# print("Angular RMSE after training:", final_eval["out_p_filt_angular_rmse"])
print("Generating visualization set")
vis_set_fname = 'saved_params/policy_viz_set.npz'
if not os.path.exists(vis_set_fname):
vis_input, vis_output, batch_size = create_policy_vis_set(
data=data,
args=args,
n_mazes=args.n_mazes,
encoding_func=encoding_func,
# maze_indices=[2, 4, 5, 6],
maze_indices=[1, 0, 5, 7],
goal_indices=[
0,
],
# TEMP: for debugging
# maze_indices=[0, ],
# goal_indices=[2, 3, 4, 5, 6, 7, 8],
# x_offset=0.25,
# y_offset=0.25,
)
np.savez(vis_set_fname,
vis_input=vis_input,
vis_output=vis_output,
batch_size=batch_size)
else:
data = np.load(vis_set_fname)
vis_input = data['vis_input']
)
final_eval = sim.evaluate(test_input, {out_p_filt: test_output}, verbose=0)
print("Loss after training:", final_eval["loss"])
print("Angular RMSE after training:", final_eval["out_p_filt_angular_rmse"])
if args.plot_vis_set:
print("Generating visualization set")
vis_input, vis_output, batch_size = create_policy_vis_set(
data=data,
args=args,
n_mazes=args.n_mazes,
encoding_func=encoding_func,
maze_indices=[0, 1, 2, 3],
goal_indices=[0, 1],
)
vis_input = np.tile(vis_input[:, None, :], (1, n_steps, 1))
# vis_output = np.tile(vis_output[:, None, :], (1, n_steps, 1))
print("Running visualization")
# viz_eval = sim.evaluate(test_input, {out_p_filt: test_output}, verbose=0)
n_batches = 4 * 2
for bi in range(n_batches):
viz_eval = sim.predict(vis_input[bi * batch_size:(bi + 1) *
batch_size])