def learn_weights(self, mode='greedy'):
"""
Runs iterations of generating samples with current weights and training
new weight vectors based on the collected samples.
- mode can be ['greedy','rl_half','rl_regression','rl_lspi']
"""
print("Beginning to learn regression weights")
self.tt.tic('learn_weights:all')
# Need to have weights set here to collect samples, so let's set
# to manual_1 to get a reasonable execution trace.
self.weights = self.load_weights(weights_mode='manual_1', force=True)
if comm_rank == 0:
print("Initial weights:")
np.set_printoptions(precision=2, suppress=True, linewidth=160)
print self.get_reshaped_weights()
# Loop until max_iterations or the error is below threshold
error = threshold = 0.001
max_iterations = 12
# early iterations should explore more than later iterations
# so do an exponential fall-off, halving every few iterations
epsilons = 0.6 * np.exp2(-np.arange(0, max_iterations + 1) / 2.)
# Collect samples (parallelized)
num_samples = 350 # actually this refers to images
dets, clses, all_samples = self.run_on_dataset(False,
num_samples,
epsilon=epsilons[0])
for i in range(0, max_iterations):
# do regression with cross-validated parameters (parallelized)
weights = None
if comm_rank == 0:
weights, error = self.regress(all_samples, mode)
self.weights = weights
try:
# write image of the weights
img_filename = opjoin(
config.get_dp_weights_images_dirname(self),
'iter_%d.png' % i)
self.write_weights_image(img_filename)
# write image of the average feature
all_features = self.construct_X_from_samples(all_samples)
avg_feature = np.mean(all_features,
0).reshape(len(self.actions),
BeliefState.num_features)
img_filename = opjoin(
config.get_dp_features_images_dirname(self),
'iter_%d.png' % i)
self.write_feature_image(avg_feature, img_filename)
except:
print("Couldn't plot, no big deal.")
print(
"""
After iteration %d (epsilon %.2f), we've trained on %d samples and
the weights and error are:""" % (i, epsilons[i], len(all_samples)))
np.set_printoptions(precision=2, suppress=True, linewidth=160)
print self.get_reshaped_weights()
print error
# after the first iteration, check if the error is small
if i > 0 and error <= threshold:
break
print(
"Now collecting more samples with the updated weights...")
safebarrier(comm)
weights = comm.bcast(weights, root=0)
self.weights = weights
new_dets, new_clses, new_samples = self.run_on_dataset(
False, num_samples, epsilon=epsilons[i + 1])
if comm_rank == 0:
# We either collect only unique samples or all samples
only_unique_samples = False
if only_unique_samples:
self.tt.tic('learn_weights:unique_samples')
for sample in new_samples:
if not (sample in all_samples):
all_samples.append(sample)
print(
"Only adding unique samples to all_samples took %.3f s"
% self.tt.qtoc('learn_weights:unique_samples'))
else:
all_samples += new_samples
if comm_rank == 0:
print("Done training regression weights! Took %.3f s total" %
self.tt.qtoc('learn_weights:all'))
# Save the weights
filename = config.get_dp_weights_filename(self)
np.savetxt(filename, self.weights, fmt='%.6f')
safebarrier(comm)
weights = comm.bcast(weights, root=0)
self.weights = weights
return weights
def learn_weights(self,mode='greedy'):
"""
Runs iterations of generating samples with current weights and training
new weight vectors based on the collected samples.
- mode can be ['greedy','rl_half','rl_regression','rl_lspi']
"""
print("Beginning to learn regression weights")
self.tt.tic('learn_weights:all')
# Need to have weights set here to collect samples, so let's set
# to manual_1 to get a reasonable execution trace.
self.weights = self.load_weights(weights_mode='manual_1',force=True)
if comm_rank==0:
print("Initial weights:")
np.set_printoptions(precision=2,suppress=True,linewidth=160)
print self.get_reshaped_weights()
# Loop until max_iterations or the error is below threshold
error = threshold = 0.001
max_iterations = 12
# early iterations should explore more than later iterations
# so do an exponential fall-off, halving every few iterations
epsilons = 0.6*np.exp2(-np.arange(0,max_iterations+1)/2.)
# Collect samples (parallelized)
num_samples = 350 # actually this refers to images
dets,clses,all_samples = self.run_on_dataset(False,num_samples,epsilon=epsilons[0])
for i in range(0,max_iterations):
# do regression with cross-validated parameters (parallelized)
weights = None
if comm_rank==0:
weights, error = self.regress(all_samples, mode)
self.weights = weights
try:
# write image of the weights
img_filename = opjoin(
config.get_dp_weights_images_dirname(self),'iter_%d.png'%i)
self.write_weights_image(img_filename)
# write image of the average feature
all_features = self.construct_X_from_samples(all_samples)
avg_feature = np.mean(all_features,0).reshape(
len(self.actions),BeliefState.num_features)
img_filename = opjoin(
config.get_dp_features_images_dirname(self),'iter_%d.png'%i)
self.write_feature_image(avg_feature, img_filename)
except:
print("Couldn't plot, no big deal.")
print("""
After iteration %d (epsilon %.2f), we've trained on %d samples and
the weights and error are:"""%(i,epsilons[i],len(all_samples)))
np.set_printoptions(precision=2,suppress=True,linewidth=160)
print self.get_reshaped_weights()
print error
# after the first iteration, check if the error is small
if i>0 and error<=threshold:
break
print("Now collecting more samples with the updated weights...")
safebarrier(comm)
weights = comm.bcast(weights,root=0)
self.weights = weights
new_dets,new_clses,new_samples = self.run_on_dataset(False,num_samples,epsilon=epsilons[i+1])
if comm_rank==0:
# We either collect only unique samples or all samples
only_unique_samples = False
if only_unique_samples:
self.tt.tic('learn_weights:unique_samples')
for sample in new_samples:
if not (sample in all_samples):
all_samples.append(sample)
print("Only adding unique samples to all_samples took %.3f s"%self.tt.qtoc('learn_weights:unique_samples'))
else:
all_samples += new_samples
if comm_rank==0:
print("Done training regression weights! Took %.3f s total"%
self.tt.qtoc('learn_weights:all'))
# Save the weights
filename = config.get_dp_weights_filename(self)
np.savetxt(filename, self.weights, fmt='%.6f')
safebarrier(comm)
weights = comm.bcast(weights,root=0)
self.weights = weights
return weights
def load_weights(self, weights_mode=None, force=False):
"""
Set self.weights to weights loaded from file or constructed from scratch
according to weights_mode and self.weights_dataset_name.
"""
if not weights_mode:
weights_mode = self.weights_mode
filename = config.get_dp_weights_filename(self)
if not force and opexists(filename):
self.weights = np.loadtxt(filename)
return self.weights
# Construct the weight vector by concatenating per-action vectors
# The features are [P(C) P(not C) H(C) 1]
weights = np.zeros((len(self.actions), BeliefState.num_features))
num_classes = len(self.dataset.classes)
# OPTION 1: the manual weights correspond to the class of the action
if weights_mode == 'manual_1':
# Set weight of 1 on the P(C) feature
weights[:, 0] = 1
# If gist mode, also set a 1 to 1-t/T, which starts at as 1.
if self.actions[0].name == 'gist':
weights[0, -1] = 1
weights = weights.flatten()
elif weights_mode == 'manual_1_fastinf':
# If gist mode, also set a 1 to 1-t/T, which starts at as 1.
start = 0
if self.actions[0].name == 'gist':
start = 1
weights[0, -1] = 1
# Set weight of 1 on the P(C_i|O) features
weights[start:start + num_classes,
1:1 + num_classes] = np.eye(num_classes)
weights = weights.flatten()
elif weights_mode in ['manual_2', 'manual_3']:
# Figure out the statistics if they aren't there
if 'naive_ap|present' not in self.actions[0].obj.config or \
'actual_ap|present' not in self.actions[0].obj.config or \
'actual_ap|absent' not in self.actions[0].obj.config:
det_configs = self.output_det_statistics()
self.test_actions = self.init_actions()
self.actions = self.test_actions
# OPTION 2: the manual weights are [naive_ap|present 0 0 0]
if weights_mode == 'manual_2':
for action in self.actions:
print action.obj.config
weights[:, 0] = [
action.obj.config['naive_ap|present']
for action in self.actions
]
# OPTION 3: the manual weights are [actual_ap|present actual_ap|absent 0 0]
elif weights_mode == 'manual_3':
weights[:, 0] = [
action.obj.config['actual_ap|present']
for action in self.actions
]
# TODO: get rid of this
# weights[:,1] = [action.obj.config['actual_ap|absent'] for action in self.actions]
else:
None # impossible
weights = weights.flatten()
elif weights_mode in ['greedy', 'rl_half', 'rl_regression', 'rl_lspi']:
weights = self.learn_weights(weights_mode)
print("DONE LEARNING WEIGHTS PLZ RUN ON TEST GODDAMNIT!!!!")
sys.exit()
else:
raise ValueError("unsupported weights_mode %s" % weights_mode)
self.weights = weights
return weights
def load_weights(self,weights_mode=None,force=False):
"""
Set self.weights to weights loaded from file or constructed from scratch
according to weights_mode and self.weights_dataset_name.
"""
if not weights_mode:
weights_mode = self.weights_mode
filename = config.get_dp_weights_filename(self)
if not force and opexists(filename):
self.weights = np.loadtxt(filename)
return self.weights
# Construct the weight vector by concatenating per-action vectors
# The features are [P(C) P(not C) H(C) 1]
weights = np.zeros((len(self.actions), BeliefState.num_features))
num_classes = len(self.dataset.classes)
# OPTION 1: the manual weights correspond to the class of the action
if weights_mode == 'manual_1':
# Set weight of 1 on the P(C) feature
weights[:,0] = 1
# If gist mode, also set a 1 to 1-t/T, which starts at as 1.
if self.actions[0].name=='gist':
weights[0,-1] = 1
weights = weights.flatten()
elif weights_mode == 'manual_1_fastinf':
# If gist mode, also set a 1 to 1-t/T, which starts at as 1.
start = 0
if self.actions[0].name=='gist':
start = 1
weights[0,-1] = 1
# Set weight of 1 on the P(C_i|O) features
weights[start:start+num_classes,1:1+num_classes] = np.eye(num_classes)
weights = weights.flatten()
elif weights_mode in ['manual_2','manual_3']:
# Figure out the statistics if they aren't there
if 'naive_ap|present' not in self.actions[0].obj.config or \
'actual_ap|present' not in self.actions[0].obj.config or \
'actual_ap|absent' not in self.actions[0].obj.config:
det_configs = self.output_det_statistics()
self.test_actions = self.init_actions()
self.actions = self.test_actions
# OPTION 2: the manual weights are [naive_ap|present 0 0 0]
if weights_mode == 'manual_2':
for action in self.actions:
print action.obj.config
weights[:,0] = [action.obj.config['naive_ap|present'] for action in self.actions]
# OPTION 3: the manual weights are [actual_ap|present actual_ap|absent 0 0]
elif weights_mode == 'manual_3':
weights[:,0] = [action.obj.config['actual_ap|present'] for action in self.actions]
# TODO: get rid of this
# weights[:,1] = [action.obj.config['actual_ap|absent'] for action in self.actions]
else:
None # impossible
weights = weights.flatten()
elif weights_mode in ['greedy','rl_half','rl_regression','rl_lspi']:
weights = self.learn_weights(weights_mode)
print("DONE LEARNING WEIGHTS PLZ RUN ON TEST GODDAMNIT!!!!")
sys.exit()
else:
raise ValueError("unsupported weights_mode %s"%weights_mode)
self.weights = weights
return weights