daavoo / dvclive-exp-tracking

Example repo to show how to start tracking experiments in DVC by adding DVCLive to your Python code.

DVCLive Experiment Tracking

Example repo to show how to start tracking experiments in DVC by adding DVCLive to your Python code.

Based on https://pytorch-lightning.readthedocs.io/en/stable/starter/introduction.html.

Instructions

1. Clone the repo:

   git clone [e-mail]:iterative/dvclive-exp-tracking.git
   

2. Install the requirements:

   pip install -r requirements.txt
   

3. Run the code in dvclive_exp_tracking.py:

   python train.py
   

4. Make some changes (for example, modify encoder_size) and run again.

5. Compare experiments:

   a. VS Code: Use the DVC extension.

   

   b. Command line:

    # Show the experiments table.
    $ dvc exp show
     ────────────────────────────────────────────────────────────────────────────────
      Experiment                 Created    train_loss   epoch   step   encoder_size
     ────────────────────────────────────────────────────────────────────────────────
      workspace                  -            0.024467       4     10   96
      main                       10:46 AM            -       -      -   -
      ├── 728aa53 [bumpy-tass]   10:53 AM     0.024467       4     10   96
      ├── ce8adea [hyoid-bosk]   10:53 AM     0.016811       4     10   80
      ├── bec3030 [agaze-gore]   10:52 AM     0.024214       4     10   32
      ├── 0b9595f [myoid-areg]   10:52 AM     0.033773       4     10   16
      └── 8466fbd [nival-flap]   10:52 AM     0.040422       0      2   64
     ────────────────────────────────────────────────────────────────────────────────
   
    # Plot the diff of all experiments.
    $ dvc plots diff $(dvc exp list --name-only)