CMU 11-775 Fall 2022 Homework 2

PDF Handout

In this homework we will perform a video classification task with visual features.

Recommended Hardware

This code template is built based on PyTorch and Pyturbo for Linux to fully utilize the computation of multiple CPU cores and GPUs. SIFT feature, K-Means, and Bag-of-Words must run on CPUs, while CNN features and MLP classifiers can run on GPUs. For GCP, an instance with 16 vCPU (e.g. n1-standard-16) and a Nvidia T4 GPU instance should be sufficient for the full pipeline. During initial debugging, you are recommended to use a smaller instance to save money, e.g., n1-standard-x (x = the number of vCPUs) with Nvidia T4 or without GPU for the SIFT part. You may not be able to start a GPU VM with a very small number of vCPUs (e.g., only 1 vCPU). You can gradually increase the number of vCPUs and see whether a GPU VM can be launched.

Install Dependencies

A set of dependencies is listed in environment.yml. You can use conda to create and activate the environment easily.

# Start from within this repo
conda env create -f environment.yml -p ./env
conda activate ./env

Dataset

You will continue using the data from Homework 1 for this homework, which you should have downloaded.

If you don't have the data, download it from AWS S3 with the following commands:

# Start from within this repo
cd ./data
# Download and decompress data (no need if you still have it from HW1)
wget https://cmu-11775-vm.s3.amazonaws.com/spring2022/11775_s22_data.zip
unzip 11775_s22_data.zip
rm 11775_s22_data.zip

Eventually, the directory structure should look like this:

• this repo
  • code
  • data
    • videos (unzipped from 11775_s22_data.zip)
    • labels (unzipped from 11775_s22_data.zip)
  • env
  • ...

Development and Debugging

Some functions in the pipeline are deliberately left blank for you to implement, where an NotImplementedError will be raised. We recommend you generate a small file list (e.g. debug.csv with 20 lines) for fast debugging during initial development. The --debug option in some scripts are also very helpful. In addition, you can enable pdb debugger upon exception

# Instead of 
python xxx.py yyy zzz
# Run
ipython --pdb xxx.py -- yyy zzz

SIFT Features

To extract SIFT features, use

python code/run_sift.py data/labels/xxx.csv

By default, features are stored under data/sift.

To train K-Means with SIFT feature for 128 clusters, use

python code/train_kmeans.py data/labels/xxx.csv data/sift 128 sift_128

By default, model weights are stored under data/kmeans.

To extract Bag-of-Words representation with the trained model, use

python code/run_bow.py data/labels/xxx.csv sift_128 data/sift

By default, features are stored under data/bow_<model_name> (e.g., data/bow_sift_128).

CNN Features

To extract CNN features, use

python code/run_cnn.py data/labels/xxx.csv

By default, features are stored under data/cnn.

3D CNN Features

To extract 3D CNN features, use

python code/run_cnn3d.py data/labels/xxx.csv

By default, features are stored under data/cnn3d.

MLP Classifier

The training script automatically and deterministically split the train_val data into training and validation, so you do not need to worry about it.

To train MLP with SIFT Bag-of-Words, run

python code/run_mlp.py sift --feature_dir data/bow_sift_128 --num_features 128

To train MLP with CNN features, run

python code/run_mlp.py cnn --feature_dir data/cnn --num_features <num_feat>

By default, training logs and predictions are stored under data/mlp/model_name/version_xxx/. You can directly submit the CSV file to Kaggle.