### Install Project Dependencies

Source: https://github.com/minghchen/carl_code/blob/master/README.md

Installs necessary packages for the project using conda and pip. It sets up a conda environment and installs PyTorch, torchvision, cudatoolkit, and other required libraries from requirements.txt.

```bash
# create conda env and install packages
conda create -y --name carl python=3.7.9

conda activate carl
# The code is tested on cuda10.1-cudnn7 and pytorch 1.6.0
conda install -y pytorch==1.6.0 torchvision==0.7.0 cudatoolkit=10.1 -c pytorch
conda install -y conda-build ipython pandas scipy pip av -c conda-forge

# install pip packages
pip install --upgrade pip
pip install -r requirements.txt
```

--------------------------------

### Launch TensorBoard for Visualization

Source: https://github.com/minghchen/carl_code/blob/master/README.md

This command starts TensorBoard, a visualization tool for TensorFlow, which is commonly used to monitor training progress and visualize results from machine learning experiments. It points to the log directory where experiment data is stored.

```bash
tensorboard --logdir=~/tmp/scl_transformer_logs
```

--------------------------------

### Start Evaluation with CARL

Source: https://github.com/minghchen/carl_code/blob/master/README.md

This command initiates the evaluation process for the CARL method. It utilizes PyTorch's distributed launch utility and specifies configuration files and log directories for the evaluation.

```bash
python -m torch.distributed.launch --nproc_per_node 1 evaluate.py --workdir ~/datasets --cfg_file ./configs/scl_transformer_config.yml --logdir ~/tmp/scl_transformer_logs
```

--------------------------------

### Prepare FineGym Dataset

Source: https://github.com/minghchen/carl_code/blob/master/README.md

Scripts to prepare the FineGym dataset. This involves downloading raw videos (potentially using youtube-dl) and then running a processing script to standardize resolution and frame rate.

```bash
python dataset_preparation/finegym_process.py
```

--------------------------------

### Download Pretrained ResNet50 Weights

Source: https://github.com/minghchen/carl_code/blob/master/README.md

Provides a link to download ResNet50 backbone weights pre-trained with BYOL. The downloaded weights should be placed in a specific directory for initialization.

```bash
# Download the pretrained weight at [pretrained_models](https://drive.google.com/drive/folders/1VwC4x5xj4Ho3bnh9wZZx--iYhIUguR-q?usp=sharing), and place it at /home/username/datasets/pretrained_models.
```

--------------------------------

### Prepare Pouring Dataset

Source: https://github.com/minghchen/carl_code/blob/master/README.md

Scripts to prepare the Pouring dataset. This involves downloading the data using a shell script and then converting TFRecords to videos.

```bash
sh dataset_preparation/download_pouring_data.sh
python dataset_preparation/tfrecords_to_videos.py
```

--------------------------------

### Prepare PennAction Dataset

Source: https://github.com/minghchen/carl_code/blob/master/README.md

Scripts to prepare the PennAction dataset. This involves downloading the original dataset and label files, then running Python scripts to convert them into TFRecords and subsequently into videos.

```bash
python dataset_preparation/penn_action_to_tfrecords.py
python dataset_preparation/tfrecords_to_videos.py
```

--------------------------------

### Train Model on FineGym Dataset

Source: https://github.com/minghchen/carl_code/blob/master/README.md

Initiates the training process for the FineGym dataset using distributed training. It specifies the working directory, configuration file, and log directory. It also provides a tip for adjusting the number of data workers to prevent CPU overload.

```bash
python -m torch.distributed.launch --nproc_per_node 1 train.py --workdir ~/datasets --cfg_file ./configs/scl_transformer_finegym_config.yml --logdir ~/tmp/scl_transformer_finegym_logs
# Tips: The default number of data worker is 4, which might causes CPU overloaded for some Machines. In this case, you can set --opt DATA.NUM_WORKERS 1.
```

--------------------------------

### Pretrain Model on Kinetics400 Dataset

Source: https://github.com/minghchen/carl_code/blob/master/README.md

Initiates the pretraining process on the Kinetics400 dataset using distributed training. It specifies the working directory, configuration file, and log directory.

```bash
python -m torch.distributed.launch --nproc_per_node 1 train.py --workdir ~/datasets --cfg_file ./configs/scl_transformer_k400_pretrain_config.yml --logdir ~/tmp/scl_transformer_k400_pretrain_logs
```

--------------------------------

### Train Model on PennAction Dataset

Source: https://github.com/minghchen/carl_code/blob/master/README.md

Initiates the training process for the PennAction dataset using distributed training. It specifies the working directory, configuration file, and log directory.

```bash
python -m torch.distributed.launch --nproc_per_node 1 train.py --workdir ~/datasets --cfg_file ./configs/scl_transformer_action_config.yml --logdir ~/tmp/scl_transformer_action_logs
```

--------------------------------

### Train Model on Pouring Dataset

Source: https://github.com/minghchen/carl_code/blob/master/README.md

Initiates the training process for the Pouring dataset using distributed training. It specifies the working directory, configuration file, and log directory. Optional arguments can adjust batch size and epochs.

```bash
python -m torch.distributed.launch --nproc_per_node 1 train.py --workdir ~/datasets --cfg_file ./configs/scl_transformer_config.yml --logdir ~/tmp/scl_transformer_logs
# The config can be changed by adding --opt TRAIN.BATCH_SIZE 1 TRAIN.MAX_EPOCHS 500
```

--------------------------------

### CARL Citation

Source: https://github.com/minghchen/carl_code/blob/master/README.md

This is the BibTeX entry for the CARL paper, providing citation details for academic referencing.

```bibtex
@inproceedings{chen2022framewise,
      title={Frame-wise Action Representations for Long Videos via Sequence Contrastive Learning},
      author={Minghao Chen and Fangyun Wei and Chong Li and Deng Cai},
      booktitle={CVPR},
      year={2022}
}
```

--------------------------------

### Python Project Dependencies

Source: https://github.com/minghchen/carl_code/blob/master/requirements.txt

This snippet lists the Python packages and their exact versions as specified in the project's requirements. These packages cover various functionalities including machine learning (tensorflow-gpu, scikit-learn, Keras-Applications), data manipulation (numpy, pandas), image processing (opencv-python, imageio), and development utilities (docker, jupyterlab).

```python
absl-py==0.13.0
Cython==0.29.24
docker==4.4.4
dtw==1.4.0
easydict==1.9
einops==0.3.0
gpustat==0.6.0
h5py==2.10.0
imageio==2.9.0
imageio-ffmpeg==0.4.4
iopath==0.1.9
jupyterlab==3.0.16
Keras-Applications==1.0.8
Markdown==3.3.4
matplotlib==3.4.2
moviepy==1.0.3
ninja==1.10.0.post3
numba==0.53.1
opencv-python==4.5.3.56
prettytable==2.1.0
PyYAML==5.4.1
scikit-build==0.11.1
scikit-learn==0.24.2
seaborn==0.11.2
simplejson==3.17.3
spacy==3.1.1
tensorboard==2.5.0
tensorboardX==2.4
tensorflow-gpu==2.3.0
timm==0.3.2
yacs==0.1.8
youtube-dl==2021.6.6
```

=== COMPLETE CONTENT === This response contains all available snippets from this library. No additional content exists. Do not make further requests.