torch-dreams 0.1.0

Feature visualization to make deep neural networks more interpretable

torch-dreams

Making deep neural networks more interpretable, one octave at a time.

pip install torch-dreams --upgrade

Contents:

• You might want to read:
  • Feature visualization by Olah, et al.
  • Google AI blog on DeepDreams
• Notebooks:
  • Quick start notebook
  • optimizing random noise to activate channels within the inceptionv3
  • optimizing images to activate channels within googlenet
• Images:
  • Full size images from notebooks

Quick start

This is a very simple example. For advanced functionalities like simultaneous optimization of channels/layers/units, check out the quick start notebook

• Importing the good stuff

import os
import matplotlib.pyplot as plt
import torchvision.models as models 
from torch_dreams.dreamer import dreamer

• Initiating torch_dreams.dreamer and selecting a layer to optimize on

model = models.inception_v3(pretrained=True)
dreamy_boi = dreamer(model)

layer = model.Mixed_5d
layers_to_use = [layer]  ## feel free to add more layers

• Showtime

os.system("wget https://raw.githubusercontent.com/Mayukhdeb/torch-dreams/master/images/noise.jpg")

out_single_layer = dreamy_boi.deep_dream(
    image_path = "noise.jpg",
    layers = layers_to_use,
    octave_scale = 1.3,
    num_octaves = 7,
    iterations = 100,
    lr = 0.9
)

plt.imshow(out_single_layer)
plt.show()

Optimizing noise to activate multiple channels simultaneously within the inceptionv3

Feature visualization through combined optimization of channels

Changes under way:

1. Expandtorch_dreams to facilitate research in neural network interpretability.