def compute_probability_vector(self, bolt_obj):
if len(self.bolt_object_list) == 0:
self.movement_time = time.time()
# First object - initialize variables to store
# Also clears out the vectors for new run
if bolt_obj.state == bolt_obj.TAP:
# Store results as they come in
self.adjective_vectors_static = dict()
self.adjective_vectors_dynamic = dict()
self.all_motion_results = dict()
self.mkl_results = dict()
# Store dictionary of strings
self.state_string = {bolt_obj.DISABLED:'disabled',
bolt_obj.THERMAL_HOLD:'thermal_hold',
bolt_obj.SLIDE:'slide',
bolt_obj.SQUEEZE:'squeeze',
bolt_obj.TAP:'tap',
bolt_obj.DONE:'done',
bolt_obj.SLIDE_FAST:'slide_fast',
bolt_obj.CENTER_GRIPPER:'center_gripper'
}
# store dictionary for detailed states
self.detailed_states = {bolt_obj.DISABLED:'MOVE_ARM_START_POSITION',
bolt_obj.SQUEEZE:'SQUEEZE_SET_PRESSURE_SLOW',
bolt_obj.THERMAL_HOLD:'HOLD_FOR_10_SECONDS',
bolt_obj.SLIDE:'SLIDE_5CM',
bolt_obj.SLIDE_FAST:'MOVE_DOWN_5CM'
}
if (len(self.bolt_object_list) < 5):
print len(self.bolt_object_list)
self.bolt_object_list.append(bolt_obj)
return
else:
self.bolt_object_list.append(bolt_obj)
motion_end_time = time.time()
print("Elapsed time was %g seconds" % (motion_end_time - self.movement_time))
start_time = time.time()
for bolt_obj_v in self.bolt_object_list:
# Get the current motion
current_motion = self.state_string[bolt_obj_v.state]
# Check if state passed in should be processed
if bolt_obj_v.state not in self.detailed_states:
print "Skipping state: %s" % current_motion
continue
else:
# Get detailed state if exists
current_detailed_state = self.detailed_states[bolt_obj_v.state]
# Check if the state is a the disabled state
if bolt_obj_v.state == bolt_obj_v.DISABLED:
self.norm_bolt_obj = upenn_features.pull_detailed_state(bolt_obj_v,current_detailed_state)
continue
else:
cur_bolt_object = upenn_features.pull_detailed_state(bolt_obj_v, current_detailed_state)
# Checks to make sure that the norm obj has been created
if not self.norm_bolt_obj:
print "Warning: there is no normalization data"
#Start storing the bolt objects
# Build the static features
static_feature_object, self.static_features_array = upenn_features.extract_static_features(cur_bolt_object, self.norm_bolt_obj)
static_feats = upenn_features.createFeatureVector(static_feature_object)
# Store all feature vectors into one large vector array
# Technically we don't need to store the static
# features by adjective - but it's cleaner this way
for classifier in self.all_classifiers:
adj = classifier['adjective']
# need to add pkl to the name b/c the hmm chain dictionary
# takes adjective.pkl as its key
hmm_adj_name = '.'.join((adj,'pkl'))
# Initialize the dictionary with adjective
if adj in self.adjective_vectors_static:
pass
else:
self.adjective_vectors_static[adj] = list()
self.adjective_vectors_dynamic[adj] = list()
# Pull out chain associated with adjective
# ordering of sensors - [pac, pdc, electrodes, tac]
sensor_hmm = ['pac', 'pdc', 'electrodes', 'tac']
hmm_features_phase = [];
for sensor in sensor_hmm:
hmm_chain = self.hmm_chains[hmm_adj_name].chains[current_detailed_state][sensor]
hmm_data = cur_bolt_object.hmm[sensor]
# fill in dyanmic features
hmm_features_phase.append(hmm_chain.score(hmm_data))
# Store the feature vector by adjective away
self.adjective_vectors_static[adj].append(static_feats)
self.adjective_vectors_dynamic[adj].append(hmm_features_phase)
# Check if all motions have been performed
# If so - feed into classifier
if len(self.adjective_vectors_dynamic[adj]) == 4:
print 'All motions received! Computing adjective scores'
for classifier in self.all_classifiers:
# Pull out which adjective we are working on
adj = classifier['adjective']
# Load training vectors for kernel creation
train_dynamic = utilities.get_all_train_test_features(adj, self.training_dynamic, train=True)
train_static = utilities.get_all_train_test_features(adj, self.training_static, train=True)
# Scale the training data
d_scaler = preprocessing.StandardScaler().fit(train_dynamic[0])
s_scaler = preprocessing.StandardScaler().fit(train_static[0])
train_dynamic_scaled = d_scaler.transform(train_dynamic[0])
train_static_scaled = s_scaler.transform(train_static[0])
# Pull out the feature vectors for static/dynamic
all_static_feats = np.hstack(self.adjective_vectors_static[adj])
all_dynamic_feats = np.hstack(self.adjective_vectors_dynamic[adj])
# Normalize the features using scaler
all_static_feats_scaled = classifier['static_scaler'].transform(all_static_feats)
all_dynamic_feats_scaled = classifier['dynamic_scaler'].transform(all_dynamic_feats)
# Create kernels for both static and dynamic
static_kernel = self.linear_kernel_test(all_static_feats_scaled, train_static_scaled, 1)
dynamic_kernel = self.linear_kernel_test(all_dynamic_feats_scaled, train_dynamic_scaled, 1)
# Merge the two kernels
alpha = classifier['alpha']
merged_kernel = (alpha)*static_kernel + (1-alpha)*dynamic_kernel
# Predict adjective with computed kernel
clf = classifier['classifier']
self.mkl_results[adj] = clf.predict(merged_kernel)
# Store off the adjectives that returned true
adjectives_found = []
for adj in self.mkl_results:
if self.mkl_results[adj] == 1:
adjectives_found.append(Adj(adj))
hardcode_adj = [Adj('hard'), Adj('metallic'), Adj('cool')]
publish_string = AdjList()
#publish_string = adjectives_found
publish_string = hardcode_adj;
# Print and publish results!
print "Results from MKL classification"
#print self.mkl_results
print str(adjectives_found)
end_time = time.time()
print("Elapsed time was %g seconds" % (end_time - start_time))
self.adjectives_pub.publish(publish_string)
def compute_probability_vector(self, bolt_obj):
# First object - initialize variables to store
# Also clears out the vectors for new run
if bolt_obj.state == bolt_obj.TAP:
# Store results as they come in
self.adjective_vectors_static = dict()
self.adjective_vectors_dynamic = dict()
self.all_motion_results = dict()
self.mkl_results = dict()
# Store dictionary of strings
self.state_string = {
bolt_obj.DISABLED: 'disabled',
bolt_obj.THERMAL_HOLD: 'thermal_hold',
bolt_obj.SLIDE: 'slide',
bolt_obj.SQUEEZE: 'squeeze',
bolt_obj.TAP: 'tap',
bolt_obj.DONE: 'done',
bolt_obj.SLIDE_FAST: 'slide_fast',
bolt_obj.CENTER_GRIPPER: 'center_gripper'
}
# store dictionary for detailed states
self.detailed_states = {
bolt_obj.DISABLED: 'MOVE_ARM_START_POSITION',
bolt_obj.SQUEEZE: 'SQUEEZE_SET_PRESSURE_SLOW',
bolt_obj.THERMAL_HOLD: 'HOLD_FOR_10_SECONDS',
bolt_obj.SLIDE: 'SLIDE_5CM',
bolt_obj.SLIDE_FAST: 'MOVE_DOWN_5CM'
}
# Get the current motion
current_motion = self.state_string[bolt_obj.state]
# Check if state passed in should be processed
if bolt_obj.state not in self.detailed_states:
return
else:
# Get detailed state if exists
current_detailed_state = self.detailed_states[bolt_obj.state]
# Check if the state is a the disabled state
if bolt_obj.state == bolt_obj.DISABLED:
self.norm_bolt_obj = upenn_features.pull_detailed_state(
bolt_obj, current_detailed_state)
return
else:
self.bolt_object = upenn_features.pull_detailed_state(
bolt_obj, current_detailed_state)
# Checks to make sure that the norm obj has been created
if not self.norm_bolt_obj:
print "Warning: there is no normalization data"
# Build the static features
static_feature_object, self.static_features_array = upenn_features.extract_static_features(
self.bolt_object, self.norm_bolt_obj)
static_feats = upenn_features.createFeatureVector(
static_feature_object)
# Store all feature vectors into one large vector array
# Technically we don't need to store the static
# features by adjective - but it's cleaner this way
for classifier in self.all_classifiers:
adj = classifier['adjective']
# need to add pkl to the name b/c the hmm chain dictionary
# takes adjective.pkl as its key
hmm_adj_name = '.'.join((adj, 'pkl'))
# Initialize the dictionary with adjective
if adj in self.adjective_vectors_static:
pass
else:
self.adjective_vectors_static[adj] = list()
self.adjective_vectors_dynamic[adj] = list()
# Pull out chain associated with adjective
# ordering of sensors - [pac, pdc, electrodes, tac]
sensor_hmm = ['pac', 'pdc', 'electrodes', 'tac']
hmm_features_phase = []
for sensor in sensor_hmm:
hmm_chain = self.hmm_chains[hmm_adj_name].chains[
current_detailed_state][sensor]
hmm_data = self.bolt_object.hmm[sensor]
# fill in dyanmic features
hmm_features_phase.append(hmm_chain.score(hmm_data))
# Store the feature vector by adjective away
self.adjective_vectors_static[adj].append(static_feats)
self.adjective_vectors_dynamic[adj].append(hmm_features_phase)
# Check if all motions have been performed
# If so - feed into classifier
if len(self.adjective_vectors_dynamic[adj]) == 4:
print 'All motions received! Computing adjective scores'
for classifier in self.all_classifiers:
# Pull out which adjective we are working on
adj = classifier['adjective']
# Load training vectors for kernel creation
train_dynamic = utilities.get_all_train_test_features(
adj, self.training_dynamic, train=True)
train_static = utilities.get_all_train_test_features(
adj, self.training_static, train=True)
# Scale the training data
d_scaler = preprocessing.StandardScaler().fit(
train_dynamic[0])
s_scaler = preprocessing.StandardScaler().fit(
train_static[0])
train_dynamic_scaled = d_scaler.transform(train_dynamic[0])
train_static_scaled = s_scaler.transform(train_static[0])
# Pull out the feature vectors for static/dynamic
all_static_feats = np.hstack(
self.adjective_vectors_static[adj])
all_dynamic_feats = np.hstack(
self.adjective_vectors_dynamic[adj])
# Normalize the features using scaler
all_static_feats_scaled = classifier[
'static_scaler'].transform(all_static_feats)
all_dynamic_feats_scaled = classifier[
'dynamic_scaler'].transform(all_dynamic_feats)
# Create kernels for both static and dynamic
static_kernel = self.linear_kernel_test(
all_static_feats_scaled, train_static_scaled, 1)
dynamic_kernel = self.linear_kernel_test(
all_dynamic_feats_scaled, train_dynamic_scaled, 1)
# Merge the two kernels
alpha = classifier['alpha']
merged_kernel = (alpha) * static_kernel + (
1 - alpha) * dynamic_kernel
# Predict adjective with computed kernel
clf = classifier['classifier']
self.mkl_results[adj] = clf.predict(merged_kernel)
# Store off the adjectives that returned true
adjectives_found = []
for adj in self.mkl_results:
if self.mkl_results[adj] == 1:
adjectives_found.append(Adj(adj))
publish_string = AdjList()
publish_string = adjectives_found
# Print and publish results!
print "Results from MKL classification"
#print self.mkl_results
print str(adjectives_found)
self.adjectives_pub.publish(publish_string)
def compute_probability_vector(self, bolt_obj):
# First object - initialize variables to store
# Also clears out the vectors for new run
if bolt_obj.state == bolt_obj.TAP:
# Store results as they come in
self.adjective_vectors = []
self.all_motion_results = dict()
self.results = dict()
# Store dictionary of strings
self.state_string = {bolt_obj.DISABLED:'disabled',
bolt_obj.THERMAL_HOLD:'thermal_hold',
bolt_obj.SLIDE:'slide',
bolt_obj.SQUEEZE:'squeeze',
bolt_obj.TAP:'tap',
bolt_obj.DONE:'done',
bolt_obj.SLIDE_FAST:'slide_fast',
bolt_obj.CENTER_GRIPPER:'center_gripper'
}
# store dictionary for detailed states
self.detailed_states = {bolt_obj.DISABLED:'MOVE_ARM_START_POSITION',
bolt_obj.SQUEEZE:'SQUEEZE_SET_PRESSURE_SLOW',
bolt_obj.THERMAL_HOLD:'HOLD_FOR_10_SECONDS',
bolt_obj.SLIDE:'SLIDE_5CM',
bolt_obj.SLIDE_FAST:'MOVE_DOWN_5CM'
}
# Get the current motion
current_motion = self.state_string[bolt_obj.state]
# Check if state passed in should be processed
if bolt_obj.state not in self.detailed_states:
return
else:
# Get detailed state if exists
current_detailed_state = self.detailed_states[bolt_obj.state]
# Check if the state is a the disabled state
if bolt_obj.state == bolt_obj.DISABLED:
self.norm_bolt_obj = upenn_features.pull_detailed_state(bolt_obj,current_detailed_state)
return
else:
self.bolt_object = upenn_features.pull_detailed_state(bolt_obj, current_detailed_state)
# Checks to make sure that the norm obj has been created
if not self.norm_bolt_obj:
print "Warning: there is no normalization data"
# Build the static features
static_feature_object, self.static_features_array = upenn_features.extract_static_features(self.bolt_object, self.norm_bolt_obj)
static_feats = upenn_features.createFeatureVector(static_feature_object)
# Add the current vector to all the features
self.adjective_vectors.append(static_feats)
print len(self.adjective_vectors)
# Check if all motions have been performed
# If so - feed into classifier
if len(self.adjective_vectors) == 4:
print 'All motions received! Computing adjective scores'
for classifier in self.all_classifiers:
# Pull out which adjective we are working on
adj = classifier['adjective']
# Pull out the feature vectors for static/dynamic
all_static_features = np.hstack(self.adjective_vectors)
# Normalize the features using scaler
all_static_feats_scaled = classifier['scaler'].transform(all_static_features)
# Predict adjective with computed kernel
clf = classifier['classifier']
self.results[adj] = clf.predict(all_static_feats_scaled)
# Store off the adjectives that returned true
adjectives_found = []
for adj in self.results:
if self.results[adj] == 1:
adjectives_found.append(Adj(adj))
publish_string = AdjList()
publish_string = adjectives_found
# Print and publish results!
print "Results from Static Only classification"
#print self.mkl_results
print str(adjectives_found)