celebalucid 0.1.5.3

Analysing Clarity's InceptionV1 network after transfer learning on CelebA dataset.

CelebA & Lucid

This package is a quick access tool of a research project of Alfred Renyi Institute of Mathematics. In our project we use lucid visualizations to get a better understanding how transfer learning manipulates a neural network. As a baseline we took Clarity's InceptionV1 network trained on imagenet and we further trained the network on the CelebA dataset.

In our project we also revisit and rethink the comparison of activations and features.

Networks

We've trained numerous neural networks. At the first stage, we trained three: one learning 100 epochs with SGD optimizer, one learning 10 epochs with Adam and one learning 100 epochs with Adam. We reference these trainings later as sgd, adam and adam_overfit, while the baseline will be referenced as imagenet. Then, we took a permuation of some sampling parameters. We defined 3 different gradient noise (the order of batches the neural network is fed) and 3+1 networks. The 3 networks were initialized with 3 different, randomly sampled weights, while the +1 is the network which were pretrained on imagenet.

	GradNoise1 (gn1)	GradNoise2 (gn2)	GradNoise3 (gn3)
Imagenet (img)	☑	☑	☑
InitNoise1 (in1)	☑	☑	☑
InitNoise2 (in2)	☑	☑	☑
InitNoise3 (in3)	☑	☑	☑

Details of the trainings:

Name	Epochs	Train loss	Train Acc	Val loss	Val acc
sgd	100	0.16	92.7%	0.18	91.9%
adam	10	0.16	92.9%	0.18	91.9%
adam_overfit	100	0.01	99.3%	0.93	90.8%
img-gn1	20	0.11	94.90%	0.21	91.63%
img-gn2	20	0.11	94.89%	0.21	91.60%
img-gn3	20	0.11	94.88%	0.21	91.49%
in1-gn1	20	0.17	92.33%	0.18	91.70%
in1-gn2	20	0.17	92.32%	0.19	91.56%
in1-gn3	20	0.17	92.29%	0.19	91.53%
in2-gn1	20	0.17	92.36%	0.18	91.73%
in2-gn2	20	0.17	92.35%	0.19	91.60%
in2-gn3	20	0.17	92.33%	0.19	91.56%
in3-gn1	20	0.17	92.32%	0.18	91.69%
in3-gn2	20	0.17	92.32%	0.19	91.59%
in3-gn3	20	0.17	92.32%	0.19	91.60%

Setup

Requirements

Python 3.6+

Install

pip install celebalucid

Usage

Load

First, load either of the networks as

from celebalucid import load_model

# Choose from 
# 'imagenet', 'sgd', 'adam', 'adam_overfit'
# 'img-gn1', 'img-gn2', 'img-gn3'
# 'in1-gn1', 'in1-gn2', 'in1-gn3'
# 'in2-gn1', 'in2-gn2', 'in2-gn3'
# 'in3-gn1', 'in3-gn2', 'in3-gn3'
model = load_model('imagenet') 

Change weights

In order to avoid the redefinition of models and unnecessary memory usage, you can update the state dictionary of a model.

model.switch_to('sgd') 

Layer information

Get accessable layers as

model.layer_info

Calling this you will see the list of available layers along with the available number of channels (neurons).

Out: [['conv2d0', 64],
      ['conv2d1', 64],
      ['conv2d2', 192],
      ['mixed3a_1x1', 64],
      ['mixed3a_3x3_bottleneck', 96],
      ...]

Lucid

Visualize a neuron with lucid as

# Default args: size=224, thresholds=[512], progress=False
img = model.lucid('mixed5a_5x5:7', progress=True)
plt.imshow(img)

Weights

You can access the weights of a network as

# Method 1
weights = model.weights.mixed3a_3x3_bottleneck.w 
bias = model.weights.mixed3a_3x3_bottleneck.b

# Method 2
weights = model.weights['mixed3a_3x3_bottleneck'].w 
bias = model.weights['mixed3a_3x3_bottleneck'].b

Activations

In order to see activations, you first need to download the test data (~300MB)

from celebalucid import build_generator

# The test data will be downloaded to this folder
DOWNLOAD_DIR = 'res/images'

# Choose dataset 'celeba' or 'imagenet'
DATASET = 'imagenet'

# This is a torch.utils.data.DataLoader object
# If you already have the dataset downloaded
# to this folder, it will skip the download
generator = build_generator(DOWNLOAD_DIR, DATASET)

Now, utilize the data generator to load images from disk, feed the input with the neural network and then you're free to analyse the activations (aka. neurons.)

for x, y in generator:
    model.stream(x)
    # Now access neurons the same way as on weights
    some_operations(model.neurons.conv2d0)

CKA similarity

The code provides a centered kernel alignment (CKA) comparison between models. In each epoch, the script runs the models on n_iters * batch_size images and then calculates the CKA value for a given layer. It does the same for each epoch and it returns with the mean cka over the epochs.

Define your CKA analyser as

from celebalucid import CKA

# Args: working_dir, dataset, models_to_compare
# Default kwargs: n_epochs=10, n_iters=64, batch_size=32
cka = CKA('res/images', 'celeba', ['imagenet', 'sgd'])

Then run your comparison on a specific layer

# Optional: set verbose=False for no progress bar
cka('mixed4a_3x3') 

Out: 0.95517987

If you want to choose models, instead of creating a new CKA object set new models in order to be GPU RAM efficient:

cka.set_models(['imagenet', 'adam']) 

Weight manipulation

You have the possibility to override a network's weights with constant or another model's weights.

imagenet = load_model('imagenet')
sgd = load_model('sgd')

# Change an entire layer to imagenet
sgd.set_weights('mixed4a_3x3', imagenet)

# Change a single neuron
sgd.set_weights('mixed4a_3x3:0', imagenet)

# Change to a constant
sgd.set_weights('mixed4a_3x3', 0)

It's good to know that by default both weights an biases change. If you know you need only one of these, you can save a bit of computation.

# Default behaviour
sgd.set_weights('mixed4a_3x3', 0, mode='both')

# Change weights only
sgd.set_weights('mixed4a_3x3', 0, mode='weight')

# Change bias only
sgd.set_weights('mixed4a_3x3', 0, mode='bias')

Credits

https://github.com/tensorflow/lucid
https://github.com/greentfrapp/lucent

References

@inproceedings{pmlr-v97-kornblith19a,
  title = {Similarity of Neural Network Representations Revisited},
  author = {Kornblith, Simon and Norouzi, Mohammad and Lee, Honglak and Hinton, Geoffrey},
  booktitle = {Proceedings of the 36th International Conference on Machine Learning},
  pages = {3519--3529},
  year = {2019},
  volume = {97},
  month = {09--15 Jun},
  publisher = {PMLR}
}