AML 2022/2023 Project - Vision and Language

Official repository for the "Vision and Language" project - Advanced Machine Learning 2022/2023 @ PoliTo

Getting Started

Make sure to have a CUDA capable device, supporting at least CUDA 11.1, installed and correctly configured on your system.

(The base code of this project has been produced using CUDA 11.3 and Python 3.9.13)

Once you have properly setup everything, make sure you are in the correct directory and run from the command line:

pip install -r requirements.txt

Dataset

1. Download PACS dataset from the portal of the course in the "project_topics" folder.
2. Place the dataset in the 'data/PACS' folder making sure that the images are organized in this way:

data/PACS/kfold/art_painting/dog/pic_001.jpg
data/PACS/kfold/art_painting/dog/pic_002.jpg
data/PACS/kfold/art_painting/dog/pic_003.jpg
...

At this point you should be able to run and edit the base code provided.

Base Code Structure

File	Description
main.py	main entry point. Contains the logic needed to properly setup and run each experiment.
parse_args.py	contains the function responsible for parsing each command line argument.
load_data.py	contains the code to load data, build splits and dataloaders.
models/base_model.py	contains the architectures used in the project.
experiments/baseline.py	contains the code to reproduce the baseline experiment (see point 1. of the project)
experiments/domain_disentangle.py	contains the skeleton code to implement the domain disentanglement experiment (see point 2. of the project)
experiments/clip_disentangle.py	contains the skeleton code to implement the disentanglement experiment using CLIP (see point 4. of the project)

Base Command Line Arguments

Argument	Description
--experiment	which experiment to run chosen among the following:
	baseline: runs the experiment needed to reproduce the baseline (see point 1. of the project)
	domain_disentangle: runs the domain disentanglement experiment (see point 2. of the project)
	clip_disentangle: runs the disentanglement experiment using CLIP (see point 4. of the project)
--target_domain	which visual domain to use as the target domain choosing among art_painting, cartoon, sketch, photo
--lr	learning rate used in the optimization procedure. Do not change it.
--max_iterations	total number of iterations of the optimization procedure. Do not change it, unless you have to reduce batch size. In that case, (max_iterations / batch_size) ratio shall be constant.
--batch_size	batch size used in the optimization procedure. The default value should be fine for GPUs with at least 4GB of dedicated GPU memory. Do not change it, unless you have to reduce batch size. In that case, (max_iterations / batch_size) ratio shall be constant.
--num_workers	total number of worker processes to spawn to speed up data loading. Tweak it according to your hardware specs.
--print_every	controls how frequently the mean training loss is displayed.
--validate_every	controls how frequently the validation procedure occurs. Do not change it.
--output_path	points to the root directory where the records/ folder (containing the results of each experiments) will be created.
--data_path	points to the directory where the PACS dataset is stored. See above (Getting Started >>> Dataset) to properly setup this argument.
--cpu	if set, the experiment will run on the CPU.
--test	if set, the experiment will skip the training procedure and just run the evaluation on the test set.

Baseline Results (see point 1. of the project)

	Art Painting → Cartoon	Art Painting → Sketch	Art Painting → Photo	Average
Baseline	59.04	58.72	94.07	70.61

CLIP Text Encoder

The following code fragment should provide an hint on how to use CLIP's text encoder.

import torch
import clip

# Load CLIP model and freeze it
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')

clip_model, _ = clip.load('ViT-B/32', device='cpu') # load it first to CPU to ensure you're using fp32 precision.
clip_model = clip_model.to(device)
clip_model.eval()
for param in clip_model.parameters():
    param.requires_grad = False

# To use the textual encoder:
description = 'a picture of a small dog playing with a ball'
tokenized_text = clip.tokenize(description).to(device)

text_features = clip_model.encode_text(tokenized_text)