Torchscan: meaningful module insights

The very useful summary method of tf.keras.Model but for PyTorch, with more useful information.

Table of Contents

• Getting Started
  • Prerequisites
  • Installation
• Usage
• Benchmark
• Technical Roadmap
• Documentation
• Contributing
• Credits
• License

Getting started

Prerequisites

• Python 3.6 (or more recent)
• pip

Installation

You can install the package using pypi as follows:

pip install torchscan

Usage

Similarly to the torchsummary implementation, torchscan brings useful module information into readable format. For nested complex architectures, you can use a maximum depth of display as follows:

from torchvision.models import densenet121
from torchscan import summary

model = densenet121().eval().cuda()
summary(model, (3, 224, 224), max_depth=2)

which would yield

__________________________________________________________________________________________
Layer                        Type                  Output Shape              Param #        
==========================================================================================
densenet                     DenseNet              (-1, 1000)                0              
├─features                   Sequential            (-1, 1024, 7, 7)          0              
|    └─conv0                 Conv2d                (-1, 64, 112, 112)        9,408          
|    └─norm0                 BatchNorm2d           (-1, 64, 112, 112)        257            
|    └─relu0                 ReLU                  (-1, 64, 112, 112)        0              
|    └─pool0                 MaxPool2d             (-1, 64, 56, 56)          0              
|    └─denseblock1           _DenseBlock           (-1, 256, 56, 56)         338,316        
|    └─transition1           _Transition           (-1, 128, 28, 28)         33,793         
|    └─denseblock2           _DenseBlock           (-1, 512, 28, 28)         930,072        
|    └─transition2           _Transition           (-1, 256, 14, 14)         133,121        
|    └─denseblock3           _DenseBlock           (-1, 1024, 14, 14)        2,873,904      
|    └─transition3           _Transition           (-1, 512, 7, 7)           528,385        
|    └─denseblock4           _DenseBlock           (-1, 1024, 7, 7)          2,186,272      
|    └─norm5                 BatchNorm2d           (-1, 1024, 7, 7)          4,097          
├─classifier                 Linear                (-1, 1000)                1,025,000      
==========================================================================================
Trainable params: 7,978,856
Non-trainable params: 0
Total params: 7,978,856
------------------------------------------------------------------------------------------
Model size (params + buffers): 30.76 Mb
Framework & CUDA overhead: 423.57 Mb
Total RAM usage: 454.32 Mb
------------------------------------------------------------------------------------------
Floating Point Operations on forward: 5.74 GFLOPs
Multiply-Accumulations on forward: 2.87 GMACs
Direct memory accesses on forward: 2.90 GDMAs
__________________________________________________________________________________________

Results are aggregated to the selected depth for improved readability.

For reference, here are explanations of a few acronyms:

• FLOPs: floating-point operations (not to be confused with FLOPS which is FLOPs per second)
• MACs: mutiply-accumulate operations (cf. wikipedia)
• DMAs: direct memory accesses (many argue that it is more relevant than FLOPs or MACs to compare model inference speeds cf. wikipedia)

## Benchmark

Below are the results for classification models supported by torchvision for a single image with 3 color channels of size 224x224 (apart from inception_v3 which uses 299x299).

The above results were produced using the scripts/benchmark.py script.

Technical roadmap

The project is currently under development, here are the objectives for the next releases:

• Support of torch.nn.Module layers: ConvTranspose, Identity.

• Package distribution: add a conda package.

• Shared parameter support (cf. discussion)

• Result exporting: add a csv export option.

• Forward pass stat support: RAM usage for each layer, on forward pass.

• Backward pass stat support: RAM usage for each layer, on backward pass.

• Support of torch.nn.Module layers: GroupNorm, Upsample, PixelShuffle, RNN, LSTM, GRU, Embedding, Transformer.

• Support of scripted modules

• Support of torch.nn functional API

• I/O handling: multiple inputs or outputs, non-tensor I/O

• Add computational graph (cf. pytorchviz)

Documentation

The full package documentation is available here for detailed specifications. The documentation was built with Sphinx using a theme provided by Read the Docs.

Contributing

Please refer to CONTRIBUTING if you wish to contribute to this project.

Credits

This project is developed and maintained by the repo owner, but the implementation was inspired or helped by the following contributions:

• Pytorch summary: existing PyTorch porting of tf.keras.Model.summary
• Torchstat: another module inspection tool
• Flops counter Pytorch: operation counter tool
• THOP: PyTorch Op counter
• Number of operations and memory estimation articles by Matthijs Hollemans, and Sicara
• Pruning Convolutional Neural Networks for Resource Efficient Inference

License

Distributed under the MIT License. See LICENSE for more information.