March 24, 2020

1203 words 6 mins read



Shape inference for PyTorch (like in Keras) + new SoTA layers!

repo name szymonmaszke/torchlayers
repo link
language Python
size (curr.) 3455 kB
stars (curr.) 330
created 2019-08-12
license MIT License

torchlayers Logo

Version Docs Tests Coverage Style PyPI Python PyTorch Docker
Version Documentation Tests codecov codebeat badge PyPI Python PyTorch Docker

torchlayers is a library based on PyTorch providing automatic shape and dimensionality inference of torch.nn layers + additional building blocks featured in current SOTA architectures (e.g. Efficient-Net).

Above requires no user intervention (except single call to similarly to the one seen in Keras.

Main functionalities:

  • Shape inference for most of torch.nn module (convolutional, recurrent, transformer, attention and linear layers)
  • Dimensionality inference (e.g. torchlayers.Conv working as torch.nn.Conv1d/2d/3d based on input shape)
  • Shape inference of custom modules (see examples section)
  • Additional Keras-like layers (e.g. torchlayers.Reshape or torchlayers.StandardNormalNoise)
  • Additional SOTA layers mostly from ImageNet competitions (e.g. PolyNet, Squeeze-And-Excitation, StochasticDepth)
  • Useful defaults ("same" padding and default kernel_size=3 for Conv, dropout rates etc.)
  • Zero overhead and torchscript support

Keep in mind this library works almost exactly like PyTorch originally. What that means is you can use Sequential, define your own networks of any complexity using torch.nn.Module, create new layers with shape inference etc.

See below to get some intuition about library.


For full functionality please check torchlayers documentation. Below examples should introduce all necessary concepts you should know.

Basic classifier

All torch.nn modules can be used through torchlayers and each module with input shape will be appropriately modified with it’s input inferable counterpart.

import torchlayers as tl

class Classifier(tl.Module):
    def __init__(self):
        self.conv1 = tl.Conv2d(64, kernel_size=6)
        self.conv2 = tl.Conv2d(128, kernel_size=3)
        self.conv3 = tl.Conv2d(256, kernel_size=3, padding=1)
        # New layer, more on that in the next example
        self.pooling = tl.GlobalMaxPool()
        self.dense = tl.Linear(10)

    def forward(self, x):
        x = torch.relu(self.conv1(x))
        x = torch.relu(self.conv2(x))
        x = torch.relu(self.conv3(x))
        return self.dense(self.pooling(x))

# Pass model and any example inputs afterwards
clf =, torch.randn(1, 3, 32, 32))

Above torchlayers.Linear(out_features=10) is used. It is “equivalent” to original PyTorch’s torch.nn.Linear(in_features=?, out_features=10) where in_features will be inferred from example input input during call.

Same thing happens with torch.nn.Conv2d(in_channels, out_channels, kernel_size, ...) which can be replaced directly by tl.Conv2d(out_channels, kernel_size, ...).

Just remember to pass example input through the network!

Simple image and text classifier in one!

  • We will use single “model” for both tasks. Firstly let’s define it using torch.nn and torchlayers:
import torch
import torchlayers as tl

# torch.nn and torchlayers can be mixed easily
model = torch.nn.Sequential(
    tl.Conv(64),  # specify ONLY out_channels
    torch.nn.ReLU(),  # use torch.nn wherever you wish
    tl.BatchNorm(),  # BatchNormNd inferred from input
    tl.Conv(128),  # Default kernel_size equal to 3
    tl.Conv(256, kernel_size=11),  # "same" padding as default
    tl.GlobalMaxPool(),  # Known from Keras
    tl.Linear(10),  # Output for 10 classes


Above would give you model’s summary like this (notice question marks for not yet inferred values):

  (0): Conv(in_channels=?, out_channels=64, kernel_size=3, stride=1, padding=same, dilation=1, groups=1, bias=True, padding_mode=zeros)
  (1): ReLU()
  (2): BatchNorm(num_features=?, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
  (3): Conv(in_channels=?, out_channels=128, kernel_size=3, stride=1, padding=same, dilation=1, groups=1, bias=True, padding_mode=zeros)
  (4): ReLU()
  (5): Conv(in_channels=?, out_channels=256, kernel_size=11, stride=1, padding=same, dilation=1, groups=1, bias=True, padding_mode=zeros)
  (6): GlobalMaxPool()
  (7): Linear(in_features=?, out_features=10, bias=True)
  • Now you can build/instantiate your model with example input (in this case MNIST-like):
mnist_model =, torch.randn(1, 3, 28, 28))
  • Or if it’s text classification you are after, same model could be built with different input shape (e.g. for text classification using 300 dimensional pretrained embedding):
# [batch, embedding, timesteps], first dimension > 1 for BatchNorm1d to work
text_model =, torch.randn(2, 300, 1))
  • Finally, you can print both models after instantiation, provided below side by-side for readability (notice different dimenstionality, e.g. Conv2d vs Conv1d after
                # TEXT CLASSIFIER                 MNIST CLASSIFIER

                Sequential(                       Sequential(
                  (0): Conv1d(300, 64)              (0): Conv2d(3, 64)
                  (1): ReLU()                       (1): ReLU()
                  (2): BatchNorm1d(64)              (2): BatchNorm2d(64)
                  (3): Conv1d(64, 128)              (3): Conv2d(64, 128)
                  (4): ReLU()                       (4): ReLU()
                  (5): Conv1d(128, 256)             (5): Conv2d(128, 256)
                  (6): GlobalMaxPool()              (6): GlobalMaxPool()
                  (7): Linear(256, 10)              (7): Linear(256, 10)
                )                                 )

As you can see both modules “compiled” into original pytorch layers.

Custom modules with shape inference capabilities

User can define any module and make it shape inferable with torchlayers.infer function:

 # Class defined with in_features
 # It might be a good practice to use _ prefix and Impl as postfix
 # to differentiate from shape inferable version
class _MyLinearImpl(torch.nn.Module):
    def __init__(self, in_features: int, out_features: int):
        self.weight = torch.nn.Parameter(torch.randn(out_features, in_features))
        self.bias = torch.nn.Parameter(torch.randn(out_features))

    def forward(self, inputs):
        return torch.nn.functional.linear(inputs, self.weight, self.bias)

MyLinear = tl.infer(_MyLinearImpl)

# Build and use just like any other layer in this library
layer, torch.randn(1, 64))
layer(torch.randn(1, 64))

By default inputs.shape[1] will be used as in_features value during initial forward pass. If you wish to use different index (e.g. to infer using inputs.shape[3]) use MyLayer = tl.infer(_MyLayerImpl, index=3) as a decorator.

Autoencoder with inverted residual bottleneck and pixel shuffle

Please check code comments and documentation if needed. If you are unsure what autoencoder is you could see this example blog post.

Below is a convolutional denoising autoencoder example for ImageNet-like images. Think of it like a demonstration of capabilities of different layers and building blocks provided by torchlayers.

# Input - 3 x 256 x 256 for ImageNet reconstruction
class AutoEncoder(torch.nn.Module):
    def __init__(self):
        self.encoder = tl.Sequential(
            tl.StandardNormalNoise(),  # Apply noise to input images
            tl.Conv(64, kernel_size=7),
            tl.activations.Swish(),  # Direct access to module .activations
            tl.AvgPool(),  # shape 64 x 128 x 128, kernel_size=2 by default
            tl.HardSwish(),  # Access simply through tl
            tl.SeparableConv(128),  # Up number of channels to 128
            tl.InvertedResidualBottleneck(),  # Default with squeeze excitation
            tl.AvgPool(),  # shape 128 x 64 x 64, kernel_size=2 by default
            tl.DepthwiseConv(256),  # DepthwiseConv easier to use
            # Pass input thrice through the same weights like in PolyNet
            tl.Poly(tl.InvertedResidualBottleneck(), order=3),
            tl.ReLU(),  # all torch.nn can be accessed via tl
            tl.MaxPool(),  # shape 256 x 32 x 32
            tl.Fire(out_channels=512),  # shape 512 x 32 x 32
            tl.MaxPool(),  # shape 512 x 16 x 16
            # Randomly switch off the last two layers with 0.5 probability
            tl.AvgPool(),  # shape 512 x 8 x 8

        # This one is more "standard"
        self.decoder = tl.Sequential(
            tl.Poly(tl.InvertedResidualBottleneck(), order=2),
            # Has ICNR initialization by default after calling `build`
            tl.ConvPixelShuffle(out_channels=512, upscale_factor=2),
            # Shape 512 x 16 x 16 after PixelShuffle
            tl.Poly(tl.InvertedResidualBottleneck(), order=3),
            tl.ConvPixelShuffle(out_channels=256, upscale_factor=2),
            # Shape 256 x 32 x 32
            tl.Poly(tl.InvertedResidualBottleneck(), order=3),
            tl.ConvPixelShuffle(out_channels=128, upscale_factor=2),
            # Shape 128 x 64 x 64
            tl.Poly(tl.InvertedResidualBottleneck(), order=4),
            tl.ConvPixelShuffle(out_channels=64, upscale_factor=2),
            # Shape 64 x 128 x 128
            tl.Dropout(),  # Defaults to 0.5 and Dropout2d for images
            tl.ConvPixelShuffle(out_channels=32, upscale_factor=2),
            # Shape 32 x 256 x 256
            # Shape 3 x 256 x 256

    def forward(self, inputs):
        return self.decoder(self.encoder(inputs))

Now one can instantiate the module and use it with torch.nn.MSELoss as per usual.

autoencoder =, torch.randn(1, 3, 256, 256))



Latest release:

pip install --user torchlayers


pip install --user torchlayers-nightly


CPU standalone and various versions of GPU enabled images are available at dockerhub.

For CPU quickstart, issue:

docker pull szymonmaszke/torchlayers:18.04

Nightly builds are also available, just prefix tag with nightly_. If you are going for GPU image make sure you have nvidia/docker installed and it’s runtime set.


If you find issue or would like to see some functionality (or implement one), please open new Issue or create Pull Request.

comments powered by Disqus