Overriding a value does not seem to inherit defaults

[mfoglio]

I am using hydra to setup a pytorch lightning project to train a neural network.
By default, I'd like to:

• Train using a GPU
• Use a batch_size of 448
• Use 64 workers
  I also want to create an optional file where I'd like to store settings that needs to be used when training locally: in fact, when running the project locally, I need to change the 3 parameters above: I want to try on the CPU using 1 worker and a batch_size=2. The idea is that by simply starting the script with the parameter local=default, the script will use all the proper settings needs to run the code locally.
  So far, I achieved everything I wanted to, except using the CPU.
  Here's a simplified structure of my hydra files:
  File config/train.yaml (entrypoint):

# @package _global_
defaults:
  - _self_
  - datamodule: vehicles
  - trainer: default
  - optional local: null  # default

File config/datamodule/vehicles.yaml:

_target_: datamodules.vehicles_datamodule.VehiclesDataModule
batch_size: 448
num_workers: 64

File config/trainer/default.yaml:

_target_: pytorch_lightning.Trainer
accelerator: gpu  # train using the GPU by default
precision: 16

File config/trainer/cpu.yaml (alternative to the default above):

defaults:
  - default

accelerator: cpu
precision: 32

File config/local/default.yaml:

# @package _global_

trainer: cpu  # without this line the training starts properly, however it would train on the GPU

datamodule:  # batch_size and num_workers are updated correctly when running the script with parameter `local=default`
  batch_size: 2
  num_workers: 1

If I add the line trainer: cpu to the config/local/default.yaml I get the following error when accessing cfg.trainer._target_:

  File "/Users/matteo/Documents/Fusus/Coding/ai-models/src/train/train_hydra.py", line 66, in train
    log.info(f"Instantiating trainer <{cfg.trainer._target_}>")
AttributeError: 'str' object has no attribute '_target_'

Why would that be? I thought that the following two line in file config/trainer/cpu.yaml would load the _target_ value from config/local/default.yaml:

defaults:
  - default

However, if I remove the trainer: cpu from file config/local/default.yaml and I run the script with parameters local=default trainer=cpu everything works fine. What am I doing wrong?

[Jasha10]

Hi @mfoglio,

The file config/local/default.yaml file that you pasted above looks like this:

# @package _global_

trainer: cpu

datamodule:
  batch_size: 2
  num_workers: 1

You'll need to change it to look something like this:

# @package _global_

defaults:
  - override /trainer: cpu

datamodule:
  batch_size: 2
  num_workers: 1