About "getting_started.ipynb"

[KarahanS]

Hi there,

Why did we use batch_size * 2 when we initialize the dataloaders in the following part:

train_loader = data.DataLoader(dataset=train_dataset, batch_size=BATCH_SIZE, shuffle=True)
train_loader_at_eval = data.DataLoader(dataset=train_dataset, batch_size=2*BATCH_SIZE, shuffle=False)
test_loader = data.DataLoader(dataset=test_dataset, batch_size=2*BATCH_SIZE, shuffle=False)

Also, why did we use the train_dataset for evaluation during training as well? Wouldn't it be a better practice to use val split of the dataset? Is there a specific reason for your choices?

[aymuos15]

1. The bigger batch size is just for faster inference.
2. Regarding evaluation, I imagine its simply to gauge train vs test.

[KarahanS]

Using the same training set for validation during training doesn't really make sense though. The idea is that we have to use samples that the model hasn't seen before and measure the metrics on it.

[aymuos15]

Yeah, that's fair. Waiting to see their reply for that I guess now haha.

[duducheng]

Hi guys,

I purely agree with @aymuos15 ‘s reply. ;)

The batch size is only for faster inference. We also monitor the training set performance because the metrics on the whole training set is more consistent than those monitored on training batches. I guess it’s a standard practice especially for research projects which need reporting in the papers.

Nevertheless, it’s only a design choice that has no direct impact on model performances. Feel free to customize your training pipeline, or you can also integrate with popular training framework like PyTorch Lightning.

Jiancheng