From 28108c905b2d14940c4aa5b7d5e97af74de421ce Mon Sep 17 00:00:00 2001 From: Sven Mika Date: Wed, 9 Dec 2020 08:03:58 +0100 Subject: [PATCH] [RLlib] Tf-eager policy bug fix: Duplicate model call in compute_gradients. (#12682) --- rllib/policy/eager_tf_policy.py | 8 -------- 1 file changed, 8 deletions(-) diff --git a/rllib/policy/eager_tf_policy.py b/rllib/policy/eager_tf_policy.py index a641f0cd0..40a2c7986 100644 --- a/rllib/policy/eager_tf_policy.py +++ b/rllib/policy/eager_tf_policy.py @@ -594,14 +594,6 @@ def build_eager_tf_policy(name, self._is_training = True with tf.GradientTape(persistent=gradients_fn is not None) as tape: - # TODO: set seq len and state-in properly - state_in = [] - for i in range(self.num_state_tensors()): - state_in.append(samples["state_in_{}".format(i)]) - self._state_in = state_in - - model_out, _ = self.model(samples, self._state_in, - samples.get("seq_lens")) loss = loss_fn(self, self.model, self.dist_class, samples) variables = self.model.trainable_variables()