JaxGCRL

https://michalbortkiewicz.github.io/JaxGCRL/
Admin
JaxGCRL is a high-performance library and benchmark for self-supervised goal-conditioned...
Tokens:1,802
Snippets:32
Trust Score:9.7
Update:6 months ago
Show doc for...
Tokens:
Raw
### Verify Installation

Source: https://michalbortkiewicz.github.io/JaxGCRL/install

Make the training script executable and run it to confirm the environment is correctly configured.

```bash
chmod +x ./scripts/train.sh; ./scripts/train.sh
```

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

### Access Help Documentation

Source: https://michalbortkiewicz.github.io/JaxGCRL/usage

Display available command-line arguments and hyperparameters.

```bash
python run.py crl --help
```

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

### Run Basic Experiment

Source: https://michalbortkiewicz.github.io/JaxGCRL/usage

Execute a standard experiment using the specified environment.

```bash
python run.py crl --env ant
```

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

### Create Conda Environment

Source: https://michalbortkiewicz.github.io/JaxGCRL/install

Use the environment.yml file to set up the required dependencies for the project.

```bash
conda env create -f environment.yml
```

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

### Run Training with Custom Contrastive Objective

Source: https://michalbortkiewicz.github.io/JaxGCRL/methods

Execute training with a newly registered contrastive objective using the `run.py` script.

```bash
python run.py crl --contrastive_loss_fn "your_loss"
```

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

### Configure Parallel Environments

Source: https://michalbortkiewicz.github.io/JaxGCRL/usage

Adjust the number of parallel environments and batch size for data collection. Ensure the replay buffer constraint is met: num_envs * (episode_length - 1) must be divisible by batch_size.

```bash
python run.py crl --env ant --num_envs 16 --batch_size 16
```

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

### Execute Multiple Experiments via Bash

Source: https://michalbortkiewicz.github.io/JaxGCRL/usage

Run a series of experiments across different seeds and contrastive loss functions.

```bash
for c_loss in binary infonce flatnce fb dp; do
    for seed in 1 2 3 4 5; do
        python run.py crl --seed ${seed} \
        --wandb_project_name crl --wandb_group contrastive_loss_experiments \ 
        --exp_name ${c_loss} \
        --contrastive_loss_fn ${c_loss} --energy_fn norm \
        --log_wandb
    done
done
```

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

### Manage Wandb Sweeps

Source: https://michalbortkiewicz.github.io/JaxGCRL/usage

Commands to initialize and execute hyperparameter sweeps using Wandb.

```bash
wandb sweep --project exemplary_sweep ./scripts/sweep.yml
```

```bash
wandb agent <previous_command_output>
```

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

### Customize Training Hyperparameters

Source: https://michalbortkiewicz.github.io/JaxGCRL/usage

Modify neural network architecture and loss function settings.

```bash
python run.py crl --env ant --h_dim 128 --n_hidden 3 --repr_dim 32
```

```bash
python run.py crl --env ant --energy_fn l2 --contrastive_loss_fn bwd_infonce
```

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

### Register New Contrastive Objective

Source: https://michalbortkiewicz.github.io/JaxGCRL/methods

Register a new contrastive objective function by adding it to the `contrastive_loss_fn` in `agents/crl/losses.py`.

```python
if name == "your_loss":
    critic_loss = ...
...
```
Tokens:
Raw