Shortcuts

Debugging

The following are flags that make debugging much easier.


fast_dev_run

This flag runs a “unit test” by running 1 training batch and 1 validation batch. The point is to detect any bugs in the training/validation loop without having to wait for a full epoch to crash.

(See: fast_dev_run argument of Trainer)

trainer = Trainer(fast_dev_run=True)

Inspect gradient norms

Logs (to a logger), the norm of each weight matrix.

(See: track_grad_norm argument of Trainer)

# the 2-norm
trainer = Trainer(track_grad_norm=2)

Log GPU usage

Logs (to a logger) the GPU usage for each GPU on the master machine.

(See: log_gpu_memory argument of Trainer)

trainer = Trainer(log_gpu_memory=True)

Make model overfit on subset of data

A good debugging technique is to take a tiny portion of your data (say 2 samples per class), and try to get your model to overfit. If it can’t, it’s a sign it won’t work with large datasets.

(See: overfit_batches argument of Trainer)

# use only 1% of training data (and use the same training dataloader (with shuffle off) in val and test)
trainer = Trainer(overfit_batches=0.01)

# or overfit a number of batches
trainer = Trainer(overfit_batches=0.01)

With this flag, the train, val, and test sets will all be the same train set. We will also replace the sampler in the training set to turn off shuffle for you.



Shorten epochs

Sometimes it’s helpful to only use a percentage of your training, val or test data (or a set number of batches). For example, you can use 20% of the training set and 1% of the validation set.

On larger datasets like Imagenet, this can help you debug or test a few things faster than waiting for a full epoch.

# use only 10% of training data and 1% of val data
trainer = Trainer(limit_train_batches=0.1, limit_val_batches=0.01)

# use 10 batches of train and 5 batches of val
trainer = Trainer(limit_train_batches=10, limit_val_batches=5)

Set the number of validation sanity steps

Lightning runs a few steps of validation in the beginning of training. This avoids crashing in the validation loop sometime deep into a lengthy training loop.

(See: num_sanity_val_steps argument of Trainer)

# DEFAULT
trainer = Trainer(num_sanity_val_steps=5)