza of - oguiza@gmail.com based on Temporal Convolutional Network (Bai, 2018).

References:

  • Bai, S., Kolter, J. Z., & Koltun, V. (2018). An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv preprint arXiv:1803.01271.
  • Official TCN PyTorch implementation: https://github.com/locuslab/TCN

class TemporalBlock[source]

TemporalBlock(ni, nf, ks, stride, dilation, padding, dropout=0.0) :: Module

Same as nn.Module, but no need for subclasses to call super().__init__

TemporalConvNet[source]

TemporalConvNet(c_in, layers, ks=2, dropout=0.0)

class TCN[source]

TCN(c_in, c_out, layers=[25, 25, 25, 25, 25, 25, 25, 25], ks=7, conv_dropout=0.0, fc_dropout=0.0) :: Module

Same as nn.Module, but no need for subclasses to call super().__init__

bs = 16
nvars = 3
seq_len = 128
c_out = 2
xb = torch.rand(bs, nvars, seq_len)
model = TCN(nvars, c_out, fc_dropout=.5)
test_eq(model(xb).shape, (bs, c_out))
model = TCN(nvars, c_out, conv_dropout=.2)
test_eq(model(xb).shape, (bs, c_out))
model = TCN(nvars, c_out)
test_eq(model(xb).shape, (bs, c_out))
model

TCN(
  (tcn): Sequential(
    (0): TemporalBlock(
      (conv1): Conv1d(3, 25, kernel_size=(7,), stride=(1,), padding=(6,))
      (chomp1): Chomp1d()
      (relu1): ReLU()
      (dropout1): Dropout(p=0.0, inplace=False)
      (conv2): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(6,))
      (chomp2): Chomp1d()
      (relu2): ReLU()
      (dropout2): Dropout(p=0.0, inplace=False)
      (net): Sequential(
        (0): Conv1d(3, 25, kernel_size=(7,), stride=(1,), padding=(6,))
        (1): Chomp1d()
        (2): ReLU()
        (3): Dropout(p=0.0, inplace=False)
        (4): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(6,))
        (5): Chomp1d()
        (6): ReLU()
        (7): Dropout(p=0.0, inplace=False)
      )
      (downsample): Conv1d(3, 25, kernel_size=(1,), stride=(1,))
      (relu): ReLU()
    )
    (1): TemporalBlock(
      (conv1): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(12,), dilation=(2,))
      (chomp1): Chomp1d()
      (relu1): ReLU()
      (dropout1): Dropout(p=0.0, inplace=False)
      (conv2): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(12,), dilation=(2,))
      (chomp2): Chomp1d()
      (relu2): ReLU()
      (dropout2): Dropout(p=0.0, inplace=False)
      (net): Sequential(
        (0): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(12,), dilation=(2,))
        (1): Chomp1d()
        (2): ReLU()
        (3): Dropout(p=0.0, inplace=False)
        (4): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(12,), dilation=(2,))
        (5): Chomp1d()
        (6): ReLU()
        (7): Dropout(p=0.0, inplace=False)
      )
      (relu): ReLU()
    )
    (2): TemporalBlock(
      (conv1): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(24,), dilation=(4,))
      (chomp1): Chomp1d()
      (relu1): ReLU()
      (dropout1): Dropout(p=0.0, inplace=False)
      (conv2): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(24,), dilation=(4,))
      (chomp2): Chomp1d()
      (relu2): ReLU()
      (dropout2): Dropout(p=0.0, inplace=False)
      (net): Sequential(
        (0): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(24,), dilation=(4,))
        (1): Chomp1d()
        (2): ReLU()
        (3): Dropout(p=0.0, inplace=False)
        (4): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(24,), dilation=(4,))
        (5): Chomp1d()
        (6): ReLU()
        (7): Dropout(p=0.0, inplace=False)
      )
      (relu): ReLU()
    )
    (3): TemporalBlock(
      (conv1): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(48,), dilation=(8,))
      (chomp1): Chomp1d()
      (relu1): ReLU()
      (dropout1): Dropout(p=0.0, inplace=False)
      (conv2): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(48,), dilation=(8,))
      (chomp2): Chomp1d()
      (relu2): ReLU()
      (dropout2): Dropout(p=0.0, inplace=False)
      (net): Sequential(
        (0): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(48,), dilation=(8,))
        (1): Chomp1d()
        (2): ReLU()
        (3): Dropout(p=0.0, inplace=False)
        (4): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(48,), dilation=(8,))
        (5): Chomp1d()
        (6): ReLU()
        (7): Dropout(p=0.0, inplace=False)
      )
      (relu): ReLU()
    )
    (4): TemporalBlock(
      (conv1): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(96,), dilation=(16,))
      (chomp1): Chomp1d()
      (relu1): ReLU()
      (dropout1): Dropout(p=0.0, inplace=False)
      (conv2): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(96,), dilation=(16,))
      (chomp2): Chomp1d()
      (relu2): ReLU()
      (dropout2): Dropout(p=0.0, inplace=False)
      (net): Sequential(
        (0): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(96,), dilation=(16,))
        (1): Chomp1d()
        (2): ReLU()
        (3): Dropout(p=0.0, inplace=False)
        (4): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(96,), dilation=(16,))
        (5): Chomp1d()
        (6): ReLU()
        (7): Dropout(p=0.0, inplace=False)
      )
      (relu): ReLU()
    )
    (5): TemporalBlock(
      (conv1): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(192,), dilation=(32,))
      (chomp1): Chomp1d()
      (relu1): ReLU()
      (dropout1): Dropout(p=0.0, inplace=False)
      (conv2): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(192,), dilation=(32,))
      (chomp2): Chomp1d()
      (relu2): ReLU()
      (dropout2): Dropout(p=0.0, inplace=False)
      (net): Sequential(
        (0): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(192,), dilation=(32,))
        (1): Chomp1d()
        (2): ReLU()
        (3): Dropout(p=0.0, inplace=False)
        (4): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(192,), dilation=(32,))
        (5): Chomp1d()
        (6): ReLU()
        (7): Dropout(p=0.0, inplace=False)
      )
      (relu): ReLU()
    )
    (6): TemporalBlock(
      (conv1): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(384,), dilation=(64,))
      (chomp1): Chomp1d()
      (relu1): ReLU()
      (dropout1): Dropout(p=0.0, inplace=False)
      (conv2): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(384,), dilation=(64,))
      (chomp2): Chomp1d()
      (relu2): ReLU()
      (dropout2): Dropout(p=0.0, inplace=False)
      (net): Sequential(
        (0): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(384,), dilation=(64,))
        (1): Chomp1d()
        (2): ReLU()
        (3): Dropout(p=0.0, inplace=False)
        (4): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(384,), dilation=(64,))
        (5): Chomp1d()
        (6): ReLU()
        (7): Dropout(p=0.0, inplace=False)
      )
      (relu): ReLU()
    )
    (7): TemporalBlock(
      (conv1): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(768,), dilation=(128,))
      (chomp1): Chomp1d()
      (relu1): ReLU()
      (dropout1): Dropout(p=0.0, inplace=False)
      (conv2): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(768,), dilation=(128,))
      (chomp2): Chomp1d()
      (relu2): ReLU()
      (dropout2): Dropout(p=0.0, inplace=False)
      (net): Sequential(
        (0): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(768,), dilation=(128,))
        (1): Chomp1d()
        (2): ReLU()
        (3): Dropout(p=0.0, inplace=False)
        (4): Conv1d(25, 25, kernel_size=(7,), stride=(1,), padding=(768,), dilation=(128,))
        (5): Chomp1d()
        (6): ReLU()
        (7): Dropout(p=0.0, inplace=False)
      )
      (relu): ReLU()
    )
  )
  (gap): GAP1d(
    (gap): AdaptiveAvgPool1d(output_size=1)
    (flatten): Flatten(full=False)
  )
  (linear): Linear(in_features=25, out_features=2, bias=True)
)