torchlurk 0.1.3

A CNN visualization library for pytorch

Introduction

CNNs do a great job at recognizing images (when appropriately trained).

Problems arise when it comes to interpret the network: although one coded the network in question and knows all the tips and tricks necessary to train it efficiently, one might ignore how it generates the output from a given image.

Torchlurk aims at helping the user in that sense: it provides an interface to visualize a Pytorch network in an efficient yet simple manner, similarly to Microscope.

All you need is the trained Pytorch network and its training set. That's it.

Installation ☕

Torchlurk is available on pip! Just run:

pip install torchlurk

Overview ☝

Documentation 📚

Torchlurk has an online documentation which gets regularly updated.

Quick Start ⌛

Your training set should follow the following structure in order for the lurker to load properly your datas:

.
├── src                  
│   ├── name_class1
│   │   ├── class1id_1.jpg
│   │   ├── class1id_2.jpg
│   │   ├── ...
│   ├── name_class2
│   │   ├── class2id_1.jpg
│   │   ├── class2id_2.jpg
│   │   ├── ...
│   ├── ...

1. Instanciation

import torchlurk
import torch

# load the trained model
your_model = ModelClass() 
your_model.load_state_dict(torch.load(PATH))

# the preprocess used for the training
preprocess = transforms.Compose(...)

# and instanciate a lurker
lurker = Lurk(your_model,
          preprocess,
          save_gen_imgs_dir='save/dir',
          save_json_path='save/dir',
          imgs_src_dir=".source/dir",
          side_size=224)

2. Layer Visualization

The layer visualization is an artificial image generated by gradient descent which aims at maximizing the acivation of a given filter: this gives useful insights on the type of texture/colors the filter in question is looking at in inputs images.

# compute the layer visualisation for a given set of layers/filters
lurker.compute_layer_viz(layer_indx = 12,filter_indexes=[7])
# OR compute it for the whole network
lurker.compute_viz()
# plot the filters
lurker.plot_filter_viz(layer_indx=12,filt_indx=7)

3. Max Activation

The max activation represents the top N images activating a given filter the most in terms of average or max score.

#compute the top activation images
lurker.compute_top_imgs(compute_max=True,compute_avg=True)
# plot them
lurker.plot_top("avg",layer_indx=12,filt_indx=7)

3.1 Deconvolution

# plot the max activating images along with their cropped areas
lurker.plot_crop(layer_indx=2,filt_indx=15)

4. Gradients

Guided GRAD CAM is another way to check what a given filter is looking at: it relies in particular in isolating a specific location in the image that excites our neurons. For more information, check this article.

#compute the gradients
lurker.compute_grads()
# plot them
lurker.plot_top("avg",layer_indx=12,filt_indx=7,plot_imgs=False,plot_grads=True)

5. Histograms

Torchlurk allows you to visualize the most activating classes of the training using histograms: a very peaked distribution is often asssociated with a specialized filter.

# display the 
lurker.plot_hist(layer_indx=12,filt_indx=7,hist_type="max",num_classes=12)

6. Serving

Torchlurk is equiped with a live update tool which allows you to visualize your computed results while coding.

#serve the application on port 5001
lurker.serve(port=5001)

lurker.end_serve()

Happy Lurking!

🕵