def api_learn_add_del_route(op):
if not op in allowed_endpoints:
abort(404)
session['username'] = database.get_username(request.form['interface'],
request.form['username'])
if session['username'] == None:
abort(403)
session['logged_in'] = True
print '@@@@@@@@', session['username']
options = {}
if not op == 'query':
options = generic_learn_route(
op, request.form,
request.files['file'] if 'file' in request.files else None)
else:
try:
# Retrieve knowledge.
if 'type' in request.form and request.form['type'] == 'text':
options['text'] = database.get_text(session['username'])
elif 'type' in request.form and request.form['type'] == 'image':
options['pictures'] = database.get_images(session['username'])
else:
options['pictures'] = database.get_images(session['username'])
options['text'] = database.get_text(session['username'])
except Exception as e:
log(e)
options['errno'] = 500
options['error'] = str(e)
if 'errno' in options:
return json.dumps(options), options['errno']
return json.dumps(options), 200
def prepare_delta_E_sums(shifted_energies, temperatures, _accuracy, options):
"""
Prepares the sum of the second part of the Z_c_m equation
"""
energies_cnt = len(shifted_energies)
temp_cnt = len(temperatures)
delta_E_sums = np.empty([energies_cnt, temp_cnt], dtype=_accuracy)
log(__name__, "Preparing delta_E_sums to speed up the calculations", options.verbose, indent=3)
for t_i in range(temp_cnt):
temperature = temperatures[t_i]
kT = _accuracy(Constants.kB * temperature)
for e_i in range(energies_cnt):
case_energies = shifted_energies[e_i]
cnt_case_energies = len(case_energies)
sum_Ediff = _accuracy(0.0)
# summing in the reverse order in order to account for small values
for i in reversed(range(cnt_case_energies)):
sum_Ediff += _accuracy(math.exp(-1.0*(case_energies[i]) / (kT)))
delta_E_sums[e_i][t_i] = sum_Ediff
return delta_E_sums
def prepare_energies(input_data_array, _accuracy, options):
"""
Prepares energy arrays for the calculations
"""
log(__name__, "Preparing the energies", options.verbose, indent=3)
cases_cnt = input_data_array.shape[0]
min_energies = np.empty(cases_cnt, dtype=_accuracy)
energies = []
shifted_energies = []
experiment_cnts = np.empty(cases_cnt, dtype=_accuracy)
for i in range(cases_cnt):
# removes all zero values from the beginning and end of the array
case_energies = np.trim_zeros(copy.deepcopy(input_data_array[i]))
energies.append(case_energies)
# it might be the first value but it might not be, thus we getting the minimum
min_energies[i] = _accuracy(np.min(case_energies))
diff_energy = case_energies - min_energies[i]
shifted_energies.append(diff_energy)
# saving the number of experiments of the stoichiometry
experiment_cnts[i] = len(case_energies)
return energies, min_energies, shifted_energies, experiment_cnts
def prepare_parameters(options):
"""
Prepares analysis parameters: temperature and mu
"""
success = True
error = ""
temperatures = None
chem_pot_range = None
# reading in the temperatures
success, error, temperatures = Utilities.get_list_of_temps(options.temps)
if not success:
return success, error
# reading in the chemical potential
if options.urange is not None:
success, error, chem_pot_range = Utilities.get_chem_pot_range(options.urange)
if not success:
return success, error
log(__name__, "Analysis will be performed at (K): %s" % (options.temps), options.verbose, indent=2)
if chem_pot_range is not None:
log(__name__, "mu parameter: %s" % (options.urange), options.verbose, indent=2)
return success, error, temperatures, chem_pot_range
def api_learn_add_del_route(op):
if not op in allowed_endpoints:
abort(404)
session['username'] = database.get_username(request.form['interface'], request.form['username'])
if session['username'] == None:
abort (403)
session['logged_in'] = True
print '@@@@@@@@', session['username']
options = {}
if not op == 'query':
options = generic_learn_route(op, request.form, request.files['file'] if 'file' in request.files else None)
else:
try:
# Retrieve knowledge.
if 'type' in request.form and request.form['type'] == 'text':
options['text'] = database.get_text(session['username'])
elif 'type' in request.form and request.form['type'] == 'image':
options['pictures'] = database.get_images(session['username'])
else:
options['pictures'] = database.get_images(session['username'])
options['text'] = database.get_text(session['username'])
except Exception as e:
log(e)
options['errno'] = 500
options['error'] = str(e)
if 'errno' in options:
return json.dumps(options), options['errno']
return json.dumps(options), 200
def get_images(self, username):
log('Retrieving all images from images_' + username)
# Notice image['data'] was encoded using Base64.
return [
image for image in self.get_image_collection(username).find(
{}, {'_id': 0})
]
def infer(self, LUCID, service_graph, text_data, image_data):
# Create the list of QueryInput.
query_input_list = []
for node in service_graph.node_list:
service = self.SERVICES[node.service_name]
data = text_data if service.input_type == 'text' else image_data
host, port = service.get_host_port()
tag_list = [host, str(port), str(len(node.to_indices))]
for to_index in node.to_indices:
tag_list.append(str(to_index))
query_input_list.append(self.create_query_input(
service.input_type, data, tag_list))
query_spec = self.create_query_spec('query', query_input_list)
# Go through all starting indices and send requests.
result = []
for start_index in service_graph.starting_indices:
service = self.SERVICES[service_graph.get_node(
start_index).service_name]
client, transport = self.get_client_transport(service)
log('Sending infer request to ' + service.name)
request = Request()
request.LUCID = str(LUCID)
request.spec = query_spec
result.append(client.infer(Request).msg)
#transport.close()
return ' '.join(result)
def send_query(self, LUCID, service_name, query_input_list):
query_spec = self.create_query_spec('query', query_input_list)
service = self.SERVICES[service_name]
client, transport = self.get_client_transport(service)
log('Sending infer request to ' + service.name)
result = client.infer(str(LUCID), query_spec)
transport.close()
return result
def __init__(self):
memcached_addr = os.environ.get('MEMCACHED_PORT_11211_TCP_ADDR')
if memcached_addr:
log('Memcached: ' + memcached_addr)
self.client = Client([(memcached_addr, 11211)])
else:
log('Memcached: localhost')
self.client = Client([('127.0.0.1', 11211)])
def send_query(self, LUCID, service_name, query_input_list):
query_spec = self.create_query_spec('query', query_input_list)
service = self.SERVICES[service_name]
client, transport = self.get_client_transport(service)
log('Sending infer request to ' + service.name)
result = client.infer(str(LUCID), query_spec)
transport.close()
return result
def __init__(self):
memcached_addr = os.environ.get('MEMCACHED_PORT_11211_TCP_ADDR')
if memcached_addr:
log('Memcached: ' + memcached_addr)
self.client = Client([(memcached_addr, 11211)])
else:
log('Memcached: localhost')
self.client = Client([('127.0.0.1', 11211)])
def __init__(self):
mongodb_addr = os.environ.get('MONGO_PORT_27017_TCP_ADDR')
if mongodb_addr:
log('MongoDB: ' + mongodb_addr)
self.db = MongoClient(mongodb_addr, 27017).lucida
else:
log('MongoDB: localhost')
self.db = MongoClient().lucida
self.users = self.db.users
def learn_image(self, LUCID, image_type, image_data, label):
for service in Config.Service.LEARNERS['image']: # add concurrency?
knowledge_input = self.create_query_input(
image_type, image_data, [label])
client, transport = self.get_client_transport(service)
log('Sending learn_image request to IMM')
client.learn(str(LUCID),
self.create_query_spec('knowledge', [knowledge_input]))
transport.close()
def learn_text(self, LUCID, text_type, text_data, text_id):
for service in Config.Service.LEARNERS['text']: # add concurrency?
knowledge_input = self.create_query_input(
text_type, text_data, [text_id])
client, transport = self.get_client_transport(service)
log('Sending learn_text request to QA')
client.learn(str(LUCID),
self.create_query_spec('knowledge', [knowledge_input]))
transport.close()
def __init__(self):
if os.environ.get('MONGO_PORT_27017_TCP_ADDR'):
log('MongoDB: ' + os.environ.get('MONGO_PORT_27017_TCP_ADDR'))
self.db = MongoClient(os.environ.get('MONGO_PORT_27017_TCP_ADDR'),
27017).lucida
else:
log('MongoDB: localhost')
self.db = MongoClient().lucida
self.users = self.db.users
def __init__(self):
mongodb_addr = os.environ.get('MONGO_PORT_27017_TCP_ADDR')
if mongodb_addr:
log('MongoDB: ' + mongodb_addr)
self.db = MongoClient(mongodb_addr, 27017).lucida
else:
log('MongoDB: localhost')
self.db = MongoClient().lucida
self.users = self.db.users
def infer_route():
options = {}
if os.environ.get('ASR_ADDR_PORT'):
options['asr_addr_port'] = os.environ.get('ASR_ADDR_PORT')
else:
options['asr_addr_port'] = 'ws://*****:*****@@@@@@@@@@', speech_input
image_input = [upload_file.read()] if upload_file else None
lucida_id = session['username']
# Check if context is saved for Lucida user
# If not, classify query, otherwise restore session
if lucida_id not in Config.SESSION:
services_needed = \
query_classifier.predict(speech_input, upload_file)
speech_input = [speech_input]
else:
services_needed = Config.SESSION[lucida_id]['graph']
Config.SESSION[lucida_id]['data']['text'].append(
speech_input)
speech_input = Config.SESSION[lucida_id]['data']['text']
options['result'] = thrift_client.infer(lucida_id, \
services_needed, speech_input, image_input)
log('Result ' + options['result'])
# Check if Calendar service is needed.
# If so, JavaScript needs to receive the parsed dates.
if services_needed.has_service('CA'):
options['dates'] = options['result']
options['result'] = None
return render_template('infer.html', **options)
else:
raise RuntimeError('Did you click the Ask button?')
except Exception as e:
log(e)
if str(e) == 'TSocket read 0 bytes':
e = 'Back-end service encountered a problem'
options['error'] = e
return render_template('infer.html', **options)
# Display.
return render_template('infer.html', **options)
def get_host_port(self):
try:
host = 'localhost'
tcp_addr = os.environ.get(self.name + '_PORT_' + str(self.port) + '_TCP_ADDR')
if tcp_addr:
log('TCP address is resolved to ' + tcp_addr)
host = tcp_addr
return host, self.port
except Exception:
raise RuntimeError('Cannot access service ' + self.name)
def learn_image(self, LUCID, image_type, image_data, image_id):
for service in Config.Service.LEARNERS['image']: # add concurrency?
knowledge_input = self.create_query_input(image_type, image_data,
[image_id])
client, transport = self.get_client_transport(service)
log('Sending learn_image request to IMM')
client.learn(
str(LUCID),
self.create_query_spec('knowledge', [knowledge_input]))
transport.close()
def learn_text(self, LUCID, text_type, text_data, text_id):
for service in Config.Service.LEARNERS['text']: # add concurrency?
knowledge_input = self.create_query_input(text_type, text_data,
[text_id])
client, transport = self.get_client_transport(service)
log('Sending learn_text request to QA')
client.learn(
str(LUCID),
self.create_query_spec('knowledge', [knowledge_input]))
transport.close()
def generic_infer_route(form, upload_file):
options = {}
if os.environ.get('ASR_ADDR_PORT'):
options['asr_addr_port'] = os.environ.get('ASR_ADDR_PORT')
else:
options['asr_addr_port'] = 'ws://localhost:' + port_dic["cmd_port"]
try:
# Deal with POST requests.
if request.method == 'POST':
if not upload_file is None and upload_file.filename != '':
check_image_extension(upload_file)
# Classify the query.
speech_input = form['speech_input'] if 'speech_input' in form \
else ''
print 'Query: ', speech_input
image_input = [upload_file.read()] if upload_file else None
lucida_id = session['username']
# Check if context is saved for Lucida user
# If not, classify query, otherwise restore session
if lucida_id not in Config.SESSION:
services_needed = query_classifier.predict(
speech_input, upload_file)
speech_input = [speech_input]
else:
services_needed = Config.SESSION[lucida_id]['graph']
Config.SESSION[lucida_id]['data']['text'].append(speech_input)
speech_input = Config.SESSION[lucida_id]['data']['text']
node = services_needed.get_node(0)
try:
options['result'] = thrift_client.infer(
lucida_id, node.service_name, speech_input, image_input)
except Exception as ex:
print "Exception raised while trying to infer", ex.message
options['error'] = str(ex)
raise
if 'result' in options:
log('Result ' + options['result'])
# Check if Calendar service is needed.
# If so, JavaScript needs to receive the parsed dates.
if services_needed.has_service('CAWF'):
options['dates'] = options['result']
options['result'] = None
else:
options['error'] = "Result was empty."
except Exception as e:
log(e)
options['errno'] = "Unknown"
options['error'] = str(e.message)
if 'code' in e and re.match("^4\d\d$", str(e.code)):
options['errno'] = e.code
if str(e) == 'TSocket read 0 bytes':
options['error'] = 'Back-end service encountered a problem'
if str(e).startswith('Could not connect to'):
options['error'] = 'Back-end service is not running'
return options
def calc_average_value(temperatures,
chem_pot_range,
prop_array,
prop_name,
Wm_array,
_accuracy,
options,
temp_depend=False):
"""
Calculates average value of a system's property
"""
success = True
error = ""
log(__name__,
"Calculating an average value of: %s" % (prop_name),
options.verbose,
indent=3)
temp_len = len(temperatures)
chem_pot_len = len(chem_pot_range)
# is the value temperature dependent
if not temp_depend:
prop_len = len(prop_array)
else:
prop_len = len(prop_array[0])
avg_array = np.zeros([temp_len, chem_pot_len], _accuracy)
# for each temperature:
for t_index in range(temp_len):
temperature = temperatures[t_index]
for mu_index in range(chem_pot_len):
prop_avg = _accuracy(0.0)
for prop_index in range(prop_len):
wm_value = Wm_array[t_index, mu_index, prop_index]
# is the value temperature dependent
if not temp_depend:
prop_value = prop_array[prop_index]
else:
prop_value = prop_array[t_index][prop_index]
prop_avg += wm_value * prop_value
avg_array[t_index, mu_index] = prop_avg
return success, error, avg_array
def get_host_port(self):
try:
host = 'localhost' # TODO - Need to replace with a dynamic service finding system - good candidate is to finish the full CREATE system
tcp_addr = os.environ.get(self.name + '_PORT_' + str(self.port) +
'_TCP_ADDR')
if tcp_addr:
log('TCP address is resolved to ' + tcp_addr)
host = tcp_addr
return host, self.port
except Exception:
raise RuntimeError('Cannot access service ' + self.name)
def get_host_port(self):
try:
host = 'localhost'
tcp_addr = os.environ.get(
self.name + '_PORT_' + str(self.port) + '_TCP_ADDR')
if tcp_addr:
log('TCP address is resolved to ' + tcp_addr)
host = tcp_addr
return host, self.port
except Exception:
raise RuntimeError('Cannot access service ' + self.name)
def distribution_analysis(chem_pot_multi, names, temperatures, chem_pot_range,
min_energies, delta_E_sums, experiment_cnts,
permutations, omega_c_arr, _accuracy, options):
"""
Performs the distribution analysis: evaluates Wm and plots it against m and mu.
"""
log(__name__, "Distribution analysis", options.verbose, indent=2)
success = True
error = ""
# Preparing Wm probabilities
Wm_array = prepare_Wm(chem_pot_multi, temperatures, chem_pot_range,
min_energies, delta_E_sums, experiment_cnts,
permutations, _accuracy, options)
# Writing Wm into a file
success, error = IO.write_Wm(temperatures, chem_pot_range, chem_pot_multi,
Wm_array)
if not success:
return success, error, Wm_array
# Plotting the Wm probabilities 3D plots
Graphs.wm_contour(temperatures, names, chem_pot_range, chem_pot_multi,
Wm_array, _accuracy, options)
# Performing analysis with respect to the distribution function (average m is the standard analysis)
# average m
average_analysis(temperatures,
chem_pot_range,
chem_pot_multi,
"m",
Wm_array,
_accuracy,
options,
temp_depend=False)
# average gamma
average_analysis(temperatures,
chem_pot_range,
omega_c_arr,
"\gamma^{c}",
Wm_array,
_accuracy,
options,
temp_depend=True)
return success, error, Wm_array
def learn_route():
options = {}
# Deal with POST requests.
if request.method == 'POST':
options = generic_learn_route(request.form['op'], request.form, request.files['file'] if 'file' in request.files else None)
try:
# Retrieve knowledge.
options['pictures'] = database.get_images(session['username'])
options['text'] = database.get_text(session['username'])
except Exception as e:
log(e)
options['errno'] = 500
options['error'] = str(e)
return render_template('learn.html', **options)
def prepare_data(data, temperatures, _accuracy, options):
"""
Prepares the data for further analysis: energy difference with respect to corresponding lowest energy, Zcm sums
"""
# first of all lets prepare the data for the calculations.
log(__name__, "Preparing the data", options.verbose, indent=2)
energies, min_energies, shifted_energies, experiment_cnts = prepare_energies(data, _accuracy, options)
delta_E_sums = prepare_delta_E_sums(shifted_energies, temperatures, _accuracy, options)
return energies, min_energies, shifted_energies, experiment_cnts, delta_E_sums
def _sendFinishedSignal(self, dest):
"""
Sends a signal to the server that it is finished its job
"""
log(__name__,
"Worker %000d | sending a FINISHED response" % (self.rank), 1)
dataPackage = {
Config._MPIRankTag: self.rank,
Config._MPISignalDataTag: Config._MPISignalFinished
}
self.comm.send(dataPackage, dest=dest, tag=Config._MPITagClient)
def _sendReadyForDataSignal(self, dest):
"""
Sends a signal to the server that it is ready to receive data
"""
log(__name__,
"Worker %000d | sending a READY 4 DATA response" % (self.rank), 1)
dataPackage = {
Config._MPIRankTag: self.rank,
Config._MPISignalDataTag: Config._MPISignalReady4Data
}
self.comm.send(dataPackage, dest=dest, tag=Config._MPITagClient)
def load(self, input_type, query_classes):
current_dir = os.path.abspath(os.path.dirname(__file__))
# If there is no or only one possible outcomes for the input type,
# there is no need to train any classifier.
if len(query_classes) <= 1:
return DummyClassifier(query_classes.keys()[0])
try:
with open(
current_dir + '/../models/dumped_classifier_' +
input_type + '.pkl', 'rb') as fid:
log('Loading model for ' + input_type)
return cPickle.load(fid)
except IOError as e:
print e
exit(1)
def load(self, input_type, query_classes):
current_dir = os.path.abspath(os.path.dirname(__file__))
# If there is no or only one possible outcomes for the input type,
# there is no need to train any classifier.
if len(query_classes) <= 1:
return DummyClassifier(query_classes.keys()[0])
try:
with open(current_dir +
'/../models/dumped_classifier_' + input_type + '.pkl',
'rb') as fid:
log('Loading model for ' + input_type)
return cPickle.load(fid)
except IOError as e:
print e
exit(1)
def learn_route():
options = {}
# Deal with POST requests.
if request.method == 'POST':
options = generic_learn_route(
request.form['op'], request.form,
request.files['file'] if 'file' in request.files else None)
try:
# Retrieve knowledge.
options['pictures'] = database.get_images(session['username'])
options['text'] = database.get_text(session['username'])
except Exception as e:
log(e)
options['errno'] = 500
options['error'] = str(e)
return render_template('learn.html', **options)
def g_c_calc_omega(chem_pot_multi, temperatures, chem_pot_range, min_energies, delta_E_sums, experiment_cnts,
permutations, _accuracy, options):
"""
Calculates omega values with respect to temperature and chemical potential (grand canonical analysis)
"""
success = True
error = ""
log(__name__, "Calculating Omega (grand canonical)", options.verbose, indent=3)
temp_len = len(temperatures)
chem_pot_len = len(chem_pot_range)
chem_pot_multi_len = len(chem_pot_multi)
global_min_energy = np.min(min_energies)
omega_arr = np.zeros([temp_len, chem_pot_len], _accuracy)
# for each temperature:
for t_index in range(temp_len):
temperature = temperatures[t_index]
kT = np.longdouble(Constants.kB * temperature)
# for each chemical potential value
for mu_index in range(chem_pot_len):
mu_value = chem_pot_range[mu_index]
sum2 = _accuracy(0.0)
# for each composition
for m_index in range(chem_pot_multi_len):
m_value = chem_pot_multi[m_index]
min_energy = min_energies[m_index]
exp_expr = _accuracy(-1.0*(min_energy - global_min_energy + m_value*mu_value) / (kT))
sum2 += np.exp(exp_expr) * (permutations[m_index]/experiment_cnts[m_index]) * delta_E_sums[m_index][t_index]
omega_value = -kT * (-global_min_energy + np.log(sum2))
omega_arr[t_index, mu_index] = omega_value
return success, error, omega_arr
def _sendFailedSignal(self, dest, failedMessage):
"""
Sends a signal to the server that the job has failed
"""
log(
__name__, "Worker %000d | sending a FAILURE response: %s" %
(self.rank, failedMessage), 1)
dataPackage = {
Config._MPIRankTag: self.rank,
Config._MPISignalDataTag: Config._MPISignalFailed,
Config._MPIMessageTag: failedMessage
}
self.comm.send(dataPackage, dest=dest, tag=Config._MPITagClient)
def predict(self, speech_input, image_input):
input_type = ''
if speech_input:
if image_input:
input_type = 'text_image'
else:
input_type = 'text'
else:
if image_input:
input_type = 'image'
else:
raise RuntimeError('Text and image cannot be both empty')
# Convert speech_input to a single-element list.
class_predicted = self.classifiers[input_type].predict([speech_input])
class_predicted = class_predicted[0] # ndarray to string
log('Query classified as ' + class_predicted)
return self.CLASSIFIER_DESCRIPTIONS[input_type][class_predicted]
def predict(self, speech_input, image_input):
input_type = ''
if speech_input:
if image_input:
input_type = 'text_image'
else:
input_type = 'text'
else:
if image_input:
input_type = 'image'
else:
raise RuntimeError('Text and image cannot be both empty')
# Convert speech_input to a single-element list.
class_predicted = self.classifiers[input_type].predict([speech_input])
class_predicted = class_predicted[0] # ndarray to string
log('Query classified as ' + class_predicted)
return self.CLASSIFIER_DESCRIPTIONS[input_type][class_predicted]
def generic_infer_route(form, upload_file):
options = {}
if os.environ.get('ASR_ADDR_PORT'):
options['asr_addr_port'] = os.environ.get('ASR_ADDR_PORT')
else:
options['asr_addr_port'] = 'ws://*****:*****@@@@@@@@@@', speech_input
image_input = [upload_file.read()] if upload_file else None
lucida_id = session['username']
# Check if context is saved for Lucida user
# If not, classify query, otherwise restore session
if lucida_id not in Config.SESSION:
services_needed = query_classifier.predict(speech_input, upload_file)
speech_input = [speech_input]
else:
services_needed = Config.SESSION[lucida_id]['graph']
Config.SESSION[lucida_id]['data']['text'].append(speech_input)
speech_input = Config.SESSION[lucida_id]['data']['text']
node = services_needed.get_node(0)
options['result'] = thrift_client.infer(lucida_id, node.service_name, speech_input, image_input)
log('Result ' + options['result'])
# Check if Calendar service is needed.
# If so, JavaScript needs to receive the parsed dates.
if services_needed.has_service('CA'):
options['dates'] = options['result']
options['result'] = None
except Exception as e:
log(e)
options['errno'] = 500
options['error'] = str(e)
if 'code' in e and re.match("^4\d\d$", str(e.code)):
options['errno'] = e.code
if str(e) == 'TSocket read 0 bytes':
options['error'] = 'Back-end service encountered a problem'
if str(e).startswith('Could not connect to'):
options['error'] = 'Back-end service is not running'
return options
def _sendStopSignal(self, dest):
"""
Sends a signal to the server that it is stopping
"""
log(__name__, "Worker %000d | received a STOP signal" % (self.rank), 1)
log(__name__, "Worker %000d | sending a QUIT response" % (self.rank),
1)
# sending QUIT message to the server
dataPackage = {
Config._MPIRankTag: self.rank,
Config._MPISignalDataTag: Config._MPISignalQuit
}
self.comm.send(dataPackage, dest=dest, tag=Config._MPITagClient)
def __init__(self, TRAIN_OR_LOAD, CLASSIFIER_DESCRIPTIONS_IN):
self.CLASSIFIER_DESCRIPTIONS = CLASSIFIER_DESCRIPTIONS_IN
self.classifiers = {}
# Each input type has its own classifier.
for input_type in self.CLASSIFIER_DESCRIPTIONS:
# query_classes represents all the possible classification outcomes
# and their needed services for a given input type.
if TRAIN_OR_LOAD == 'train':
self.classifiers[input_type] = self.train(input_type,
self.CLASSIFIER_DESCRIPTIONS[input_type])
elif TRAIN_OR_LOAD == 'load':
self.classifiers[input_type] = self.load(input_type,
self.CLASSIFIER_DESCRIPTIONS[input_type])
else:
raise RuntimeError(
'TRAIN_OR_LOAD must be either "train" or "load"')
log('@@@@@ Summary of classifiers:')
log(str(self.classifiers))
def __init__(self, TRAIN_OR_LOAD, CLASSIFIER_DESCRIPTIONS_IN):
self.CLASSIFIER_DESCRIPTIONS = CLASSIFIER_DESCRIPTIONS_IN
self.classifiers = {}
# Each input type has its own classifier.
for input_type in self.CLASSIFIER_DESCRIPTIONS:
# query_classes represents all the possible classification outcomes
# and their needed services for a given input type.
if TRAIN_OR_LOAD == 'train':
self.classifiers[input_type] = self.train(
input_type, self.CLASSIFIER_DESCRIPTIONS[input_type])
elif TRAIN_OR_LOAD == 'load':
self.classifiers[input_type] = self.load(
input_type, self.CLASSIFIER_DESCRIPTIONS[input_type])
else:
raise RuntimeError(
'TRAIN_OR_LOAD must be either "train" or "load"')
log('@@@@@ Summary of classifiers:')
log(str(self.classifiers))
def c_calc_gamma(temperatures, min_energies, delta_E_sums, experiment_cnts, permutations, _accuracy, options):
"""
Calculates omega values with respect to temperature (canonical analysis)
"""
success = True
error = ""
log(__name__, "Calculating Omega (canonical)", options.verbose, indent=3)
temp_len = len(temperatures)
comp_len = len(min_energies)
omega_arr = np.zeros([temp_len, comp_len], _accuracy)
# for each temperature:
for t_index in range(temp_len):
temperature = temperatures[t_index]
kT = np.longdouble(Constants.kB * temperature)
# for each composition
for c_index in range(comp_len):
# calculating Z^c_m
# Pm
if options.permCalc:
Pm = _accuracy(calc_permutation(m_value, mm_value, _accuracy))
else:
Pm = _accuracy(permutations[c_index])
# Nm
Nm = experiment_cnts[c_index]
#Z_cm = np.exp(-1.0*(min_energies[c_index]) / kT) * (Pm/Nm) * delta_E_sums[c_index][t_index]
#omega_value = - kT * np.log(Z_cm)
omega_value = - kT * (-1.0*(min_energies[c_index]) / kT + np.log((Pm/Nm)) + np.log(delta_E_sums[c_index][t_index]))
omega_arr[t_index, c_index] = omega_value / Constants.gamma_c_m_coef
return success, error, omega_arr
def infer_route():
options = {}
if os.environ.get('ASR_ADDR_PORT'):
options['asr_addr_port'] = os.environ.get('ASR_ADDR_PORT')
else:
options['asr_addr_port'] = 'ws://*****:*****@@@@@@@@@@', form['speech_input']
# Classify the query.
services_needed = \
query_classifier.predict(form['speech_input'], upload_file)
options['result'] = thrift_client.infer(
session['username'], services_needed, form['speech_input'],
upload_file.read())
log('Result ' + options['result'])
# Check if Calendar service is needed.
# If so, JavaScript needs to receive the parsed dates.
if services_needed.has_service('CA'):
options['dates'] = options['result']
options['result'] = None
return render_template('infer.html', **options)
else:
raise RuntimeError('Did you click the Ask button?')
except Exception as e:
log(e)
if str(e) == 'TSocket read 0 bytes':
e = 'Back-end service encountered a problem'
options['error'] = e
return render_template('infer.html', **options)
# Display.
return render_template('infer.html', **options)
def infer_route():
options = {}
if os.environ.get("ASR_ADDR_PORT"):
options["asr_addr_port"] = os.environ.get("ASR_ADDR_PORT")
else:
options["asr_addr_port"] = "ws://*****:*****@@@@@@@@@@", form["speech_input"]
# Classify the query.
services_needed = query_classifier.predict(form["speech_input"], upload_file)
options["result"] = thrift_client.infer(
session["username"], services_needed, form["speech_input"], upload_file.read()
)
log("Result " + options["result"])
# Check if Calendar service is needed.
# If so, JavaScript needs to receive the parsed dates.
if services_needed.has_service("CA"):
options["dates"] = options["result"]
options["result"] = None
return render_template("infer.html", **options)
else:
raise RuntimeError("Did you click the Ask button?")
except Exception as e:
log(e)
if str(e) == "TSocket read 0 bytes":
e = "Back-end service encountered a problem"
options["error"] = e
return render_template("infer.html", **options)
# Display.
return render_template("infer.html", **options)
def infer(self, LUCID, service_graph, text_data, image_data):
# Create the list of QueryInput.
query_input_list = []
for node in service_graph.node_list:
service = self.SERVICES[node.service_name]
data = text_data if service.input_type == 'text' else image_data
host, port = service.get_host_port()
tag_list = [host, str(port), str(len(node.to_indices))]
for to_index in node.to_indices:
tag_list.append(str(to_index))
query_input_list.append(self.create_query_input(
service.input_type, data, tag_list))
query_spec = self.create_query_spec('query', query_input_list)
# Go through all starting indices and send requests.
result = []
for start_index in service_graph.starting_indices:
service = self.SERVICES[service_graph.get_node(
start_index).service_name]
client, transport = self.get_client_transport(service)
log('Sending infer request to ' + service.name)
result.append(client.infer(str(LUCID), query_spec))
transport.close()
return ' '.join(result)
def count_images(self, username):
log('Retrieving the number of images from images_' + username)
return self.get_image_collection(username).count()
def __init__(self, SERVICES):
self.SERVICES = SERVICES
log('Pre-configured services: ' + str(SERVICES))
def get_text(self, username):
log('Retrieving text from text_' + username)
return [text for text in self.get_text_collection(username).find()]
def generic_learn_route(op, form, upload_file):
options = {}
username = session['username']
try:
# Add image knowledge.
if op == 'add_image':
image_type = 'image'
label = form['label']
# Check the uploaded image.
if upload_file.filename == '':
raise RuntimeError('Empty file is not allowed')
check_image_extension(upload_file)
# Check the label of the image.
check_text_input(label)
# Check whether the user can add one more image.
database.check_add_image(username)
# Generate the id.
image_data = upload_file.read()
image_id = hashlib.md5(username +
str(datetime.datetime.now())).hexdigest()
# Send the image to IMM.
upload_file.close()
thrift_client.learn_image(username, image_type, image_data,
image_id)
# Add the image into the database.
database.add_image(username, image_data, label, image_id)
# Delete image knowledge.
elif op == 'delete_image':
image_type = 'unlearn'
image_id = form['image_id']
# Send the unlearn request to IMM.
thrift_client.learn_image(username, image_type, '', image_id)
# Delete the image from the database.
database.delete_image(username, image_id)
# Add text knowledge.
elif op == 'add_text' or op == 'add_url':
text_type = 'text' if op == 'add_text' else 'url'
text_data = form['knowledge']
# Check the text knowledge.
check_text_input(text_data)
# Check whether the user can add one more piece of text.
database.check_add_text(username)
# Generate the id.
text_id = hashlib.md5(username + text_data +
str(datetime.datetime.now())).hexdigest()
# Send the text to QA.
thrift_client.learn_text(username, text_type,
text_data, text_id)
# Add the text knowledge into the database.
database.add_text(username, text_type, text_data, text_id)
# Delete text knowledge.
elif op == 'delete_text':
text_type = 'unlearn'
text_id = form['text_id']
# Send the unlearn request to QA.
thrift_client.learn_text(username, text_type, '', text_id)
# Delete the text from into the database.
database.delete_text(username, text_id)
else:
raise RuntimeError('Did you click the button?')
except Exception as e:
log(e)
options['errno'] = 500
options['error'] = str(e)
if 'code' in e and re.match("^4\d\d$", str(e.code)):
options['errno'] = e.code
if str(e) == 'TSocket read 0 bytes':
options['error'] = 'Back-end service encountered a problem'
if str(e).startswith('Could not connect to'):
options['error'] = 'Back-end service is not running'
return options
def get_images(self, username):
log('Retrieving all images from images_' + username)
# Notice image['data'] was encoded using Base64.
return [image for image in self.get_image_collection(username).find()]
def train(self, input_type, query_classes):
log('********************** ' + input_type + ' **********************')
current_dir = os.path.abspath(os.path.dirname(__file__))
# If there is no or only one possible outcomes for the input type,
# there is no need to train any classifier.
if len(query_classes) <= 1:
return DummyClassifier(query_classes.keys()[0])
# Build DataFrame by going through all data files.
data = DataFrame({'text': [], 'class': []})
for query_class_name in query_classes:
path = current_dir + '/../data/' + query_class_name + '.txt'
log('Opening ' + path)
lines = [line.rstrip('\n') for line in open(path)]
rows = []
index = []
for text in lines:
if text in index:
log('duplicate in ' + path + ": " + text)
exit(1)
rows.append({'text': text, 'class': query_class_name})
index.append(text)
data = data.append(DataFrame(rows, index))
# Build the pipeline.
pipeline = Pipeline([
('count_vectorizer', CountVectorizer(ngram_range = (1, 2))),
# ('classifier', PassiveAggressiveClassifier())
('classifier', LinearSVC())
])
# Train and k-fold cross-validate. Introduce randomness.
data = data.reindex(numpy.random.permutation(data.index))
k_fold = KFold(n=len(data), n_folds=6)
scores = []
for train_indices, test_indices in k_fold:
train_text = data.iloc[train_indices]['text'].values
train_y = data.iloc[train_indices]['class'].values.astype(str)
test_text = data.iloc[test_indices]['text'].values
test_y = data.iloc[test_indices]['class'].values.astype(str)
pipeline.fit(train_text, train_y)
predictions = pipeline.predict(test_text)
score = f1_score(test_y, predictions,
pos_label=None if len(query_classes) == 2 else 1,
average='weighted')
scores.append(score)
log('Total documents classified:' + str(len(data)))
log('Score:' + str(sum(scores) / len(scores)))
# Save the classifier,
if not os.path.exists(current_dir + '/../models'):
os.makedirs(current_dir + '/../models')
with open(current_dir + '/../models/dumped_classifier_' +
input_type + '.pkl', 'wb') as fid:
log('Saving model for ' + input_type)
cPickle.dump(pipeline, fid)
return pipeline
def learn_route():
options = {}
username = session['username']
try:
form = request.form
# Deal with POST requests.
if request.method == 'POST':
# If the request does not contain an "op" field.
if not 'op' in request.form:
raise RuntimeError('Did you click the button?')
# Add image knowledge.
elif form['op'] == 'add_image':
image_type = 'image'
label = form['label']
# Check the uploaded image.
upload_file = request.files['file']
if upload_file.filename == '':
raise RuntimeError('Empty file is not allowed')
check_image_extension(upload_file)
# Check the label of the image.
check_text_input(label)
# Check whether the user can add one more image.
database.check_add_image(username)
# Send the image to IMM.
image_data = upload_file.read()
upload_file.close()
thrift_client.learn_image(username, image_type, image_data,
label)
# Add the image into the database.
database.add_image(username, image_data, label)
# Delete image knowledge.
elif form['op'] == 'delete_image':
image_type = 'unlearn'
label = form['label']
# Send the unlearn request to IMM.
thrift_client.learn_image(username, image_type, '', label)
# Delete the image from the database.
database.delete_image(username, label)
# Add text knowledge.
elif form['op'] == 'add_text' or form['op'] == 'add_url':
text_type = 'text' if form['op'] == 'add_text' else 'url'
text_data = form['knowledge']
# Check the text knowledge.
check_text_input(text_data)
# Check whether the user can add one more piece of text.
database.check_add_text(username)
# Generate the id.
text_id = hashlib.md5(username + text_data
+ str(datetime.datetime.now())).hexdigest()
# Send the text to QA.
thrift_client.learn_text(username, text_type,
text_data, text_id)
# Add the text knowledge into the database.
database.add_text(username, text_type, text_data, text_id)
# Delete text knowledge.
elif form['op'] == 'delete_text':
text_type = 'unlearn'
text_id = form['text_id']
# Send the unlearn request to QA.
thrift_client.learn_text(username, text_type, '', text_id)
# Delete the text from into the database.
database.delete_text(username, text_id)
else:
raise RuntimeError('Did you click the button?')
except Exception as e:
log(e)
if str(e) == 'TSocket read 0 bytes':
e = 'Back-end service encountered a problem'
options['error'] = e
try:
# Retrieve knowledge.
options['pictures'] = database.get_images(username)
options['text'] = database.get_text(username)
except Exception as e:
log(e)
options['error'] = e
return render_template('learn.html', **options)