`seq_lens` is missing from trained policy TF signature in Ray 1.4

RLlib
1

Hi Ray team,

I’ve used Ray 1.3 and now I’m testing Ray 1.4 to upgrade Ray version for my product.
During my test, I realized that I couldn’t find seq_lens input information from the trained model. I’m not sure this is on purpose or not.

is my simplified test code.

with Ray 1.4

saved_model_cli show --dir /tmp/PPO/cartpole/PPO_CartPole-v0_98860_00000_0_num_sgd_iter\=20\,sgd_minibatch_size\=512\,train_batch_size\=1000_2021-06-14_11-01-51/model --tag_set serve --signature_def serving_default

I couldn’t find sqs_lens
however with Ray 1.3

saved_model_cli show --dir /tmp/PPO/cartpole/PPO_CartPole-v0_fd622_00000_0_num_sgd_iter\=10\,sgd_minibatch_size\=128\,train_batch_size\=1000_2021-06-14_11-04-40/model --tag_set serve --signature_def serving_default

It has

 inputs['seq_lens'] tensor_info:
      dtype: DT_INT32
      shape: (-1)
      name: default_policy/seq_lens:0

I was wondering if Ray 1.4 doesn’t include sqs_lens in the trained TF model on purpose or not.

Thank you.

1 Like
2

Hi @sven1977
Do you have any idea about it?

Thank you.

1 Like