added ability to pass an arch name to learner instead of class

[oguiza]

No description provided.

[dnth]

I can imagine we will be able to do this soon with tsai and optuna? :)

model = trial.suggest_categorical("model", ["LSTM", "InceptionTimePlus", "TSTPlus", "MLP"])

[oguiza]

I can imagine we will be able to do this soon with tsai and optuna? :)

model = trial.suggest_categorical("model", ["LSTM", "InceptionTimePlus", "TSTPlus", "MLP"])

Exactly! That's the intent of this new enhancement. 👍🏼

[dnth]

I tried something like this once upon a time using eval. But not sure if its got any down side?

    nf = trial.suggest_categorical('num_filters', [32, 64, 96]) 
    learning_rate = trial.suggest_float("learning_rate", 1e-5, 1e-2, log=True)  
    model = trial.suggest_categorical('model', ['InceptionTime', 'InceptionTimePlus'])

    batch_tfms = TSStandardize(by_sample=True)
    learn = TSClassifier(X, y, splits=splits, bs=[64, 128], batch_tfms=batch_tfms,
                         arch=eval(model), arch_config={'nf':nf},
                         metrics=accuracy, cbs=FastAIPruningCallback(trial))

[oguiza]

I'm not sure about eval.

But with the current approach you can try this (I haven't tested it but I think it should work):

nf = trial.suggest_categorical('num_filters', [32, 64, 96]) 
learning_rate = trial.suggest_float("learning_rate", 1e-5, 1e-2, log=True)  
model = trial.suggest_categorical('model', ['InceptionTime', 'InceptionTimePlus'])

batch_tfms = TSStandardize(by_sample=True)
learn = TSClassifier(X, y, splits=splits, bs=[64, 128], batch_tfms=batch_tfms,
                     arch=model, arch_config={'nf':nf},
                     metrics=accuracy, cbs=FastAIPruningCallback(trial))

There shouldn't be any downsides with this approach.