Miscellaneous

This contains a set of experiments.


source

InputWrapper

 InputWrapper (arch, c_in, c_out, seq_len, new_c_in=None,
               new_seq_len=None, **kwargs)

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

from tsai.models.TST import *
xb = torch.randn(16, 1, 1000)
model = InputWrapper(TST, 1, 4, 1000, 10, 224)
test_eq(model.to(xb.device)(xb).shape, (16,4))

source

ResidualWrapper

 ResidualWrapper (model)

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


source

RecursiveWrapper

 RecursiveWrapper (model, n_steps, anchored=False)

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

xb = torch.randn(16, 1, 20)
model = RecursiveWrapper(TST(1, 1, 20), 5)
test_eq(model.to(xb.device)(xb).shape, (16, 5))