Yu Zhao

Change `tf.device` to `tf.v1.compat.device` will solve this.

I found `btgym/research/casual_conv/networks.py` have the following code: ``` alignments = attention_mechanism( query_state, attention_mechanism.initial_alignments(tf.shape(inputs)[0], dtype=tf.float32) ) ``` Because `LuongAttention` has a significant change in tensorflow 2.0. So we should update this...

and `lstm_network` in `algorithms/nn/networks.py` also need some updates, because the `dropoutwrapper` will cause error in tensorflow2, use `dropout` parameter instead, according to https://stackoverflow.com/questions/62989175/layernormlstmcell-object-has-no-attribute-zero-state-in-tf-2-2 and https://github.com/tensorflow/tensorflow/issues/29129

And this ``` Traceback (most recent call last): File "/Users/cmal/btgym/btgym/algorithms/aac.py", line 409, in __init__ self.network = pi_global = self._make_policy('global') File "/Users/cmal/btgym/btgym/algorithms/aac.py", line 842, in _make_policy network = self.policy_class(**self.policy_kwargs) File "/Users/cmal/btgym/btgym/algorithms/policy/stacked_lstm.py",...

Has this feature been implemented? Also, is it possible to specify an `:id` to element?

Same problem: ``` $uname -a Darwin ====== 20.2.0 Darwin Kernel Version 20.2.0: Wed Dec 2 20:39:59 PST 2020; root:xnu-7195.60.75~1/RELEASE_X86_64 x86_64 i386 MacBookPro15,2 Darwin $ java -version java version "1.8.0_251" Java(TM)...

Some question..


solved: 

same problem.