elif sys.argv[2] == 'adam_L2':
from acidano.utils.optim import adam_L2 as Optimization_method
elif sys.argv[2] == 'rmsprop':
from acidano.utils.optim import rmsprop as Optimization_method
elif sys.argv[2] == 'sgd_nesterov':
from acidano.utils.optim import sgd_nesterov as Optimization_method
else:
raise ValueError(sys.argv[2] + " is not an optimization method")
############################################################
# System paths
############################################################
logging.info('System paths')
SOURCE_DIR = os.getcwd()
result_folder = SOURCE_DIR + u'/../Results/' + unit_type + '/' + script_param['temporal_granularity'] + '/' +\
'quantization_' + str(script_param['quantization']) + '/' + Optimization_method.name() + '/' + Model_class.name()
# Check if the result folder exists
if not os.path.isdir(result_folder):
os.makedirs(result_folder)
# Count number of files in the folder and name this hparam run by the number
hparam_run_counter = 0
while(os.path.isdir(result_folder + '/hrun_' + str(hparam_run_counter))):
hparam_run_counter += 1
result_folder = result_folder + '/hrun_' + str(hparam_run_counter)
os.mkdir(result_folder)
script_param['result_folder'] = result_folder
# Data : .pkl files
data_folder = SOURCE_DIR + '/../Data'
def train_hopt(max_evals, csv_file_path):
# Create/reinit csv file
open(csv_file_path, 'w').close()
logger_hopt.info((u'WITH HYPERPARAMETER OPTIMIZATION').encode('utf8'))
logger_hopt.info((u'**** Model : ' + Model_class.name()).encode('utf8'))
logger_hopt.info((u'**** Optimization technic : ' + Optimization_method.name()).encode('utf8'))
logger_hopt.info((u'**** Temporal granularity : ' + temporal_granularity).encode('utf8'))
if binary_unit:
logger_hopt.info((u'**** Binary unit (intensity discarded)').encode('utf8'))
else:
logger_hopt.info((u'**** Real valued unit (intensity taken into consideration)').encode('utf8'))
logger_hopt.info((u'**** Quantization : ' + str(quantization)).encode('utf8'))
# Define hyper-parameter search space for the model
# Those are given by the static methods get_param_dico and get_hp_space
model_space = Model_class.get_hp_space()
optim_space = Optimization_method.get_hp_space()
space = {'model': model_space, 'optim': optim_space}
# Get the headers (i.e. list of hyperparameters tuned for printing and
# save purposes)
header = model_space.keys() + optim_space.keys() + ['accuracy']
global run_counter
run_counter = 0
def run_wrapper(params):
global run_counter
run_counter += 1
logger_hopt.info(('\n').encode('utf8'))
logger_hopt.info((u'#'*40).encode('utf8'))
logger_hopt.info((u'# Config : {}'.format(run_counter)).encode('utf8'))
# Build model and optim dico ################
model_param = params['model']
optim_param = params['optim']
#############################################
# Weights plotted and stored in a folder ####
# Same for generated midi sequences #########
weights_folder = result_folder + '/' + str(run_counter) + '/' + 'weights'
if not os.path.isdir(weights_folder):
os.makedirs(weights_folder)
generated_folder = result_folder + '/' + str(run_counter) + '/generated_sequences'
if not os.path.isdir(generated_folder):
os.makedirs(generated_folder)
model_folder = result_folder + '/' + str(run_counter) + '/model'
if not os.path.isdir(model_folder):
os.makedirs(model_folder)
#############################################
# Load data #################################
time_load_0 = time.time()
piano_train, orchestra_train, train_index, \
piano_valid, orchestra_valid, valid_index, \
piano_test, orchestra_test, test_index, generation_index \
= load_data(LOCAL_SCRATCH + '/Data',
model_param['temporal_order'],
model_param['batch_size'],
binary_unit=binary_unit,
skip_sample=1,
logger_load=logger_load)
time_load_1 = time.time()
logger_load.info('TTT : Loading data took {} seconds'.format(time_load_1-time_load_0))
##############
# visualize_mat(piano_train.get_value(), 'DEBUG', 'piano_train')
# visualize_mat(orchestra_train.get_value(), 'DEBUG', 'orchestra_train')
# visualize_mat(piano_test.get_value(), 'DEBUG', 'piano_test')
# visualize_mat(orchestra_test.get_value(), 'DEBUG', 'orchestra_test')
# visualize_mat(piano_valid.get_value(), 'DEBUG', 'piano_valid')
# visualize_mat(orchestra_valid.get_value(), 'DEBUG', 'orchestra_valid')
##############
# For large datasets
# http://deeplearning.net/software/theano/tutorial/aliasing.html
# use borrow=True (avoid copying the whole matrix) ?
# Load as much as the GPU can handle, train then load other
# part of the dataset using shared_variable.set_value(new_value)
#############################################
# Train #####################################
time_train_0 = time.time()
model, dico_res = train(piano_train, orchestra_train, train_index,
piano_valid, orchestra_valid, valid_index,
model_param, optim_param, max_iter, weights_folder)
time_train_1 = time.time()
logger_train.info('TTT : Training data took {} seconds'.format(time_train_1-time_train_0))
error = -dico_res['accuracy'] # Search for a min
#############################################
# Generate ##################################
time_generate_0 = time.time()
generate(model,
piano_test, orchestra_test, generation_index,
generation_length, seed_size, quantization_write,
generated_folder, logger_generate)
time_generate_1 = time.time()
logger_generate.info('TTT : Generating data took {} seconds'.format(time_generate_1-time_generate_0))
#############################################
# Save ######################################
save_model_file = open(model_folder + '/model.pkl', 'wb')
pickle.dump(model, save_model_file, protocol=pickle.HIGHEST_PROTOCOL)
#############################################
# log
logger_hopt.info((u'# Accuracy : {}'.format(dico_res['accuracy'])).encode('utf8'))
logger_hopt.info((u'###################\n').encode('utf8'))
# Write the result in result.csv
with open(csv_file_path, 'ab') as csvfile:
writer = csv.DictWriter(csvfile, delimiter=',', fieldnames=header)
writer.writerow(dico_res)
return error
# Calling get_param_dico with None return an empty dictionary,
# Useful to get the header of hparam
with open(csv_file_path, 'ab') as csvfile:
# Write headers if they don't already exist
writerHead = csv.writer(csvfile, delimiter=',')
writerHead.writerow(header)
best = fmin(run_wrapper, space, algo=tpe.suggest, max_evals=max_evals)
return best
def train(piano_train, orchestra_train, train_index,
piano_valid, orchestra_valid, valid_index,
model_param, optim_param, max_iter, weights_folder):
############################################################
############################################################
############################################################
# model_param and optim_param are dictionaries
# If you use train directly, bypassing the hparameter loop,
# be careful that the keys match the constructor arguments of both model and optimizer
# Log them
logger_train.info((u'##### Model parameters').encode('utf8'))
for k, v in model_param.iteritems():
logger_train.info((u'# ' + k + ' : {}'.format(v)).encode('utf8'))
logger_train.info((u'##### Optimization parameters').encode('utf8'))
for k, v in optim_param.iteritems():
logger_train.info((u'# ' + k + ' : {}'.format(v)).encode('utf8'))
logger_generate.info((u'##### Generation parameters').encode('utf8'))
logger_generate.info((u'# generation_length : ' + str(generation_length)).encode('utf8'))
logger_generate.info((u'# seed_size : ' + str(seed_size)).encode('utf8'))
logger_generate.info((u'# quantization_write : ' + str(quantization_write)).encode('utf8'))
################################################################
################################################################
################################################################
# DATA
piano_dim = piano_train.get_value().shape[1]
orchestra_dim = orchestra_train.get_value().shape[1]
n_train_batches = len(train_index)
n_val_batches = len(valid_index)
logger_load.info((u'##### Data').encode('utf8'))
logger_load.info((u'# n_train_batch : {}'.format(n_train_batches)).encode('utf8'))
logger_load.info((u'# n_val_batch : {}'.format(n_val_batches)).encode('utf8'))
################################################################
################################################################
################################################################
# MODEL
# dimensions dictionary
dimensions = {'batch_size': model_param['batch_size'],
'temporal_order': model_param['temporal_order'],
'piano_dim': piano_dim,
'orchestra_dim': orchestra_dim}
model = Model_class(model_param, dimensions)
# Define an optimizer
optimizer = Optimization_method(optim_param)
############################################################
############################################################
############################################################
# COMPILE FUNCTIONS
# Compilation of the training function is encapsulated in the class since the 'givens'
# can vary with the model
train_iteration = model.get_train_function(piano_train, orchestra_train, optimizer, name='train_iteration')
# Same for the validation
validation_error = model.get_validation_error(piano_valid, orchestra_valid, name='validation_error')
############################################################
############################################################
############################################################
# TRAINING
logger_train.info("#")
logger_train.info("# Training")
epoch = 0
OVERFITTING = False
DIVERGING = False
val_tab = np.zeros(max(1,max_iter))
while (not OVERFITTING
and not DIVERGING
and epoch!=max_iter):
# go through the training set
train_cost_epoch = []
train_monitor_epoch = []
for batch_index in xrange(n_train_batches):
this_cost, this_monitor = train_iteration(train_index[batch_index])
# Keep track of cost
train_cost_epoch.append(this_cost)
train_monitor_epoch.append(this_monitor)
# Validation
# For binary unit, it's an accuracy measure.
# For real valued units its a gaussian centered value with variance 1
accuracy = []
for batch_index in xrange(n_val_batches):
_, _, accuracy_batch = validation_error(valid_index[batch_index])
accuracy += [accuracy_batch]
# Early stopping criterion
# Note that sum_{i=0}^{n} der = der(n) - der(0)
# So mean over successive derivatives makes no sense
# 1/ Get the mean derivative between 5 and 10 =
# \sum_{i=validation_order}^{validation_order+initial_derivative_length} E(i) - E(i-validation_order) / validation_order
#
# 2/ At each iteration, compare the mean derivative over the last five epochs :
# \sum_{i=0}^{validation_order} E(t)
mean_accuracy = 100 * np.mean(accuracy)
val_tab[epoch] = mean_accuracy
if epoch == initial_derivative_length-1:
ind = np.arange(validation_order-1, initial_derivative_length)
increase_reference = (val_tab[ind] - val_tab[ind-validation_order+1]).sum() / (validation_order * len(ind))
if increase_reference <= 0:
# Early stop if the model didn't really improved over the first iteration
DIVERGING = True
elif epoch >= initial_derivative_length:
ind = np.arange(epoch - check_derivative_length + 1, epoch+1)
derivative_mean = (val_tab[ind] - val_tab[ind-validation_order+1]).sum() / (validation_order * len(ind))
# Mean derivative is less than 10% of increase reference
if derivative_mean < 0.1 * increase_reference:
OVERFITTING = True
# Monitor learning
logger_train.info(('Epoch : {} , Monitor : {} , Cost : {} , Valid acc : {}'
.format(epoch, np.mean(train_monitor_epoch), np.mean(train_cost_epoch), mean_accuracy))
.encode('utf8'))
if DIVERGING:
logger_train.info('DIVERGING !!')
elif OVERFITTING:
logger_train.info('OVERFITTING !!')
# Plot weights every ?? epoch
if((epoch%20==0) or (epoch<5) or OVERFITTING or DIVERGING):
weights_folder_epoch = weights_folder + '/' + str(epoch)
if not os.path.isdir(weights_folder_epoch):
os.makedirs(weights_folder_epoch)
model.save_weights(weights_folder_epoch)
epoch += 1
# Return results
best_accuracy = np.amax(val_tab)
dico_res = model_param
dico_res.update(optim_param)
dico_res['accuracy'] = best_accuracy
return model, dico_res
else:
raise ValueError(sys.argv[2] + " is not an optimization method")
# System paths
MAIN_DIR = os.getcwd().decode('utf8') + u'/'
if 'LSCRATCH' in os.environ.keys():
# We are on Guillimin
LOCAL_SCRATCH = os.environ['LSCRATCH']
else:
LOCAL_SCRATCH = '..'
data_folder = LOCAL_SCRATCH + u'/Data'
if ONLY_BUILD_DB:
result_folder = MAIN_DIR + u'../Results/' + 'Only_build_db'
else:
result_folder = MAIN_DIR + u'../Results/' + temporal_granularity + '/' + Optimization_method.name() + '/' + Model_class.name()
result_file = result_folder + u'/hopt_results.csv'
if ONLY_BUILD_DB:
log_file_path = result_folder + '/build_db.log'
else:
log_file_path = result_folder + '/' + Model_class.name() + u'.log'
# Fixed hyper parameter
max_evals = 50 # number of hyper-parameter configurations evaluated
max_iter = 100 # nb max of iterations when training 1 configuration of hparams
# Config is set now, no need to modify source below for standard use
# Validation
validation_order = 5
initial_derivative_length = 20
check_derivative_length = 5
