When trying to create a neural network and optimize it using Pytorch, I am getting
ValueError: optimizer got an empty parameter list
Here is the code.
import torch.nn as nn import torch.nn.functional as F from os.path import dirname from os import getcwd from os.path import realpath from sys import argv class NetActor(nn.Module): def __init__(self, args, state_vector_size, action_vector_size, hidden_layer_size_list): super(NetActor, self).__init__() self.args = args self.state_vector_size = state_vector_size self.action_vector_size = action_vector_size self.layer_sizes = hidden_layer_size_list self.layer_sizes.append(action_vector_size) self.nn_layers = [] self._create_net() def _create_net(self): prev_layer_size = self.state_vector_size for next_layer_size in self.layer_sizes: next_layer = nn.Linear(prev_layer_size, next_layer_size) prev_layer_size = next_layer_size self.nn_layers.append(next_layer) def forward(self, torch_state): activations = torch_state for i,layer in enumerate(self.nn_layers): if i != len(self.nn_layers)-1: activations = F.relu(layer(activations)) else: activations = layer(activations) probs = F.softmax(activations, dim=-1) return probs and then the call
self.actor_nn = NetActor(self.args, 4, 2, [128]) self.actor_optimizer = optim.Adam(self.actor_nn.parameters(), lr=args.learning_rate) gives the very informative error
ValueError: optimizer got an empty parameter list
I find it hard to understand what exactly in the network's definition makes the network have parameters.
I am following and expanding the example I found in Pytorch's tutorial code.
I can't really tell the difference between my code and theirs that makes mine think it has no parameters to optimize.
How to make my network have parameters like the linked example?
1 Answer
Your NetActor does not directly store any nn.Parameter. Moreover, all other layers it eventually uses in forward are stored as a simple list in self.nn_layers.
If you want self.actor_nn.parameters() to know that the items stored in the list self.nn_layers may contain trainable parameters, you should work with containers.
Specifically, making self.nn_layers to be a nn.ModuleList instead of a simple list should solve your problem:
self.nn_layers = nn.ModuleList() 
