Issue in DDPG setup #7
Description
Hi Sri,
I have been trying run training with DDPG set as the algorithm. Keras has been giving me this error:
Input 0 is incompatible with layer conv2d_1: expected ndim=4, found ndim=3
I think there is an issue in the input data provided to the Neural Network. In the top-level file for this algorithm (algorithms/continuous/ddpg/ddpg_airsim.py) we obtain depth, grey image, velocity and position data dimensions, lines 53-60. We attempt this by indexing into env.state(). This state function is defined in gym_airsim/envs/AirGym.py on line 195, and we can see that specifically for DDPG, depth data is directly returned, rather than an array of data whose indices are the aforementioned parameters; indexing into the depth data results in all the parameters storing the dimensions of one row of the depth data (I believe it is [1, 256]). On checking the version history of the file I see that in your original git commit you returned all these parameters, before changing it to what is presently there; I wondered if you had made changes to return all data as one array, and then missed an update somewhere, so I thought I'd check with you about this.
As an aside, I'd like to mention that there was an issue during the environment initialization in the if statement while setting action_space, should be "DDPG" instead of "DDPG ". Thought I'd mention it here for record keeping.
Thanks,
Ashwin