Torch-Dreams

Making neural networks more interpretable, for research and art.

pip install torch-dreams 

Quick start

Make sure you also check out the quick start colab notebook and the docs for more interesting examples.

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

model = models.inception_v3(pretrained=True)
dreamy_boi = dreamer(model,  device = 'cuda', quiet =  False)

image_param = dreamy_boi.render(
    layers = [model.Mixed_5b],
)

plt.imshow(image_param.rgb)
plt.show()

You can also optimize activations from multiple models simultaneously

First, let's pick 2 models and specify which layers we'd want to work with

from torch_dreams.model_bunch import ModelBunch

bunch = ModelBunch(
    model_dict = {
        'inception': models.inception_v3(pretrained=True).eval(),
        'resnet':    models.resnet18(pretrained= True).eval()
    }
)

layers_to_use = [
            bunch.model_dict['inception'].Mixed_6a,
            bunch.model_dict['resnet'].layer2[0].conv1
        ]

dreamy_boi = dreamer(model = bunch, quiet= True, device= 'cuda')

Then define a custom_func which determines which exact activations of the models we have to optimize

def custom_func(layer_outputs):
    loss =   layer_outputs[0].mean()*2.0 + layer_outputs[1][89].mean() 
    return -loss

Run the optimization

image_param = dreamy_boi.render(
    layers = layers_to_use,
    custom_func= custom_func,
    iters= 100
)

plt.imshow(image_param.rgb)
plt.show()

Visualize individual channels

layers_to_use = [model.Mixed_6b.branch1x1.conv]

def make_custom_func(layer_number = 0, channel_number= 0): 
    def custom_func(layer_outputs):
        loss = layer_outputs[layer_number][channel_number].mean()
        return -loss
    return custom_func

my_custom_func = make_custom_func(layer_number= 0, channel_number = 119)

image_param = dreamy_boi.render(
    layers = layers_to_use,
    custom_func = my_custom_func,
)
plt.imshow(image_param.rgb)
plt.show()

Acknowledgements

• amFOSS
• Gene Kogan