tf.summary.scalar(
"theta_loss",
tf.reduce_mean(params.THETA_SCALING * error[:, :, params.THETA_IND]))
tf.summary.scalar("x_loss_rel",
tf.reduce_mean(error_relative[:, :, params.X_IND]))
tf.summary.scalar("y_loss_rel",
tf.reduce_mean(error_relative[:, :, params.Y_IND]))
tf.summary.scalar("theta_loss_rel",
tf.reduce_mean(error_relative[:, :, params.THETA_IND]))
return loss
if __name__ == "__main__":
# Read the input data
t_chunks, state_chunks, control_chunks, p_chunks = read_data.read_chunks(
params.TRAIN_DIR)
t_chunks_val, state_chunks_val, control_chunks_val, p_chunks_val = read_data.read_chunks(
params.VALIDATION_DIR)
# Compute the average time step
h = np.mean(np.diff(t_chunks[0]))
# Make placeholders
h_ph = tf.placeholder(tf.float32, shape=(), name="h")
training_ph = tf.placeholder(tf.bool, name="training")
state_batch_ph = tf.placeholder(tf.float32,
shape=(params.BATCH_SIZE,
params.STATE_STEPS, params.STATES),
name="state_batch")
control_batch_ph = tf.placeholder(tf.float32,
shape=(params.BATCH_SIZE,
scale=None,
headwidth=3.,
width=0.002,
color=next(colors))
plt.show()
if __name__ == "__main__":
"""
Test basic plotting.
"""
import read_data
# Read the data
t_chunks, state_chunks, control_chunks, p_chunks = read_data.read_chunks(
params.TRAIN_DIR)
# Plot a couple states
states = []
for i in range(4):
states.append(np.expand_dims(state_chunks[i], axis=0))
plot_states(states)
# Plot the states zoomed in
states = []
chunk = np.expand_dims(state_chunks[0], axis=0)
for i in range(3):
states.append(chunk[:, i * params.STATE_STEPS:(i + 1) *
params.STATE_STEPS])
plot_states(states)
feed_dict[self.i_ph] = i
feed_dict[self.control_check_batch_ph] = control_check_batch
feed_dict[self.state_batch_ph] = state_batch
feed_dict[self.control_batch_ph] = control_batch
i, state_batch, control_batch = self.sess.run(
(self.i, self.state_batch, self.control_batch),
feed_dict=feed_dict)
return i, state_batch, control_batch
if __name__ == "__main__":
m = TestDriftModel()
t_chunks, state_chunks, control_chunks, p_chunks = read_data.read_chunks(
params.VALIDATION_DIR)
for i in range(10):
# Make a random input batch
state_batch, control_batch, state_check_batch, control_check_batch = process_data.random_batch(
state_chunks, control_chunks, p_chunks)
# state by simply integrating out the state differences.
loss_base = state_check_batch[:, -1] - (
state_batch[:, -1] + params.CHECK_STEPS *
(state_batch[:, -1] - state_batch[:, -2]))
# Use the learned model
i, state_batch, control_batch = m.compute_f(0, state_batch,
control_batch,
control_check_batch)