Conv1D Option for Networks

[kochlisGit]

Hello,
I have noticed that EncodingNetwork has set its conv_type to 2d by default, but there is no way to actually change this parameter. Am i missing anything?

@gin.configurable
class EncodingNetwork(network.Network):
  """Feed Forward network with CNN and FNN layers."""

  def __init__(self,
               input_tensor_spec,
               preprocessing_layers=None,
               preprocessing_combiner=None,
               conv_layer_params=None,
               fc_layer_params=None,
               dropout_layer_params=None,
               activation_fn=tf.keras.activations.relu,
               weight_decay_params=None,
               kernel_initializer=None,
               batch_squash=True,
               dtype=tf.float32,
               name='EncodingNetwork',
               conv_type=CONV_TYPE_2D):

This is the constructor of the Encoding Network. However, no Network allows to change the conv_type.... It is always set on 2D.
Thanks

[sibyjackgrove]

I don't think there is an option for 2D. I had the same requirement and ended up defining my own 1D CNN network and passing it as a preprocessing layer when creating the Actor networ.

preprocessing_layers = tf.keras.models.Sequential([tf.keras.layers.Conv1D(filters=16,kernel_size=1, activation='relu'),
                                                                  tf.keras.layers.Conv1D(filters=32,kernel_size=1, activation='relu'),
                                                                  tf.keras.layers.Flatten(),
                                                                  tf.keras.layers.Dense(units=64, activation='relu'),])

actor_net = actor_distribution_rnn_network.ActorDistributionRnnNetwork(
          tf_env.observation_spec(),
          tf_env.action_spec(),
          preprocessing_layers = preprocessing_layers,
          preprocessing_combiner=None,
          input_fc_layer_params=actor_fc_layers,
          output_fc_layer_params=None,
          lstm_size=lstm_size)

[kochlisGit]

Well, I found a way around this.... I copied the ActorDistributionNetwork class and just passed the argument to the encoding network and it worked! However, it's such a shame that tf-agents, which also supported by tensorflow, doesn't support such simple thing...

[sguada]

For customized networks you need to create your own network, similar to what sibyjackgrove mentioned.

I'm confused, it seems that EncodingNetwork allows 1D Convolutions.

[kochlisGit]

@sguada
Hello,
It does allow 1D convolutions. However, the ActionDistributionNetwork class does not allow you to pass CONV_TYPE_1D to the parameter "conv_type" of the encoding network.

[sguada]

Can you make a PR to allow passing conv_type from ActionDistributionNetwork to EncodingNetwork?