I am trying to make a simple BC experiment using RL-Coach and a custom Gym environment. Based on the idea of this I am trying to predict the target variable based on f_1, f_2 and f_3. The Gym environment I am using is:
import random
import numpy as np
import pandas as pd
import gym
from gym import spaces
from sklearn.preprocessing import MinMaxScaler
class TestEnvOne(gym.Env):
def __init__(self, max_time):
super(TestEnvOne, self).__init__()
self.max_time = max_time
f_1 = np.sin(np.arange(self.max_time))
f_2 = np.cos(np.arange(self.max_time))
f_3 = np.tan(np.arange(self.max_time))
target = np.roll(f_1, 1) + np.roll(f_2, 2) + np.roll(f_3, 3)
df = pd.DataFrame({'target': target, 'f_1': f_1, 'f_2': f_2, 'f_3': f_3})
list_col = ['target', 'f_1', 'f_2', 'f_3']
df[list_col] = MinMaxScaler().fit_transform(df[list_col])
self.df = df
self.start_step = 0
self.current_step = 0
# Actions
self.action_space = spaces.Box(low=np.array([0]), high=np.array([1]), dtype=np.float32)
self.observation_space = gym.spaces.dict.Dict(
{'measurements': spaces.Box(low=0.0, high=1.1, shape=(3,), dtype=np.float32),
'desired_goal': spaces.Box(low=np.array([0]), high=np.array([1]), dtype=np.float32)
})
self.reward_range = (-1, 1)
def _next_observation(self):
measurements = np.array([
self.df.loc[self.current_step, 'f_1'],
self.df.loc[self.current_step, 'f_2'],
self.df.loc[self.current_step, 'f_3']
])
frame = {'desired_goal': self.df.loc[self.current_step, 'target'].reshape(-1, 1),
'measurements': measurements
}
return frame
def step(self, action):
self.current_step += 1
if self.current_step >= len(self.df.loc[:, 'target'].values):
self.current_step = 0
obs = self._next_observation()
reward = (obs['desired_goal'] - action)[0][0]
done = (self.current_step == self.start_step)
returning_value = {'measurements': obs['measurements'], 'desired_goal': obs['desired_goal']}
all = returning_value, reward, done, {}
return all
def reset(self):
# Set the current step to a random point within the data frame
self.start_step = random.randint(0, len(self.df.loc[:, 'target'].values) - 1)
self.current_step = self.start_step
observation = self._next_observation()
return observation
def render(self, mode='human', close=False):
# Render the environment to the screen
print(f'Step: {self.current_step}')
print(f'Target: {self.df.loc[self.current_step, "target"]}')
def seed(self, seed=None):
self.seed_value = seed
return [seed]
The preset I am using is based on Doom Basic BC as the following:
Running using command line: coach -p presets/PruebaPresetBC.py
I am getting a exception in the improve phase:
Traceback (most recent call last):
File "/home/meteo/coach_env/bin/coach", line 8, in <module>
sys.exit(main())
File "/home/meteo/coach_env/lib/python3.6/site-packages/rl_coach/coach.py", line 777, in main
launcher.launch()
File "/home/meteo/coach_env/lib/python3.6/site-packages/rl_coach/coach.py", line 226, in launch
self.run_graph_manager(graph_manager, args)
File "/home/meteo/coach_env/lib/python3.6/site-packages/rl_coach/coach.py", line 612, in run_graph_manager
self.start_single_threaded(task_parameters, graph_manager, args)
File "/home/meteo/coach_env/lib/python3.6/site-packages/rl_coach/coach.py", line 674, in start_single_threaded
start_graph(graph_manager=graph_manager, task_parameters=task_parameters)
File "/home/meteo/coach_env/lib/python3.6/site-packages/rl_coach/coach.py", line 88, in start_graph
graph_manager.improve()
File "/home/meteo/coach_env/lib/python3.6/site-packages/rl_coach/graph_managers/graph_manager.py", line 547, in improve
self.train_and_act(self.steps_between_evaluation_periods)
File "/home/meteo/coach_env/lib/python3.6/site-packages/rl_coach/graph_managers/graph_manager.py", line 481, in train_and_act
self.act(EnvironmentSteps(1))
File "/home/meteo/coach_env/lib/python3.6/site-packages/rl_coach/graph_managers/graph_manager.py", line 447, in act
result = self.top_level_manager.step(None)
File "/home/meteo/coach_env/lib/python3.6/site-packages/rl_coach/level_manager.py", line 245, in step
action_info = acting_agent.act()
File "/home/meteo/coach_env/lib/python3.6/site-packages/rl_coach/agents/agent.py", line 851, in act
action = self.choose_action(curr_state)
File "/home/meteo/coach_env/lib/python3.6/site-packages/rl_coach/agents/imitation_agent.py", line 43, in choose_action
prediction = self.networks['main'].online_network.predict(self.prepare_batch_for_inference(curr_state, 'main'))
File "/home/meteo/coach_env/lib/python3.6/site-packages/rl_coach/architectures/tensorflow_components/architecture.py", line 547, in predict
output = self.sess.run(outputs, feed_dict)
File "/home/meteo/coach_env/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 950, in run
run_metadata_ptr)
File "/home/meteo/coach_env/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1149, in _run
str(subfeed_t.get_shape())))
ValueError: Cannot feed value of shape (1, 3) for Tensor 'main_level/agent/main/online/network_0/measurements/measurements:0', which has shape '(?, 0)'
I think it is trying to use a measurement (which are three values) into a zero sized network. I don't know where this zero shape comes from. Any idea? Thanks
I am trying to make a simple BC experiment using RL-Coach and a custom Gym environment. Based on the idea of this I am trying to predict the target variable based on f_1, f_2 and f_3. The Gym environment I am using is:
The preset I am using is based on Doom Basic BC as the following:
Running using command line: coach -p presets/PruebaPresetBC.py
I am getting a exception in the improve phase:
I think it is trying to use a measurement (which are three values) into a zero sized network. I don't know where this zero shape comes from. Any idea? Thanks