Stockroom is a platform to version models, data, parameters, experiment artifacts etc. alongside git versioned source code.
- It is easy. The APIs are very similar to dictionaries in python
- It works alongside Git - in case you need to version source code as well. It's OK if you don't.
- High performance, thanks to the amazing hangar library
- Integration with PyTorch and its ecosystem, so that you don't need to write the complex pipeline code.
One important motivation behind the initial design of stockroom is to avoid users
learning another tool for versioning. We try to make the APIs as minimal and familiar
as possible. Similar to other versioning tools, stockroom let "git" does
and rely on "git" to move between branches/commits. But unlike other tools, we channel
your data access through the smart API so that we don't need to move the huge data files
around when you traverse between commits.
$ pip install stockroom
from stockroom import StockRoom stock = StockRoom(enable_write=True) model.load_state_dict(stock.model['resnet50']) for e in range(epochs): for i in range(limit): optimizer.zero_grad() x, y = stock.data['dataset_name', i] out = model(x) loss = criterion(out, y) loss.backward() optimizer.step() if loss < previous_loss: stock.experiment['loss'] = loss.item() stock.model['resnet50'] = model.state_dict() stock.commit('adding a better model')