Training Models hub.fit(...)

[davidbuniat]

The problem: User wants to train a bunch of models on bunch generated datasets. How she can do that?

import hub
...
model = model()

hub.fit(model, dataset, optimizer, local=True) # <- train locally
hub.fit(..., local=False) # <- train remotely

Train multiple models

for ds in [ds1, ds2, ds3, ...]:
    model = model()
    hub.fit(model1, ds)

Steps

1. Serialize/Deserialize models (start with PyTorch[Lightning])
   saving and loading
2. Collect and store logs/models
3. Retrieve the model
4. Visualize the logs

Roadmap

1. Store and Load Models (have previously done)
2. Store Log[Tensor]

{
    "train_accuracy": ...
    "train_loss": ...
    "val_loss": ... 
    "val_acc": ...
} # <- json object

logs = hub.Dataset(dtype={"train_acc": float, "train_loss": float, ...}, shape=[epoch],) 

for i in range(epoch):
   acc, loss = train_step(ds, model)
   train_acc, train_loss = val_step(ds, model)
   logs["train_loss"][i] = acc
   logs["train_loss"][i] = loss
   ...

3. Write .fit(...) that only trains a model locally. This will help to understand what are the parameters that need to be inserted into the function. Take a look at Pytorch Lightning
4. Remote training of the models, (connect to backend and Hyper)