Epsilon greedy policy error when generating new action from Dict Action space.

[JaCoderX]

I'm trying to convert a custom gym project (called BTgym) to work as a tf-agent env.
the original observation space and the action space are both gym.spaces.Dict.
but for the moment I have simplified the observation space so I can fit the env to run using the same code of DQN tutorial example (as a proof of concept). so the modified spaces are as follows:

Observation Spec:
BoundedTensorSpec(shape=(6, 1, 5), dtype=tf.float32, name='observation/external', minimum=array(-100., dtype=float32), maximum=array(100., dtype=float32))
Action Spec:
OrderedDict([('default_asset', BoundedTensorSpec(shape=(), dtype=tf.int64, name='action/default_asset', minimum=array(0), maximum=array(3)))]) 

Error occur under Training the agent section, when performing collect_step():

Traceback (most recent call last):
  File "/home/jack/envTest.py", line 251, in <module>
    collect_step(train_env, agent.collect_policy, replay_buffer)
  File "/home/jack/envTest.py", line 209, in collect_step
    action_step = policy.action(time_step)
  File "/home/jack/tf_agents/policies/tf_policy.py", line 278, in action
    step = action_fn(time_step=time_step, policy_state=policy_state, seed=seed)
  File "/home/jack/tf_agents/utils/common.py", line 131, in with_check_resource_vars
    return fn(*fn_args, **fn_kwargs)
  File "/home/jack/tf_agents/policies/epsilon_greedy_policy.py", line 106, in _action
    random_action.action)
  File "/home/jack/anaconda3/envs/deep/lib/python3.6/site-packages/tensorflow_core/python/util/dispatch.py", line 180, in wrapper
    return target(*args, **kwargs)
  File "/home/jack/anaconda3/envs/deep/lib/python3.6/site-packages/tensorflow_core/python/ops/array_ops.py", line 3753, in where
    return gen_math_ops.select(condition=condition, x=x, y=y, name=name)
  File "/home/jack/anaconda3/envs/deep/lib/python3.6/site-packages/tensorflow_core/python/ops/gen_math_ops.py", line 9430, in select
    condition, x, y, name=name, ctx=_ctx)
  File "/home/jack/anaconda3/envs/deep/lib/python3.6/site-packages/tensorflow_core/python/ops/gen_math_ops.py", line 9462, in select_eager_fallback
    _attr_T, _inputs_T = _execute.args_to_matching_eager([x, y], _ctx)
  File "/home/jack/anaconda3/envs/deep/lib/python3.6/site-packages/tensorflow_core/python/eager/execute.py", line 257, in args_to_matching_eager
    t, dtype, preferred_dtype=default_dtype, ctx=ctx))
  File "/home/jack/anaconda3/envs/deep/lib/python3.6/site-packages/tensorflow_core/python/framework/ops.py", line 1296, in internal_convert_to_tensor
    ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)
  File "/home/jack/anaconda3/envs/deep/lib/python3.6/site-packages/tensorflow_core/python/framework/constant_op.py", line 286, in _constant_tensor_conversion_function
    return constant(v, dtype=dtype, name=name)
  File "/home/jack/anaconda3/envs/deep/lib/python3.6/site-packages/tensorflow_core/python/framework/constant_op.py", line 227, in constant
    allow_broadcast=True)
  File "/home/jack/anaconda3/envs/deep/lib/python3.6/site-packages/tensorflow_core/python/framework/constant_op.py", line 235, in _constant_impl
    t = convert_to_eager_tensor(value, ctx, dtype)
  File "/home/jack/anaconda3/envs/deep/lib/python3.6/site-packages/tensorflow_core/python/framework/constant_op.py", line 96, in convert_to_eager_tensor
    return ops.EagerTensor(value, ctx.device_name, dtype)
ValueError: Attempt to convert a value (DictWrapper({'default_asset': <tf.Tensor: id=132834, shape=(1,), dtype=int64, numpy=array([0])>})) with an unsupported type (<class 'tensorflow.python.training.tracking.data_structures._DictWrapper'>) to a Tensor.

it seems that epsilon greedy policy have some problem with the Dict action space when trying to generate action action_step = policy.action(time_step)
as both DQN and random agents seems to work fine for producing actions, Dict space in epsilon greedy policy seem not to be supported.

any idea on how to resolve this?

[JaCoderX]

I'm still not sure on how to add support for Dict action space.

The error occur on trying to get the action from greedy_action (greedy_policy is source)
action = tf.compat.v1.where(cond, greedy_action.action, random_action.action)

[ebrevdo]

That line needs to be rewritten as:

action = tf.nest.map_structure(lambda g, r: tf.compat.v1.where(cond, g, r), greedy_action.action, random_action.action)

Report back and let us know if this works. We can patch it on our end.

[JaCoderX]

@ebrevdo,
I tested it on my end and it works well.
thank you :)