Useful additional layers for PyTorch.
torchmore library is a small library of layers and utilities
for writing PyTorch models for image recognition, OCR, and other applications.
flex library performs simple size inference. It does so by wrapping up individual layers in a wrapper that instantiates the layer only when dimensional data is available. The wrappers can be removed later and the model turned into one with only completely standard modules. That looks like this:
from torch import nn from torchmore import layers, flex noutput = 10 model = nn.Sequential( layers.Input("BDHW"), flex.Conv2d(100), flex.BatchNorm(), nn.ReLU(), flex.Conv2d(100), flex.BatchNorm(), nn.ReLU(), layers.Reshape([1, [2, 3, 4]]), flex.Full(100), flex.BatchNorm(), nn.ReLU(), flex.Full(noutput) ) flex.shape_inference(model, (1, 1, 28, 28))
flex library provides wrappers for the following layers right now:
You can use
Flex directly. The following two layers are identical:
layer1 = flex.Conv2d(100) layer2 = flex.Flex(lambda x: nn.Conv2d(x.size(1), 100))
That is, you can easily turn any layer into a
Flex layer that way even if it isn't in the library.
Input layer is a handy little layer that reorders input dimensions, checks size ranges and value ranges, and automatically transfers data to the current device on which the model runs.
For example, consider the following
layers.Input("BHWD", "BDHW", range=(0, 1), sizes=[None, 1, None, None]),
- the input is in "BHWD" order and will get reordered to "BDHW"
- input values must be in the interval $[0, 1]$
- input tensors must have $D=1$
- input tensors are transferred to the same device as weights for the model
Note that if the input tensor has a
.order attribute, that will be used to reorder the input dimensions into the desired dimensions. This allows the model to accept inputs in multiple orders. Consider
model = nn.Sequential( layers.Input("BHWD", "BDHW", range=(0, 1), sizes=[None, 1, None, None]), ... ) a = torch.rand((1, 100, 150, 1)) b = a.permute(0, 3, 1, 2) b.order = "BDHW" assert model(a) == model(b)
Reorder layer reorders axes just like
Tensor.permute does, but it does so in a way that documents better what is going on. Consider the following code fragment:
layers.Reorder("BDL", "LBD"), flex.LSTM(100, bidirectional=True), layers.Reorder("LBD", "BDL"), flex.Conv1d(noutput, 1), layers.Reorder("BDL", "BLD")
The letters themselves are arbitrary, but common choices are "BDLHW". This is likely clearer than a sequence of permutations.
For module-based networks, it's convenient to add functions. The
Fun layer permits that, as in:
layers.Fun("lambda x: x.permute(2, 0, 1)")
Note that since functions are specified as strings, this can be pickled.
layers.LSTM: a trivial LSTM layer that simply dicards the state output
layers.BDL_LSTM: an LSTM variant that is a drop-in replacement for a
layers.BDHW_LSTM: an MDLSTM variant that is a drop-in replacement for a
layers.BDHW_LSTM_to_BDH: a rowwise LSTM, reducing dimension by 1
These may be occasionally useful:
layers.Info(info="", every=1000000): prints info about the activations
layers.CheckSizes(...): checks the sizes of tensors propagated through
layers.CheckRange(...): checks the ranges of values
layers.Permute(...): axis permutation (like x.permute)
layers.Reshape(...): tensor reshaping, with the option of combining axes
layers.View(...): equivalent of x.view
layers.Parallel: run two modules in parallel and stack the results
layers.SimplePooling2d: wrapped up max pooling/unpooling
layers.AcrossPooling2d: wrapped up max pooling/unpooling with convolution
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
|Filename, size||File type||Python version||Upload date||Hashes|
|Filename, size torchmore-0.1.0-py3-none-any.whl (13.1 kB)||File type Wheel||Python version py3||Upload date||Hashes View|
|Filename, size torchmore-0.1.0.tar.gz (13.4 kB)||File type Source||Python version None||Upload date||Hashes View|
Hashes for torchmore-0.1.0-py3-none-any.whl