def log_model_predictions(self, itr, all_logs):
# model predictions
import matplotlib.pyplot as plt
self.fig = plt.figure()
# sample actions
action_sequence = self.agent.actor.sample_action_sequences(num_sequences=1, horizon=10) #20 reacher
action_sequence = action_sequence[0]
# calculate and log model prediction error
mpe, true_states, pred_states = utils.calculate_mean_prediction_error(self.env, action_sequence, self.agent.dyn_models, self.agent.actor.data_statistics)
assert self.params['agent_params']['ob_dim'] == true_states.shape[1] == pred_states.shape[1]
ob_dim = self.params['agent_params']['ob_dim']
ob_dim = 2*int(ob_dim/2.0) ## skip last state for plotting when state dim is odd
# plot the predictions
self.fig.clf()
for i in range(ob_dim):
plt.subplot(ob_dim/2, 2, i+1)
plt.plot(true_states[:,i], 'g')
plt.plot(pred_states[:,i], 'r')
self.fig.suptitle('MPE: ' + str(mpe))
self.fig.savefig(self.params['logdir']+'/itr_'+str(itr)+'_predictions.png', dpi=200, bbox_inches='tight')
# plot all intermediate losses during this iteration
all_losses = np.array([log['Training Loss'] for log in all_logs])
np.save(self.params['logdir']+'/itr_'+str(itr)+'_losses.npy', all_losses)
self.fig.clf()
plt.plot(all_losses)
self.fig.savefig(self.params['logdir']+'/itr_'+str(itr)+'_losses.png', dpi=200, bbox_inches='tight')
def log_model_predictions(self, itr, all_logs):
# model predictions
import matplotlib.pyplot as plt
self.fig = plt.figure()
# sample actions
action_sequence = self.agent.actor.sample_action_sequences(
num_sequences=1, horizon=10
) # 20 reacher
action_sequence = action_sequence[0]
# calculate and log model prediction error
mpe, true_states, pred_states = utils.calculate_mean_prediction_error(
self.env,
action_sequence,
self.agent.dyn_models,
self.agent.actor.data_statistics,
)
assert (
self.params["agent_params"]["ob_dim"]
== true_states.shape[1]
== pred_states.shape[1]
)
ob_dim = self.params["agent_params"]["ob_dim"]
ob_dim = 2 * int(
ob_dim / 2.0
) ## skip last state for plotting when state dim is odd
# plot the predictions
self.fig.clf()
for i in range(ob_dim):
plt.subplot(ob_dim // 2, 2, i + 1)
plt.plot(true_states[:, i], "g")
plt.plot(pred_states[:, i], "r")
self.fig.suptitle("MPE: " + str(mpe))
self.fig.savefig(
self.params["logdir"] + "/itr_" + str(itr) + "_predictions.png",
dpi=200,
bbox_inches="tight",
)
# plot all intermediate losses during this iteration
all_losses = np.array([log["Training Loss"] for log in all_logs])
np.save(self.params["logdir"] + "/itr_" + str(itr) + "_losses.npy", all_losses)
self.fig.clf()
plt.plot(all_losses)
self.fig.savefig(
self.params["logdir"] + "/itr_" + str(itr) + "_losses.png",
dpi=200,
bbox_inches="tight",
)
