torchparse: PyTorch cfg Model Parser

Simple (and for now, sequential) PyTorch model parser. Allowes to define a model in a cfg file for easier iteration.

Flow

A more detailed example with a custom module in demo.ipynb.

Features

• Don't have to worry about layer I/O dimensions
• Easily define dimension reshapes between layers
• Repeat block syntax for less typing
• Get a neat nn.ModuleDict back with the desired nn.Sequentials

Contents

• Installation
• Supported modules
• Detailed Usage
• Improvements

Installation

HTTPS

pip install git+https://github.com/ksanjeevan/torchparse.git

SSH

pip install git+ssh://[email protected]/ksanjeevan/torchparse.git

Verify:

>> from torchparse import parse_cfg, get_sample_cfg
>> parse_cfg(get_sample_cfg(), in_shape=(3,100,100))

ModuleDict(
  (convs): Sequential(
    (conv2d_0): Conv2d(3, 32, kernel_size=(3, 3), stride=(1, 1))
...

Supported modules

Implemented layers (nn.Module):

• Linear
• LSTM, GRU, RNN
• (Avg,Max)Pool(1,2)d
• Conv(1,2)d
• BatchNorm(1,2)d
• ReLU, ELU, LeakyReLU, Sigmoid
• Dropout
• Adaptive(Max,Avg)Pool(1,2)d
• ConvTranspose(1,2)d
• Upsample
• Unpooling

Detailed Usage

[moddims]: account for Tensor manipulations

Allows to incorporate in the cfg file any tranpose or reshape that will occur in the forward call, since this will affect the intermmediate shapes.

permute

For example if in forward():

...
# (batch, height, width, channel) -> (batch, channel, height, width)
x = x.permute([0,3,1,2])
...

then in .cfg add:

...
[moddims]
	permute=[2,0,1]
...

Since torchparse doesn't consider batch dimension (should it? maybe... but not for now).

This can also be used when dropping a dimension. e.g. in a many-to-one RNN might do something like:

...
# (batch, time, feature) -> (batch, feature)
x = x[:,-1]
...

then in .cfg add:

...
[moddims]
	permute=[1]
...

(Doesn't care if we choose the last input of the RNN, only that the time dimension is not there anymore, i.e. keep dimension 1).

---

collapse

For example if in forward():

...
# (batch, time, freq, channel) -> (batch, time, freq*channel)
batch, time = x.size()[:2]
x = x.view(batch, time, -1)
...

then in .cfg add:

...
[moddims]
	collapse=[1,2]
...

---

[_module]: sub-module sequential blocks

Even for a sequential model there might be transformations applied in the forward call that aren't defined in the nn.Module (e.g.: example above where the conv_module will be seperatley defined from the recur_module since the foraward call will deal with the reshapes, packing sequences, etc.).

For now only allow shallow submodules (i.e. every .cfg can have any number of named sequential submodules).

torchparse.parse_cfg will return an nn.ModuleDict. If no submodules are explicitly defined, the nn.ModuleDict will only have one key (main) mapping to the defined nn.Sequential.

REPEATx: for repeating blocks of layers

If the model has blocks of layers that repeat with the same paramater values (like in the example above or in example.py), they can be written in the .cfg as:

...
REPEATx3
    [conv2d]
        out_channels=64
        kernel_size=3
        stride=1
        padding=valid
    [batchnorm2d]
    [elu]
    [maxpool2d]
        kernel_size=4
        stride=4
END
...

Improvements

• Non _module cfg handling
• Block repetitions
• Skip connections
• Allow .cfg to include input shape