I need to train a model that was saved in .params. I used this post but it works for inference and doesn’t train.
My code looks like:
sym, arg_params, aux_params = mx.model.load_checkpoint(
prefix, 0)
# Dropping the loss from model
new_sym = sym.get_children()[0]
net = gluon.nn.SymbolBlock(outputs=new_sym, inputs=mx.sym.var("data"))
net.initialize(mx.init.Normal(0.002), ctx=ctx)
Then I define loss, optimizer etc. Everything works but the model weights won’t change.
Also when I replaced my model by a single conv layer, it works. So I guess something freezes my weights in the way I load my net.