feddg (ELCFS) meta learning implementation is wrong

[mynameisissac]

CCST/federated/fed_run.py

Lines 112 to 122 in 87c3e44

	fast_weights = OrderedDict((name, param - torch.mul(meta_step_size, torch.clamp(grad, 0-clip_value, clip_value))) for
	((name, param), grad) in
	zip(model.named_parameters(), grads))
	learner = copy.deepcopy(model)
	learner.load_state_dict(fast_weights, strict=False)
	output_outer = learner(image_freq)
	del fast_weights
	loss_outer = loss_fun(output_outer, class_l)
	loss = loss_inner + loss_outer

Your meta learning implementation is wrong, load_state_dict() does not preserve the computation graph so loss_outer will never propagate gradient back to the original model parameters. be careful :)