Edit on GitHub

Hydra Composition

Hydra is a framework to configure complex applications. DVC supports Hydra's config composition as a way to configure experiment runs.

At the moment you must explicitly enable this feature with:

$ dvc config hydra.enabled True

How it works

Upon dvc exp run, Hydra will be used to compose a single params.yaml before executing the experiment. Its parameters will configure the underlying DVC pipeline that contains your experiment.

DVC pipelines can run multiple stages with different shell commands instead of a single script, and offer advanced features like templating and foreach stages.

Setting up Hydra

First you need a conf/ directory for Hydra config groups. For example:

โ”œโ”€โ”€ config.yaml
โ”œโ”€โ”€ dataset
โ”‚   โ”œโ”€โ”€ imagenette.yaml
โ”‚   โ””โ”€โ”€ imagewoof.yaml
โ””โ”€โ”€ train
    โ”œโ”€โ”€ baseline.yaml
    โ”œโ”€โ”€ model
    โ”‚   โ”œโ”€โ”€ alexnet.yaml
    โ”‚   โ”œโ”€โ”€ efficientnet.yaml
    โ”‚   โ””โ”€โ”€ resnet.yaml
    โ””โ”€โ”€ optimizer
        โ”œโ”€โ”€ adam.yaml
        โ””โ”€โ”€ sgd.yaml

You also need a conf/config.yaml file defining the Hydra defaults list:

  - dataset: imagenette
  - train/model: resnet
  - train/optimizer: sgd

Use dvc config hydra options to change the default locations for the config groups directory and defaults list file.

Now let's look at what the resulting params.yaml could be based on the setup above:

  url: https://s3.amazonaws.com/fast-ai-imageclas/imagenette2-160.tgz
  output_folder: imagenette
    name: ResNet
    size: 50
    weights: ResNet50_Weights.IMAGENET1K_V2
    name: SGD
    lr: 0.001
    momentum: 0.9

Running experiments

Let's build an experimental pipeline with 2 stages. The first one downloads a dataset and uses the parameters defined in the dataset section of params.yaml. The second stage trains an ML model and uses the rest of the parameters (entire train group).

      - wget ${dataset.url} -O tmp.tgz
      - mkdir -p ${dataset.output_folder}
      - tar zxvf tmp.tgz -C ${dataset.output_folder}
      - rm tmp.tgz
      - ${dataset.output_folder}
    cmd: python train.py
      - ${dataset.output_folder}
      - train

We parametrize the shell commands above (mkdir, tar, wget) as well as output and dependency paths (outs, deps) using templating (${} expression).

You can use dvc.api.params_show() to load params in Python code. For other languages, use dictionary unpacking or a YAML parsing library.

We can now use dvc exp run --set-param to modify the config composition on-the-fly, for example loading the model config from train/model/efficientnet.yaml (instead of resnet.yaml from the defaults list):

$ dvc exp run --set-param 'train/model=efficientnet'
Stage 'setup-dataset' didn't change  skipping
Running stage 'train':
> python train.py
--model.name 'EfficientNet' --model.size 'b0' --model.weights 'ResNet50_Weights.IMAGENET1K_V2'
--optimizer.name 'SGD' --optimizer.lr 0.001 --optimizer.momentum 0.9

Add -vv (dvc exp run -vv --set-param 'train/model=efficientnet') if you need to debug the composed values.

We can also modify specific values from any config section:

$ dvc exp run --set-param 'train.optimizer.lr=0.1'
Stage 'setup-dataset' didn't change, skipping
Running stage 'train':
> python train.py
--model.name 'EfficientNet' --model.size 'b0' --model.weights 'ResNet50_Weights.IMAGENET1K_V2'
--optimizer.name 'SGD' --optimizer.lr 0.01 --optimizer.momentum 0.9

We can also load multiple config groups in an experiments queue, for example to run a grid search of ML hyperparameters:

$ dvc exp run --queue \
              -S 'train/optimizer=adam,sgd' \
              -S 'train/model=resnet,efficientnet'

Queueing with overrides '{'params.yaml': ['optimizer=adam', 'model=resnet']}'.
Queued experiment 'ed3b4ef' for future execution.
Queueing with overrides '{'params.yaml': ['optimizer=adam', 'model=efficientnet']}'.
Queued experiment '7a10d54' for future execution.
Queueing with overrides '{'params.yaml': ['optimizer=sgd', 'model=resnet']}'.
Queued experiment '0b443d8' for future execution.
Queueing with overrides '{'params.yaml': ['optimizer=sgd', 'model=efficientnet']}'.
Queued experiment '0a5f20e' for future execution.

$ dvc queue start

Note that DVC keeps a cache of all runs, so many permutations will be completed without actually having to run the experiment. In the above example, the experiment with ['optimizer=sgd', 'model=resnet'] will not waste computing time because the results are already in the run cache. You can confirm this with dvc queue logs:

$ dvc queue logs 0b443d8
Stage 'setup-dataset' didn't change, skipping
Stage 'train' didn't change, skipping

๐Ÿ› Found an issue? Let us know! Or fix it:

Edit on GitHub

โ“ Have a question? Join our chat, we will help you:

Discord Chat