conv_to_fcnet accidentally embeds into smaller space
eugenevinitsky opened this issue · 1 comments
eugenevinitsky commented
The final line:
output = slim.fully_connected(
last_layer,
num_outputs,
weights_initializer=normc_initializer(0.01),
activation_fn=None,
scope="fc_out")
embeds the inputs into num_outputs, but the last hidden_dim should actually go into the LSTM, not into this final hidden layer
eugenevinitsky commented
This is not actually true, the way rllib does things it actually ignores the final output layer before wrapping the LSTM