print("Env has ", action_size, " actions.") measurement_size = 3 # [Battery, posion, food] timesteps = [1, 2, 4, 8, 16, 32] # For long horizon: [4,8,16,32,64,128] goal_size = measurement_size * len(timesteps) img_rows, img_cols = 84, 84 #KOE: Think this is still correct. # Convert image into Black and white #KOETODO Not quite sure what happens here - I'm making images black/white, so what is the point? img_channels = 3 # KOE: If I want to change this, I have to also edit the frame stacking when forming s_t state_size = (img_rows, img_cols, img_channels) agent = DFPAgent(state_size, measurement_size, action_size, timesteps) agent.model = Networks.dfp_network(state_size, measurement_size, goal_size, action_size, len(timesteps), agent.learning_rate) #x_t = game_state.screen_buffer # 480 x 640 #x_t = preprocessImg(initial_observation, size=(img_rows, img_cols)) #np.save("input_output_examples/initial_obs.npy", initial_observation) #np.save("input_output_examples/preprocessed_obs.npy", x_t) #KOE: Preprocessing to get black and white. #KOE: Not sure what is going on here. 4 images in a row? #s_t = np.stack(([x_t]*4), axis=2) # It becomes 64x64x4 s_t = initial_observation s_t = np.expand_dims(s_t, axis=0) # 1x64x64x4 #np.save("input_output_examples/stacked_obs.npy", s_t)