AML 2023/2024 Project - Activation Shaping for Domain Adaptation

"Activation Shaping for Domain Adaptation" - Vincenzo Dalia, Mario Todaro, Fabio Barbieri

Base Code Structure

PLEASE NOTE THAT PROJECT STRUCTURE HAS (slightly) CHANGED FROM THE ORIGINAL

In the following, you can find a brief description of the included files.

File	Description
[...]
models/resnet.py	contains the resnet18 baseline
models/as_module.py	contains the custom Activation Shaping Module
models/ras_resnet.py	contains the architecture that performs random shaping with the custom module
models/da_resnet.py	contains the architecture that performs domain adaptation with the custom module

Running The Experiments

To run the experiments you can use, the provided launch scripts. Example usage:

# Eventually give script permissions first, then
./launch_scripts/baseline.sh cartoon
./launch_scripts/random_maps.sh sketch --mask_ratio 0.9
./launch_scripts/domain_adapt.sh photo --layers 3.1.2 4.1.bn2 --topK --tk_treshold 0.5 --no_binarize

In the following, you can find a brief description of the relevant command line arguments when launching the experiments.

Command Line Arguments (Hyperparameters)

Argument	Description
--experiment	the identifier of the experiment to run. (Eg. baseline)
--experiment_name	the name of the path where to save the checkpoint and logs of the experiment. (Eg. baseline/cartoon)
[...]
--layers	The layers after which to hook the activation shaping module. Any number of strings following this pattern: RESNET_LAYER.LEVEL CONV_NUM, for example: 2.0.1 corresponds to resnet layer2.0.conv1. To hook a relu layer, use the pattern : 2.0.r, for layer2.0.relu. To hook the avgpool, use "avgpool". To hook the first convolution, resnet conv1, use : 1. Invalid/Duplicated layers are ignored.
--topK	if set, topK shaping of activation map is performed (for both experiments)
--tk_treshold	if set, and the topK flag is set, this controls the K in the topK shaping of activation maps
--no_binarize	if set, the activation shaping module will not binarize the masks
--mask_ratio	it controls the number (%) of ones in the random masks (random_maps experiment only), defaults to 1
--use_bernoulli	if the experiment is random_maps, it sets the number of 1s drawing from a mask_ratio bernoulli dist. This will make random_maps faster to run, but results may be different from setting exact number of 0s and 1s. Furthermore, this does not create a truly random maps, but a random tensor composed of just 0s and 1s

Note: you must use --flag running the scripts to pass other args then the experiment name. so:

./launch_scripts/domain_adapt.sh cartoon --layers 2.1.2 4.1.2 --topK --no_binarize  --tk_treshold=0.5

./launch_scripts/random_maps.sh cartoon --layers 2.1.2 --mask_ratio 0.9

Baseline Results (see point 0. of the project)

	Art Painting → Cartoon	Art Painting → Sketch	Art Painting → Photo	Average
Baseline	54.52	40.44	95.93	63.63

Domain adaptation most improving experiments

	Art Painting → Cartoon	Art Painting → Sketch	Art Painting → Photo	Average
4.1.2, 4.1.bn2, topK90, eps = 0.1	60,58	49,27	94,73	68,19

Random shaping most improving experiments

	Art Painting → Cartoon	Art Painting → Sketch	Art Painting → Photo	Average
2.1.2, mask_ratio = 0.9	59,34	52,46	95,33	69,04