Closed nitsanluke closed 5 years ago
PR #310 adds input size to the printed network summary.
@galnov Sorry to re-open this again. I found that the prints give the same output for both LSTM and FC middleware layers. Is it something that is done purposefully.
eg:
Input Embedder: observation
Input size = [84, 84, 4]
Input Normalization (scale = 255.0, offset = 0.0)
Convolution (num filters = 32, kernel size = 8, stride = 4)
Activation (type = <function relu at 0x1318082f0>)
Convolution (num filters = 64, kernel size = 4, stride = 2)
Activation (type = <function relu at 0x1318082f0>)
Convolution (num filters = 64, kernel size = 3, stride = 1)
Activation (type = <function relu at 0x1318082f0>)
Middleware:
Dense (num outputs = 512)
Activation (type = <function relu at 0x1318082f0>)
Output Head: v_values_head
Dense (num outputs = 1)
Output Head: policy_values_head
Dense (num outputs = 6)
Softmax
coach -r -p Atari_A3C_LSTM --print_networks_summary -lvl air_raid
print_networks_summary
option from the visualization params only provides the input dense layer size. I would like to know whether can get the actual input shape of the environment observation ? I'm working on my own environment and I want to verify whether my environment is correctly set up to get the correct observation shape.