tsai

tsai copied to clipboard

Context: I am using the Flower framework along with a custom gMLP model to perform federated learning. I encountered an issue when trying to use the get_X_preds method with the gMLP model. It seems that the gMLP model does not have a get_X_preds method, which results in an error.

Error Message: I am encountering the following error message when attempting to run federated learning with the FlowerClient class: AttributeError: 'gMLP' object has no attribute 'get_X_preds'

Steps to Reproduce: To reproduce the issue, you can follow these steps:

1. Create a FlowerClient class similar to the one mentioned below:

model = gMLP(2, 1, seq_len=seq_length).to(device)
class FlowerClient(fl.client.NumPyClient):
    def __init__(self, model, data):
        self.model = model
        self.X_train, self.y_train, self.X_test, self.y_test = data
        splits = RandomSplitter(valid_pct=0.2, seed=42)(range_of(self.X_train))
        tfms = [None, [TSStandardize(by_sample=True, by_var=True)]]
        batch_tfms = [TSStandardize(by_sample=True, by_var=True)]
        train_dls = get_ts_dls(self.X_train, self.y_train, splits=splits, tfms=tfms, batch_tfms=batch_tfms, bs=64)
        self.learn = ts_learner(train_dls, self.model, metrics=[mae, rmse])

    def get_parameters(self, config):
        parameters = [p.detach().cpu().numpy() for p in self.model.parameters()]
        return parameters

    def fit(self, parameters, config):      
        with torch.no_grad():
            for i, (param, param_np) in enumerate(zip(self.model.parameters(), parameters)):
                param.copy_(torch.Tensor(param_np))
        
        self.learn.fit_one_cycle(2, 1e-3)       
        updated_parameters = self.get_parameters(self.model)
        return updated_parameters, len(self.X_train), {}

    def evaluate(self, parameters, config):       
        with torch.no_grad():
            for i, (param, param_np) in enumerate(zip(self.model.parameters(), parameters)):
                param.copy_(torch.Tensor(param_np).to(device))
                
        probas, targets, preds = self.learn.get_X_preds(self.X_test, self.y_test, with_decoded=True)
        probas = probas.to(device)
        targets = targets.to(device)
        preds = preds.to(device)
        mae_score = mae(preds, targets)
        rmse_score = rmse(preds, targets)
        return {"rmse": float(rmse_score), "mae": float(mae_score)}

2. Use a gMLP model as the self.model in the FlowerClient.
3. Execute federated learning rounds with Flower using this client.

Expected Behavior: I expected the federated learning process to work correctly with the gMLP model and FlowerClient, including the evaluation step that uses the get_X_preds method.

Actual Behavior: Instead, I encountered the AttributeError mentioned above, indicating that the gMLP model does not have a get_X_preds method.

Additional Information: Python version: 3.8 PyTorch version: 2.0.1 Flower version: 1.5.0 Tsai version: 0.3.7

Please let me know if there are any specific details or additional information that should be included in the issue. Thank you for your assistance in resolving this problem.