jettify / pytorch-inspect

torch-inspect

torch-inspect -- collection of utility functions to inspect low level information of neural network for PyTorch

Features

• Provides helper function summary that prints Keras style model summary.
• Provides helper function inspect that returns object with network summary information for programmatic access.
• RNN/LSTM support.
• Library has tests and reasonable code coverage.

Simple example

import torch.nn as nn
import torch.nn.functional as F
import torch_inspect as ti

class SimpleNet(nn.Module):
    def __init__(self):
        super(SimpleNet, self).__init__()
        self.conv1 = nn.Conv2d(1, 6, 3)
        self.conv2 = nn.Conv2d(6, 16, 3)
        self.fc1 = nn.Linear(16 * 6 * 6, 120)
        self.fc2 = nn.Linear(120, 84)
        self.fc3 = nn.Linear(84, 10)

    def forward(self, x):
        x = F.max_pool2d(F.relu(self.conv1(x)), (2, 2))
        x = F.max_pool2d(F.relu(self.conv2(x)), 2)
        x = x.view(-1, self.num_flat_features(x))
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        return x

    def num_flat_features(self, x):
        size = x.size()[1:]
        num_features = 1
        for s in size:
            num_features *= s
        return num_features


  net = SimpleNet()
  ti.summary(net, (1, 32, 32))

Will produce following output:

----------------------------------------------------------------
        Layer (type)               Output Shape         Param #
================================================================
            Conv2d-1           [100, 6, 30, 30]              60
            Conv2d-2          [100, 16, 13, 13]             880
            Linear-3                 [100, 120]          69,240
            Linear-4                  [100, 84]          10,164
            Linear-5                  [100, 10]             850
================================================================
Total params: 81,194
Trainable params: 81,194
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.39
Forward/backward pass size (MB): 6.35
Params size (MB): 0.31
Estimated Total Size (MB): 7.05
----------------------------------------------------------------

For programmatic access to network information there is inspect function:

info = ti.inspect(net, (1, 32, 32))
print(info)

[LayerInfo(name='Conv2d-1', input_shape=[100, 1, 32, 32], output_shape=[100, 6, 30, 30], trainable_params=60, non_trainable_params=0),
 LayerInfo(name='Conv2d-2', input_shape=[100, 6, 15, 15], output_shape=[100, 16, 13, 13], trainable_params=880, non_trainable_params=0),
 LayerInfo(name='Linear-3', input_shape=[100, 576], output_shape=[100, 120], trainable_params=69240, non_trainable_params=0),
 LayerInfo(name='Linear-4', input_shape=[100, 120], output_shape=[100, 84], trainable_params=10164, non_trainable_params=0),
 LayerInfo(name='Linear-5', input_shape=[100, 84], output_shape=[100, 10], trainable_params=850, non_trainable_params=0)]

Installation

Installation process is simple, just:

$ pip install torch-inspect

Requirements

• Python 3.6+
• PyTorch 1.0+

References and Thanks

This package is based on pytorch-summary and PyTorch issue . Compared to pytorch-summary, pytorch-inspect has support of RNN/LSTMs, also provides programmatic access to the network summary information. With a bit more modular structure and presence of tests it is easier to extend and support more features.