l-berg/landscapes

landscapes

Exploring loss landscapes

Installation

anaconda simplifies dependency management. To install, execute:

conda env create -f torch-land.yml
conda activate torch-land
export PYTHONPATH=.

Scripts

All scripts are run from the projects root directory and further specify their usage when called with the -h flag.

Training

Training a model whose loss landscapes we want to investigate later:

python src/scripts/train.py resnet fashion-mnist

Computing loss landscapes

After having trained a model, to compute loss landscapes (= losses over a 2-dimensional parameter subspace) use the gird.py script, e.g.

python src/scripts/grid.py grid9 resnet fashion-mnist --grid_width=9

Visualizing the results

Now, that the loss values have been computed, visualize the landscapes using 2d heatmaps by calling the visualize.py script with the same parameters.

python src/scripts/visualize.py grid9 resnet fashion-mnist --grid_width=9

Experiments

The commands to run the experiments are documented in the files experiments_run.sh and experiments_visualize.sh.

The landscapes are computed using a pair of random filter-normalized vectors that perturb the model's parameters. The losses correspond to a training step, meaning only a single mini-batch.

We use three pairs of perturbation vectors and the training-set's first three mini-batches of 256 images.

For visualization, we can either look at heatmaps or contour-plots (using the --contour flag on visualize.py):

Some results

Training progress: ResNet on CIFAR-10 before training and after the first and ninth episode

Perturbing only a single layer or even a single conv-filter

First (convolutional) layer:

Last (fully connected) layer:

First filter in first layer:

Different activation functions before and after overfitting: ReLU, sigmoid and tanh

ReLU

sigmoid

tanh

Training progress on Fashion-MNIST

Network architecture: Resnet14 vs VGG11

before overfitting (after 1 epoch)

with overfitting (after 9 epochs)

Zooming in 100x (VGG one shows only noise)