dnnviewer 0.1.0.dev6

Deep Neural Network inspection: view weights, gradients and activations

Deep Neural Network viewer

A dashboard to inspect deep neural network models, providing interactive view on the layer and unit weights and gradients, as well as activation maps.

Current version is targeted at the image classification. However, coming version will target more diverse tasks.

This project is for learning and teaching purpose, do not try to display a network with hundreds of layers.

Install

1. Clone the repository git clone https://github.com/tonio73/dnnviewer.git on your local machine
2. Using Conda or Pip, install the dependencies within your environment following the file environment.yml
3. Run dnnviewer.py with one of the examples below, or with you own model (see below for capabilities and limitations)
4. Access the web application at http://127.0.0.1:8050

Running the program

Currently accepted input formats are Keras Sequential models written to file in Checkpoint format or HDF5. A series of checkpoints along training epochs is also accepted as exemplified below.

Test data is provided by Keras.

Loading a single model

Keras models are loaded from Checkpoint or HDF5 format with option --model-keras <file>

CIFAR-10 Convolutional neural network

$ ./dnnviewer.py --model-keras stimuli/models/CIFAR-10_CNN5.h5 --test-dataset cifar-10

Larger model:

$ ./dnnviewer.py --model-keras stimuli/models/CIFAR-10_LeNetLarge.030.h5 --test-dataset cifar-10

MNIST Convolutional neural network based on LeNet5

$ ./dnnviewer.py --model-keras stimuli/models/MNIST_LeNet60.h5 --test-dataset mnist

MNIST Dense only neural network

$ ./dnnviewer.py --model-keras stimuli/models/MNIST_dense128.h5 --test-dataset mnist

Loading several epochs of a model

Series of models along training epochs are loaded using the argument --sequence-keras <path> and the pattern {epoch} within the provided path. See below on how to generate these checkpoints.

Fashion MNIST convolutionnal network

$ ./dnnviewer.py --sequence-keras "stimuli/models/FashionMNIST_checkpoints/model1_{epoch}" --test-dataset fashion-mnist

Generating the models

From Tensorflow 2.0 Keras

Note: Only Sequential models are currently supported.

Save a single model

Use the save()method of keras.models.Model class the output file format is either Tensorflow Checkpoint or HDF5 based on the extension.

model1.save('models/MNIST_LeNet60.h5')

Save models during training

The Keras standard callback tensorflow.keras.callbacks.ModelCheckpoint is saving the model every epoch or a defined period of epochs:

from tensorflow.keras.callbacks import ModelCheckpoint

callbacks = [
    ModelCheckpoint(
        filepath='checkpoints_cnn-mnistfashion/model1_{epoch}',
        save_best_only=False,
        verbose=1)
]

hist1 = model1.fit(train_images, train_labels, 
                   epochs=nEpochs, validation_split=0.2, batch_size=batch_size,
                   verbose=0, callbacks=callbacks)

Current capabilities

• Load Tensorflow Keras Sequential models and create a display of the network
• Targeted at image classification task (assume image as input, class as output)
• Display series of models over training epochs
• Interactive display and unit weights through connections within the network and histograms
• Supported layers
  • Dense
  • Convolution 2D
  • Flatten
  • Input
• Ignored layers (no impact on the representation)
  • Dropout, ActivityRegularization, SpatialDropout1D/2D/3D
  • All pooling layers
  • BatchNormalization
  • Activation
• Unsupported layers
  • Reshape, Permute, RepeatVector, Lambda, ActivityRegularization, Masking
  • Recurrent layers (LSTM, GRU...)
  • Embedding layers
  • Merge layers

Software architecture

The application is based on Dash for the user interface management (which is generating ReactJs components), and Plotly for the data visualization.

Code is separated in two branches:

• The Graphical representation
• Adapters to read existing models, currently only supporting Keras

Code quality

• Code is PEP8 compliant, thanks to flake8 and Intellij for watching
• There are a few pytest unit tests in unittests/

Run unit test

From project root directory:

$ python -m pytest

Code linting

$ flake8  --max-line-length=120