You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Add a script for some basic testing such as finding batch size, testing just the dataset or just the network.
Motivation
When we are developing a new model we want to test parts of our pipeline before pursuing coding. Having a tool to just test some parts (e.g. : dataset, network creation/forward) and some setting tests before starting training (e.g. : batch size) would ease the use of the framework.
To test networks we would need datasets to generate empty items (so we don't need to load an entire dataset) so we can feed the network with batches really fast.
The text was updated successfully, but these errors were encountered:
@Ricocotam
After discussing with @MicaelCarvalho, we think that it is a great idea.
For instance, it would be useful to test that our model is able to overfit a small part of the dataset, or to check the gradient.
However, it's not a priority for now to us.
Feel free to suggest any crazy ideas or implementations which is not something that pytest already does.
Feature
Add a script for some basic testing such as finding batch size, testing just the dataset or just the network.
Motivation
When we are developing a new model we want to test parts of our pipeline before pursuing coding. Having a tool to just test some parts (e.g. : dataset, network creation/forward) and some setting tests before starting training (e.g. : batch size) would ease the use of the framework.
Example of use
I imagine something like :
python -m bootstrap.test --batch_size 50 80 100 --options /path/to/options
python -m bootstrap.test --network forward --options /path/to/options
python -m bootstrap.test --dataset --options /path/to/options
To test networks we would need datasets to generate empty items (so we don't need to load an entire dataset) so we can feed the network with batches really fast.
The text was updated successfully, but these errors were encountered: