def train_path_planning_network():
"""Train Path Planning Network
Trains an Evolino LSTM neural network for long-term path planning for
use in the surgical simulator.
Returns:
A copy of the fully-trained path planning neural network.
"""
# Build up the list of files to use as training set
training_dir = constants.G_TRAINING_DATA_DIR
# Find all data files in the training data directory
training_files = pathutils.list_data_files(training_dir)
# Get the training data and place it into a dataset
training_dataset = None
# Store all training set ratings
ratings = np.array([])
for training_file in training_files:
training_data = datastore.retrieve(training_file)
# Normalize the time input of the data
training_data = pathutils.normalize_time(training_data, t_col=constants.G_TIME_IDX)
# Add this data sample to the training dataset
training_dataset = datastore.list_to_dataset(
training_data[:,constants.G_RNN_INPUT_IDX:constants.G_RNN_INPUT_IDX+constants.G_RNN_NUM_INPUTS],
training_data[:,constants.G_RNN_OUTPUT_IDX:constants.G_RNN_OUTPUT_IDX+constants.G_RNN_NUM_OUTPUTS],
dataset=training_dataset
)
# Store the rating of the data
this_rating = training_data[1:,constants.G_RATING_IDX]
ratings = np.hstack((ratings, this_rating))
# Get the starting point information for testing
output_start_idx = constants.G_RNN_OUTPUT_IDX
output_end_idx = output_start_idx + constants.G_RNN_NUM_OUTPUTS
output_initial_condition = training_data[0,output_start_idx:output_end_idx]
# Generate the time sequence input data for testing
time_steps = constants.G_RNN_GENERATED_TIME_STEPS
t_input = np.linspace(start=0.0, stop=1.0, num=time_steps)
t_input = np.reshape(t_input, (len(t_input), 1))
gate_start_idx = constants.G_GATE_IDX
gate_end_idx = gate_start_idx + constants.G_NUM_GATE_INPUTS
# Pull the gate data from the last training dataset
gate_data = training_data[0:1,gate_start_idx:gate_end_idx]
gate_data = np.tile(gate_data, (time_steps, 1))
# Build up a full ratings matrix
nd_ratings = None
for rating in ratings:
this_rating = rating * np.ones((1, constants.G_RNN_NUM_OUTPUTS))
if nd_ratings is None:
nd_ratings = this_rating
else:
nd_ratings = np.vstack((nd_ratings, this_rating))
# Create network and trainer
print('>>> Building Network...')
net = PathPlanningNetwork()
print('>>> Initializing Trainer...')
trainer = PathPlanningTrainer(
evolino_network=net,
dataset=training_dataset,
nBurstMutationEpochs=10,
importance=nd_ratings
)
# Begin the training iterations
fitness_list = []
max_fitness = None
max_fitness_epoch = None
# Draw the generated path plot
fig = plt.figure(1, facecolor='white')
testing_axis = fig.add_subplot(111, projection='3d')
fig.show()
idx = 0
current_convergence_streak = 0
while True:
print('>>> Training Network (Iteration: %3d)...' % (idx+1))
trainer.train()
# Determine fitness of this network
current_fitness = trainer.evaluation.max_fitness
fitness_list.append(current_fitness)
print('>>> FITNESS: %f' % current_fitness)
# Determine if this is the minimal error network
if max_fitness is None or max_fitness < current_fitness:
# This is the minimum, record it
max_fitness = current_fitness
max_fitness_epoch = idx
# Draw the generated path after training
print('>>> Testing Network...')
generated_output = net.extrapolate(t_input, [output_initial_condition], len(t_input)-1)
generated_output = np.vstack((output_initial_condition, generated_output))
generated_input = np.hstack((t_input, gate_data))
# Smash together the input and output
generated_data = np.hstack((generated_input, generated_output))
print('>>> Drawing Generated Path...')
pathutils.display_path(testing_axis, generated_data, title='Generated Testing Path')
plt.draw()
if current_fitness > constants.G_RNN_CONVERGENCE_THRESHOLD:
# We've encountered a fitness higher than threshold
current_convergence_streak += 1
else:
# Streak ended. Reset the streak counter
current_convergence_streak = 0
if current_convergence_streak == constants.G_RNN_REQUIRED_CONVERGENCE_STREAK:
print('>>> Convergence Achieved: %d Iterations' % idx)
break
elif idx == constants.G_RNN_MAX_ITERATIONS - 1:
print('>>> Reached maximum iterations (%d)' % constants.G_RNN_MAX_ITERATIONS)
break
idx += 1
# Draw the iteration fitness plot
plt.figure(facecolor='white')
plt.cla()
plt.title('Fitness of RNN over %d Iterations' % (idx-1))
plt.xlabel('Training Iterations')
plt.ylabel('Fitness')
plt.grid(True)
plt.plot(range(len(fitness_list)), fitness_list, 'r-')
plt.annotate('local max', xy=(max_fitness_epoch, fitness_list[max_fitness_epoch]),
xytext=(max_fitness_epoch, fitness_list[max_fitness_epoch]+0.01),
arrowprops=dict(facecolor='black', shrink=0.05))
plt.show()
# Return a full copy of the trained neural network
return copy.deepcopy(net)
def train_lt_network():
"""Train Long-Term Network
Trains an Evolino LSTM neural network for long-term path planning for
use in the surgical simulator.
Returns:
A copy of the fully-trained path planning neural network.
"""
# Build up the list of files to use as training set
filepath = '../../results/'
training_filenames = [
'sample1.dat',
'sample2.dat',
'sample3.dat',
'sample4.dat',
]
testing_filename = 'sample5.dat'
# Get the training data and place it into a dataset
training_dataset = None
# Store all training set ratings
ratings = np.array([])
for training_filename in training_filenames:
training_file = filepath + training_filename
training_data = datastore.retrieve(training_file)
# Normalize the time input of the data
training_data = pathutils.normalize_time(training_data,
t_col=constants.G_TIME_IDX)
# Add this data sample to the training dataset
training_dataset = datastore.list_to_dataset(
training_data[:,
constants.G_LT_INPUT_IDX:constants.G_LT_INPUT_IDX +
constants.G_LT_NUM_INPUTS],
training_data[:,
constants.G_LT_OUTPUT_IDX:constants.G_LT_OUTPUT_IDX +
constants.G_LT_NUM_OUTPUTS],
dataset=training_dataset)
# Store the rating of the data
this_rating = training_data[1:, constants.G_RATING_IDX]
ratings = np.hstack((ratings, this_rating))
# Get testing data
testing_file = filepath + testing_filename
testing_data = datastore.retrieve(testing_file)
testing_data = pathutils.normalize_time(testing_data,
t_col=constants.G_TIME_IDX)
# Store the testing data in a datastore object
testing_dataset = datastore.list_to_dataset(
testing_data[:, constants.G_LT_INPUT_IDX:constants.G_LT_INPUT_IDX +
constants.G_LT_NUM_INPUTS],
testing_data[:, constants.G_LT_OUTPUT_IDX:constants.G_LT_OUTPUT_IDX +
constants.G_LT_NUM_OUTPUTS],
dataset=None)
# Build up a full ratings matrix
nd_ratings = None
for rating in ratings:
this_rating = rating * np.ones((1, constants.G_LT_NUM_OUTPUTS))
if nd_ratings is None:
nd_ratings = this_rating
else:
nd_ratings = np.vstack((nd_ratings, this_rating))
# Create network and trainer
print('>>> Building Network...')
net = network.LongTermPlanningNetwork()
print('>>> Initializing Trainer...')
trainer = network.LongTermPlanningTrainer(evolino_network=net,
dataset=training_dataset,
nBurstMutationEpochs=20,
importance=nd_ratings)
# Begin the training iterations
fitness_list = []
max_fitness = None
max_fitness_epoch = None
# Draw the generated path plot
fig = plt.figure(1, facecolor='white')
testing_axis = fig.add_subplot(111, projection='3d')
fig.show()
# Define paramters for convergence
CONVERGENCE_THRESHOLD = -0.000005
REQUIRED_CONVERGENCE_STREAK = 20
idx = 0
current_convergence_streak = 0
while True:
print('>>> Training Network (Iteration: %3d)...' % (idx + 1))
trainer.train()
# Determine fitness of this network
current_fitness = trainer.evaluation.max_fitness
fitness_list.append(current_fitness)
print('>>> FITNESS: %f' % current_fitness)
# Determine if this is the minimal error network
if max_fitness is None or max_fitness < current_fitness:
# This is the minimum, record it
max_fitness = current_fitness
max_fitness_epoch = idx
# Draw the generated path after training
print('>>> Testing Network...')
washout_ratio = 1.0 / len(testing_data)
_, generated_output, _ = net.calculateOutput(
testing_dataset, washout_ratio=washout_ratio)
generated_input = testing_data[:len(generated_output), :constants.
G_TOTAL_NUM_INPUTS]
# Smash together the input and output
generated_data = np.hstack((generated_input, generated_output))
print('>>> Drawing Generated Path...')
pathutils.display_path(testing_axis,
generated_data,
testing_data,
title='Generated Testing Path')
plt.draw()
if current_fitness > CONVERGENCE_THRESHOLD:
# We've encountered a fitness higher than threshold
current_convergence_streak += 1
else:
# Streak ended. Reset the streak counter
current_convergence_streak = 0
if current_convergence_streak == REQUIRED_CONVERGENCE_STREAK:
print('>>> Convergence Achieved: %d Iterations' % idx)
break
idx += 1
# Draw the iteration fitness plot
plt.figure(facecolor='white')
plt.cla()
plt.title('Fitness of RNN over %d Iterations' % (idx - 1))
plt.xlabel('Training Iterations')
plt.ylabel('Fitness')
plt.grid(True)
plt.plot(range(len(fitness_list)), fitness_list, 'r-')
plt.annotate('local max',
xy=(max_fitness_epoch, fitness_list[max_fitness_epoch]),
xytext=(max_fitness_epoch,
fitness_list[max_fitness_epoch] + 0.01),
arrowprops=dict(facecolor='black', shrink=0.05))
plt.show()
# Return a full copy of the trained neural network
return copy.deepcopy(net)
def main():
path_file = '../../neuralsim/generated.dat' #'../../results/sample5.dat'
path = datastore.retrieve(path_file)
# A list of the the segments of the optimized path
segments = pathutils._detect_segments(path)
# The new path generated by original path and corrective algorithm
new_path = None
x_path_offset = np.array([0.0, 0.0, 0.0]) # [m]
v_curr = np.array([0.0, 0.0, 0.0]) # [m/s]
a_max = constants.G_MAX_ACCEL # [m/s^2]
for i, _ in enumerate(path):
if i == len(path) - 1:
continue
# Detect current segment
seg_idx = 0
for j in range(len(segments)):
if i <= segments[j]:
seg_idx = j
break
# Get current time and position
t_curr = pathutils.get_path_time(path, i) * t_total
t_next = pathutils.get_path_time(path, i + 1) * t_total
dt = (t_next - t_curr)
x_curr = pathutils.get_path_tooltip_pos(path, i) + x_path_offset
x_next = pathutils.get_path_tooltip_pos(path, i + 1) + x_path_offset
# Get the expected gate position at this timestep
x_gate_expected = pathutils.get_path_gate_pos(path, segments[seg_idx],
seg_idx)
# Get current gate position
x_gate_actual = generate_gate_pos(t_curr, path, seg_idx)
dx_gate = x_gate_actual - (x_gate_expected + x_path_offset)
# Calculate the new position with positional change from target to gate
x_new = x_next + dx_gate
# Calculate the new velocity
v_new = (x_new - x_curr) / dt
# Calculate the new acceleration
a_new = (v_new - v_curr) / dt
# Calculate the acceleration vector norm
a_new_norm = np.linalg.norm(a_new)
# Limit the norm vector
a_new_norm_clipped = np.clip(a_new_norm, -a_max, a_max)
# Determine the ratio of the clipped norm
if a_new_norm != 0:
ratio_unclipped = a_new_norm_clipped / a_new_norm
else:
ratio_unclipped = 0.0
# Scale the acceleration vector by this ratio
a_new = a_new * ratio_unclipped
# Calculate the new change in velocity
dv_new = a_new * dt
v_new = v_curr + dv_new
# Calculate the new change in position
dx_new = v_new * dt
x_new = x_curr + dx_new
# Store the next movement for later
if new_path is None:
new_path = x_new
else:
new_path = np.vstack((new_path, x_new))
# Store this velocity for the next time step
v_curr = v_new
# Recalculate the current offset
x_path_offset += x_new - x_next
# Plot the inputted path
fig = plt.figure(facecolor='white')
axis = fig.gca(projection='3d')
pos_start_idx = constants.G_POS_IDX
pos_end_idx = pos_start_idx + constants.G_NUM_POS_DIMS
full_path = path[:-1].copy()
full_path[:, pos_start_idx:pos_end_idx] = new_path
pathutils.display_path(axis, full_path, title='Path')
plt.show()
return
def train_lt_network():
"""Train Long-Term Network
Trains an Evolino LSTM neural network for long-term path planning for
use in the surgical simulator.
Returns:
A copy of the fully-trained path planning neural network.
"""
# Build up the list of files to use as training set
filepath = "../../results/"
training_filenames = ["sample1.dat", "sample2.dat", "sample3.dat", "sample4.dat"]
testing_filename = "sample5.dat"
# Get the training data and place it into a dataset
training_dataset = None
# Store all training set ratings
ratings = np.array([])
for training_filename in training_filenames:
training_file = filepath + training_filename
training_data = datastore.retrieve(training_file)
# Normalize the time input of the data
training_data = pathutils.normalize_time(training_data, t_col=constants.G_TIME_IDX)
# Add this data sample to the training dataset
training_dataset = datastore.list_to_dataset(
training_data[:, constants.G_LT_INPUT_IDX : constants.G_LT_INPUT_IDX + constants.G_LT_NUM_INPUTS],
training_data[:, constants.G_LT_OUTPUT_IDX : constants.G_LT_OUTPUT_IDX + constants.G_LT_NUM_OUTPUTS],
dataset=training_dataset,
)
# Store the rating of the data
this_rating = training_data[1:, constants.G_RATING_IDX]
ratings = np.hstack((ratings, this_rating))
# Get testing data
testing_file = filepath + testing_filename
testing_data = datastore.retrieve(testing_file)
testing_data = pathutils.normalize_time(testing_data, t_col=constants.G_TIME_IDX)
# Store the testing data in a datastore object
testing_dataset = datastore.list_to_dataset(
testing_data[:, constants.G_LT_INPUT_IDX : constants.G_LT_INPUT_IDX + constants.G_LT_NUM_INPUTS],
testing_data[:, constants.G_LT_OUTPUT_IDX : constants.G_LT_OUTPUT_IDX + constants.G_LT_NUM_OUTPUTS],
dataset=None,
)
# Build up a full ratings matrix
nd_ratings = None
for rating in ratings:
this_rating = rating * np.ones((1, constants.G_LT_NUM_OUTPUTS))
if nd_ratings is None:
nd_ratings = this_rating
else:
nd_ratings = np.vstack((nd_ratings, this_rating))
# Create network and trainer
print(">>> Building Network...")
net = network.LongTermPlanningNetwork()
print(">>> Initializing Trainer...")
trainer = network.LongTermPlanningTrainer(
evolino_network=net, dataset=training_dataset, nBurstMutationEpochs=20, importance=nd_ratings
)
# Begin the training iterations
fitness_list = []
max_fitness = None
max_fitness_epoch = None
# Draw the generated path plot
fig = plt.figure(1, facecolor="white")
testing_axis = fig.add_subplot(111, projection="3d")
fig.show()
# Define paramters for convergence
CONVERGENCE_THRESHOLD = -0.000005
REQUIRED_CONVERGENCE_STREAK = 20
idx = 0
current_convergence_streak = 0
while True:
print(">>> Training Network (Iteration: %3d)..." % (idx + 1))
trainer.train()
# Determine fitness of this network
current_fitness = trainer.evaluation.max_fitness
fitness_list.append(current_fitness)
print(">>> FITNESS: %f" % current_fitness)
# Determine if this is the minimal error network
if max_fitness is None or max_fitness < current_fitness:
# This is the minimum, record it
max_fitness = current_fitness
max_fitness_epoch = idx
# Draw the generated path after training
print(">>> Testing Network...")
washout_ratio = 1.0 / len(testing_data)
_, generated_output, _ = net.calculateOutput(testing_dataset, washout_ratio=washout_ratio)
generated_input = testing_data[: len(generated_output), : constants.G_TOTAL_NUM_INPUTS]
# Smash together the input and output
generated_data = np.hstack((generated_input, generated_output))
print(">>> Drawing Generated Path...")
pathutils.display_path(testing_axis, generated_data, testing_data, title="Generated Testing Path")
plt.draw()
if current_fitness > CONVERGENCE_THRESHOLD:
# We've encountered a fitness higher than threshold
current_convergence_streak += 1
else:
# Streak ended. Reset the streak counter
current_convergence_streak = 0
if current_convergence_streak == REQUIRED_CONVERGENCE_STREAK:
print(">>> Convergence Achieved: %d Iterations" % idx)
break
idx += 1
# Draw the iteration fitness plot
plt.figure(facecolor="white")
plt.cla()
plt.title("Fitness of RNN over %d Iterations" % (idx - 1))
plt.xlabel("Training Iterations")
plt.ylabel("Fitness")
plt.grid(True)
plt.plot(range(len(fitness_list)), fitness_list, "r-")
plt.annotate(
"local max",
xy=(max_fitness_epoch, fitness_list[max_fitness_epoch]),
xytext=(max_fitness_epoch, fitness_list[max_fitness_epoch] + 0.01),
arrowprops=dict(facecolor="black", shrink=0.05),
)
plt.show()
# Return a full copy of the trained neural network
return copy.deepcopy(net)
def main():
path_file = '../../neuralsim/generated.dat'#'../../results/sample5.dat'
path = datastore.retrieve(path_file)
# A list of the the segments of the optimized path
segments = pathutils._detect_segments(path)
# The new path generated by original path and corrective algorithm
new_path = None
x_path_offset = np.array([0.0, 0.0, 0.0]) # [m]
v_curr = np.array([0.0, 0.0, 0.0]) # [m/s]
a_max = constants.G_MAX_ACCEL # [m/s^2]
for i, _ in enumerate(path):
if i == len(path) - 1:
continue
# Detect current segment
seg_idx = 0
for j in range(len(segments)):
if i <= segments[j]:
seg_idx = j
break
# Get current time and position
t_curr = pathutils.get_path_time(path, i) * t_total
t_next = pathutils.get_path_time(path, i+1) * t_total
dt = (t_next - t_curr)
x_curr = pathutils.get_path_tooltip_pos(path, i) + x_path_offset
x_next = pathutils.get_path_tooltip_pos(path, i+1) + x_path_offset
# Get the expected gate position at this timestep
x_gate_expected = pathutils.get_path_gate_pos(path, segments[seg_idx], seg_idx)
# Get current gate position
x_gate_actual = generate_gate_pos(t_curr, path, seg_idx)
dx_gate = x_gate_actual - (x_gate_expected + x_path_offset)
# Calculate the new position with positional change from target to gate
x_new = x_next + dx_gate
# Calculate the new velocity
v_new = (x_new - x_curr) / dt
# Calculate the new acceleration
a_new = (v_new - v_curr) / dt
# Calculate the acceleration vector norm
a_new_norm = np.linalg.norm(a_new)
# Limit the norm vector
a_new_norm_clipped = np.clip(a_new_norm, -a_max, a_max)
# Determine the ratio of the clipped norm
if a_new_norm != 0:
ratio_unclipped = a_new_norm_clipped / a_new_norm
else:
ratio_unclipped = 0.0
# Scale the acceleration vector by this ratio
a_new = a_new * ratio_unclipped
# Calculate the new change in velocity
dv_new = a_new * dt
v_new = v_curr + dv_new
# Calculate the new change in position
dx_new = v_new * dt
x_new = x_curr + dx_new
# Store the next movement for later
if new_path is None:
new_path = x_new
else:
new_path = np.vstack((new_path, x_new))
# Store this velocity for the next time step
v_curr = v_new
# Recalculate the current offset
x_path_offset += x_new - x_next
# Plot the inputted path
fig = plt.figure(facecolor='white')
axis = fig.gca(projection='3d')
pos_start_idx = constants.G_POS_IDX
pos_end_idx = pos_start_idx + constants.G_NUM_POS_DIMS
full_path = path[:-1].copy()
full_path[:,pos_start_idx:pos_end_idx] = new_path
pathutils.display_path(axis, full_path, title='Path')
plt.show()
return
