def train_net(config):
theano.config.on_unused_input = 'warn'
image_sz = 227
x = T.ftensor4('x')
y = T.lvector('y')
## BUILD NETWORK ##
model = AlexNet(config)
layers = model.layers
batch_size = model.batch_size
errors = model.errors
## COMPILE FUNCTIONS ##
(train_model, validate_model, train_error, learning_rate, shared_x,
shared_y, rand_arr, vels) = compile_models(model, config, flag_top_5=True)
images = np.random.rand(batch_size, 3, image_sz,
image_sz).astype(np.float32)
labels = np.random.randint(0, 1000, size=batch_size).astype(np.int32)
shared_x.set_value(images)
shared_y.set_value(labels)
time_theano_run(validate_model, 'Forward')
time_theano_run(train_model, 'Forward-Backward')
def time_alexnet():
config = Config()
# Build network
model = AlexNet(config)
batch_size = model.batch_size
# Compile forward and forward-backward functions
(forward_step, forward_backward_step, shared_x, shared_y) = \
compile_models(model, config)
num_warmup_iters = config.num_warmup_iters
num_timing_iters = config.num_timing_iters
forward_times = np.zeros(num_timing_iters)
forward_backward_times = np.zeros(num_timing_iters)
num_iterations = num_warmup_iters + num_timing_iters
for minibatch_index in range(num_iterations):
if num_warmup_iters <= minibatch_index < num_warmup_iters + num_timing_iters:
forward_time, forward_backward_time = \
time_model(
forward_step, forward_backward_step, shared_x,
shared_y, config, batch_size)
forward_times[minibatch_index - num_warmup_iters] = forward_time
forward_backward_times[minibatch_index-num_warmup_iters] = \
forward_backward_time
else: # Do not measure time for the warmup iterations.
time_model(forward_step, forward_backward_step, shared_x, shared_y,
config, batch_size)
# print forward_times
# print forward_backward_times
print("forward time %.4f +- %.4f ms batch_size %d" %
(np.mean(forward_times), np.std(forward_times), batch_size))
print("gradient computation time %.4f +- %.4f ms batch_size %d" %
(np.mean(forward_backward_times), np.std(forward_backward_times),
batch_size))
def train_net(config):
# UNPACK CONFIGS
(flag_para_load, train_filenames, val_filenames, train_labels, val_labels,
img_mean) = unpack_configs(config)
# pycuda set up
drv.init()
dev = drv.Device(int(config['gpu'][-1]))
ctx = dev.make_context()
if flag_para_load:
# zmq set up
sock = zmq.Context().socket(zmq.PAIR)
sock.connect('tcp://*****:*****@ iter = ', num_iter
print 'training cost:', cost_ij
if config['print_train_error']:
print 'training error rate:', train_error()
if flag_para_load and (count < len(minibatch_range)):
load_send_queue.put('calc_finished')
############### Test on Validation Set ##################
DropoutLayer.SetDropoutOff()
this_validation_error, this_validation_loss = get_val_error_loss(
rand_arr,
shared_x,
shared_y,
val_filenames,
val_labels,
flag_para_load,
img_mean,
batch_size,
validate_model,
send_queue=load_send_queue,
recv_queue=load_recv_queue)
print('epoch %i: validation loss %f ' % (epoch, this_validation_loss))
print('epoch %i: validation error %f %%' %
(epoch, this_validation_error * 100.))
val_record.append([this_validation_error, this_validation_loss])
np.save(config['weights_dir'] + 'val_record.npy', val_record)
DropoutLayer.SetDropoutOn()
############################################
# Adapt Learning Rate
step_idx = adjust_learning_rate(config, epoch, step_idx, val_record,
learning_rate)
# Save weights
if epoch % config['snapshot_freq'] == 0:
save_weights(layers, config['weights_dir'], epoch)
np.save(config['weights_dir'] + 'lr_' + str(epoch) + '.npy',
learning_rate.get_value())
save_momentums(vels, config['weights_dir'], epoch)
print('Optimization complete.')
def validate_performance(config):
# UNPACK CONFIGS
(flag_para_load, train_filenames, val_filenames,
train_labels, val_labels, img_mean) = unpack_configs(config)
if flag_para_load:
# pycuda and zmq set up
drv.init()
dev = drv.Device(int(config['gpu'][-1]))
ctx = dev.make_context()
sock = zmq.Context().socket(zmq.PAIR)
sock.connect('tcp://localhost:{0}'.format(config['sock_data']))
load_send_queue = config['queue_t2l']
load_recv_queue = config['queue_l2t']
else:
load_send_queue = None
load_recv_queue = None
import theano.sandbox.cuda
theano.sandbox.cuda.use(config['gpu'])
import theano
theano.config.on_unused_input = 'warn'
from layers import DropoutLayer
from alex_net import AlexNet, compile_models
import theano.misc.pycuda_init
import theano.misc.pycuda_utils
# # BUILD NETWORK ##
model = AlexNet(config)
layers = model.layers
batch_size = model.batch_size
# # COMPILE FUNCTIONS ##
(train_model, validate_model, train_error, learning_rate,
shared_x, shared_y, rand_arr, vels) = compile_models(model, config,
flag_top_5=True)
print '... training'
if flag_para_load:
# pass ipc handle and related information
gpuarray_batch = theano.misc.pycuda_utils.to_gpuarray(
shared_x.container.value)
h = drv.mem_get_ipc_handle(gpuarray_batch.ptr)
sock.send_pyobj((gpuarray_batch.shape, gpuarray_batch.dtype, h))
load_send_queue.put(img_mean)
load_epoch = config['load_epoch']
load_weights(layers, config['weights_dir'], load_epoch)
DropoutLayer.SetDropoutOff()
this_validation_error, this_validation_error_top_5, this_validation_loss = \
get_val_error_loss(rand_arr, shared_x, shared_y,
val_filenames, val_labels,
flag_para_load,img_mean,
batch_size, validate_model,
send_queue=load_send_queue,
recv_queue=load_recv_queue,
flag_top_5=True)
print('validation error %f %%' %
(this_validation_error * 100.))
print('top 5 validation error %f %%' %
(this_validation_error_top_5 * 100.))
print('validation loss %f ' %
(this_validation_loss))
return this_validation_error, this_validation_loss
## BUILD NETWORK ##
model = AlexNet(config)
layers = model.layers
batch_size = model.batch_size
## LOAD DATASET.
with Timer('loading training data'):
train_set = load_lmdb(train_file, img_mean)
with Timer('loading test data'):
test_set = load_lmdb(test_file, img_mean)
## COMPILE FUNCTIONS ##
(train_model, validate_model, train_error, learning_rate,
shared_x, shared_y, rand_arr, vels) = compile_models(model, config)
######################### TRAIN MODEL ################################
print '... training'
# Start Training Loop
epoch = 0
step_idx = 0
val_record = []
while epoch < config['n_epochs']:
epoch = epoch + 1
if config['resume_train'] and epoch == 1:
def validate_performance(config):
# UNPACK CONFIGS
(flag_para_load, train_filenames, val_filenames, train_labels, val_labels,
img_mean) = unpack_configs(config)
if flag_para_load:
# pycuda and zmq set up
drv.init()
dev = drv.Device(int(config['gpu'][-1]))
ctx = dev.make_context()
sock = zmq.Context().socket(zmq.PAIR)
sock.connect('tcp://localhost:{0}'.format(config['sock_data']))
load_send_queue = config['queue_t2l']
load_recv_queue = config['queue_l2t']
else:
load_send_queue = None
load_recv_queue = None
import theano.sandbox.cuda
theano.sandbox.cuda.use(config['gpu'])
import theano
theano.config.on_unused_input = 'warn'
from layers import DropoutLayer
from alex_net import AlexNet, compile_models
import theano.misc.pycuda_init
import theano.misc.pycuda_utils
# # BUILD NETWORK ##
model = AlexNet(config)
layers = model.layers
batch_size = model.batch_size
# # COMPILE FUNCTIONS ##
(train_model, validate_model, train_error, learning_rate, shared_x,
shared_y, rand_arr, vels) = compile_models(model, config, flag_top_5=True)
print '... training'
if flag_para_load:
# pass ipc handle and related information
gpuarray_batch = theano.misc.pycuda_utils.to_gpuarray(
shared_x.container.value)
h = drv.mem_get_ipc_handle(gpuarray_batch.ptr)
sock.send_pyobj((gpuarray_batch.shape, gpuarray_batch.dtype, h))
load_send_queue.put(img_mean)
load_epoch = config['load_epoch']
load_weights(layers, config['weights_dir'], load_epoch)
DropoutLayer.SetDropoutOff()
this_validation_error, this_validation_error_top_5, this_validation_loss = \
get_val_error_loss(rand_arr, shared_x, shared_y,
val_filenames, val_labels,
flag_para_load,img_mean,
batch_size, validate_model,
send_queue=load_send_queue,
recv_queue=load_recv_queue,
flag_top_5=True)
print('validation error %f %%' % (this_validation_error * 100.))
print('top 5 validation error %f %%' %
(this_validation_error_top_5 * 100.))
print('validation loss %f ' % (this_validation_loss))
return this_validation_error, this_validation_loss
def train_net(config):
# UNPACK CONFIGS
(flag_para_load, train_filenames, val_filenames,
train_labels, val_labels, img_mean) = unpack_configs(config)
# pycuda set up
drv.init()
dev = drv.Device(int(config['gpu'][-1]))
ctx = dev.make_context()
if flag_para_load:
# zmq set up
sock = zmq.Context().socket(zmq.PAIR)
sock.connect('tcp://*****:*****@ iter = ', num_iter
print 'training cost:', cost_ij
if config['print_train_error']:
print 'training error rate:', train_error()
if flag_para_load and (count < len(minibatch_range)):
load_send_queue.put('calc_finished')
############### Test on Validation Set ##################
DropoutLayer.SetDropoutOff()
this_validation_error, this_validation_loss = get_val_error_loss(
rand_arr, shared_x, shared_y,
val_filenames, val_labels,
flag_para_load, img_mean,
batch_size, validate_model,
send_queue=load_send_queue, recv_queue=load_recv_queue)
print('epoch %i: validation loss %f ' %
(epoch, this_validation_loss))
print('epoch %i: validation error %f %%' %
(epoch, this_validation_error * 100.))
val_record.append([this_validation_error, this_validation_loss])
np.save(config['weights_dir'] + 'val_record.npy', val_record)
DropoutLayer.SetDropoutOn()
############################################
# Adapt Learning Rate
step_idx = adjust_learning_rate(config, epoch, step_idx,
val_record, learning_rate)
# Save weights
if epoch % config['snapshot_freq'] == 0:
save_weights(layers, config['weights_dir'], epoch)
np.save(config['weights_dir'] + 'lr_' + str(epoch) + '.npy',
learning_rate.get_value())
save_momentums(vels, config['weights_dir'], epoch)
print('Optimization complete.')
def train_net(config):
# UNPACK CONFIGS
(flag_para_load, train_filenames, val_filenames,
train_labels, val_labels, img_mean) = unpack_configs(config)
if flag_para_load:
# zmq set up
sock = zmq.Context().socket(zmq.PAIR)
sock.connect('tcp://*****:*****@ iter = %i" % (num_iter))
logger.info("training cost: %lf" % (cost_ij))
if config['print_train_error']:
logger.info('training error rate: %lf' % train_error())
if flag_para_load and (count < len(minibatch_range)):
load_send_queue.put('calc_finished')
############### Test on Validation Set ##################
#"""
DropoutLayer.SetDropoutOff()
result_list = get_val_error_loss(
rand_arr, shared_x, shared_y,
val_filenames, val_labels,
flag_para_load, img_mean,
batch_size, validate_model,
send_queue=load_send_queue,
recv_queue=load_recv_queue,
flag_top_5=flag_top5)
logger.info(('epoch %i: validation loss %f ' %
(epoch, result_list[-1])))
if flag_top5:
logger.info(('epoch %i: validation error (top 1) %f %%, (top5) %f %%' %
(epoch, result_list[0] * 100., result_list[1] * 100.)))
else:
logger.info(('epoch %i: validation error %f %%' %
(epoch, result_list[0] * 100.)))
val_record.append(result_list)
np.save(config['weights_dir'] + 'val_record.npy', val_record)
DropoutLayer.SetDropoutOn()
############################################
# Adapt Learning Rate
step_idx = adjust_learning_rate(config, epoch, step_idx,
val_record, learning_rate)
# Save weights
if epoch % config['snapshot_freq'] == 0:
save_weights(layers, config['weights_dir'], epoch)
np.save(config['weights_dir'] + 'lr_' + str(epoch) + '.npy',
learning_rate.get_value())
save_momentums(vels, config['weights_dir'], epoch)
#"""
print('Optimization complete.')
def train_net(config, private_config):
# UNPACK CONFIGS
(flag_para_load, flag_datalayer, train_filenames, val_filenames,
train_labels, val_labels, img_mean) = \
unpack_configs(config, ext_data=private_config['ext_data'],
ext_label=private_config['ext_label'])
gpu_send_queue = private_config['queue_gpu_send']
gpu_recv_queue = private_config['queue_gpu_recv']
# pycuda and zmq set up
drv.init()
dev = drv.Device(int(private_config['gpu'][-1]))
ctx = dev.make_context()
sock_gpu = zmq.Context().socket(zmq.PAIR)
if private_config['flag_client']:
sock_gpu.connect('tcp://*****:*****@ iter = ', num_iter
print 'training cost:', cost_ij
if config['print_train_error']:
error_ij = train_error()
gpu_send_queue.put(error_ij)
that_error = gpu_recv_queue.get()
error_ij = (error_ij + that_error) / 2.
if private_config['flag_verbose']:
print 'training error rate:', error_ij
if flag_para_load and (count < len(minibatch_range)):
load_send_queue.put('calc_finished')
############### Test on Validation Set ##################
DropoutLayer.SetDropoutOff()
this_val_error, this_val_loss = get_val_error_loss(
rand_arr,
shared_x,
shared_y,
val_filenames,
val_labels,
flag_datalayer,
flag_para_load,
batch_size,
validate_model,
send_queue=load_send_queue,
recv_queue=load_recv_queue)
# report validation stats
gpu_send_queue.put(this_val_error)
that_val_error = gpu_recv_queue.get()
this_val_error = (this_val_error + that_val_error) / 2.
gpu_send_queue.put(this_val_loss)
that_val_loss = gpu_recv_queue.get()
this_val_loss = (this_val_loss + that_val_loss) / 2.
if private_config['flag_verbose']:
print('epoch %i: validation loss %f ' % (epoch, this_val_loss))
print('epoch %i: validation error %f %%' %
(epoch, this_val_error * 100.))
val_record.append([this_val_error, this_val_loss])
if private_config['flag_save']:
np.save(config['weights_dir'] + 'val_record.npy', val_record)
DropoutLayer.SetDropoutOn()
############################################
# Adapt Learning Rate
step_idx = adjust_learning_rate(config, epoch, step_idx, val_record,
learning_rate)
# Save Weights, only one of them will do
if private_config['flag_save']:
if epoch % config['snapshot_freq'] == 0:
save_weights(layers, config['weights_dir'], epoch)
np.save(config['weights_dir'] + 'lr_' + str(epoch) + '.npy',
learning_rate.get_value())
save_momentums(vels, config['weights_dir'], epoch)
print('Optimization complete.')
def train_net(config):
# UNPACK CONFIGS
(flag_para_load, train_filenames, val_filenames,
train_labels, val_labels, img_mean) = unpack_configs(config)
if flag_para_load:
# zmq set up
sock = zmq.Context().socket(zmq.PAIR)
sock.connect('tcp://*****:*****@iter " + str(count)
if count == 1:
s = time.time()
if count == 20:
e = time.time()
print "time per 20 iter:", (e - s)
logger.info("time per 20 iter: %f" % (e - s))
cost_ij = train_model_wrap(train_model, shared_x,
shared_y, rand_arr, img_mean,
count, minibatch_index,
minibatch_range, batch_size,
train_filenames, train_labels,
flag_para_load,
config['batch_crop_mirror'],
send_queue=load_send_queue,
recv_queue=load_recv_queue)
if num_iter % config['print_freq'] == 0:
#print 'training @ iter = ', num_iter
#print 'training cost:', cost_ij
logger.info("training @ iter = %i" % (num_iter))
logger.info("training cost: %lf" % (cost_ij))
if config['print_train_error']:
logger.info('training error rate: %lf' % train_error())
#print 'training error rate:', train_error()
if flag_para_load and (count < len(minibatch_range)):
load_send_queue.put('calc_finished')
############### Test on Validation Set ##################
#"""
DropoutLayer.SetDropoutOff()
# result_list = [ this_validation_error, this_validation_error_top5, this_validation_loss ]
# or
# result_list = [ this_validation_error, this_validation_loss ]
result_list = get_val_error_loss(
#this_validation_error, this_validation_loss = get_val_error_loss(
rand_arr, shared_x, shared_y,
val_filenames, val_labels,
flag_para_load, img_mean,
batch_size, validate_model,
send_queue=load_send_queue,
recv_queue=load_recv_queue,
flag_top_5=flag_top5)
logger.info(('epoch %i: validation loss %f ' %
(epoch, result_list[-1])))
#print('epoch %i: validation loss %f ' %
# (epoch, this_validation_loss))
if flag_top5:
logger.info(('epoch %i: validation error (top 1) %f %%, (top5) %f %%' %
(epoch, result_list[0] * 100., result_list[1] * 100.)))
else:
logger.info(('epoch %i: validation error %f %%' %
(epoch, result_list[0] * 100.)))
#print('epoch %i: validation error %f %%' %
# (epoch, this_validation_error * 100.))
val_record.append(result_list)
#val_record.append([this_validation_error, this_validation_loss])
np.save(config['weights_dir'] + 'val_record.npy', val_record)
DropoutLayer.SetDropoutOn()
############################################
# Adapt Learning Rate
step_idx = adjust_learning_rate(config, epoch, step_idx,
val_record, learning_rate)
# Save weights
if epoch % config['snapshot_freq'] == 0:
save_weights(layers, config['weights_dir'], epoch)
np.save(config['weights_dir'] + 'lr_' + str(epoch) + '.npy',
learning_rate.get_value())
save_momentums(vels, config['weights_dir'], epoch)
#"""
print('Optimization complete.')
def train_net(config, private_config):
# UNPACK CONFIGS
(flag_para_load, train_filenames, val_filenames,
train_labels, val_labels, img_mean) = \
unpack_configs(config, ext_data=private_config['ext_data'],
ext_label=private_config['ext_label'])
gpu_send_queue = private_config['queue_gpu_send']
gpu_recv_queue = private_config['queue_gpu_recv']
# pycuda and zmq set up
drv.init()
dev = drv.Device(int(private_config['gpu'][-1]))
ctx = dev.make_context()
sock_gpu = zmq.Context().socket(zmq.PAIR)
if private_config['flag_client']:
sock_gpu.connect('tcp://*****:*****@ iter = ', num_iter
print 'training cost:', cost_ij
if config['print_train_error']:
error_ij = train_error()
gpu_send_queue.put(error_ij)
that_error = gpu_recv_queue.get()
error_ij = (error_ij + that_error) / 2.
if private_config['flag_verbose']:
print 'training error rate:', error_ij
if flag_para_load and (count < len(minibatch_range)):
load_send_queue.put('calc_finished')
if count%20 == 0:
e = time.time()
print "time per 20 iter:", (e - s)
############### Test on Validation Set ##################
DropoutLayer.SetDropoutOff()
this_val_error, this_val_loss = get_val_error_loss(
rand_arr, shared_x, shared_y,
val_filenames, val_labels,
flag_para_load, img_mean,
batch_size, validate_model,
send_queue=load_send_queue, recv_queue=load_recv_queue)
# report validation stats
gpu_send_queue.put(this_val_error)
that_val_error = gpu_recv_queue.get()
this_val_error = (this_val_error + that_val_error) / 2.
gpu_send_queue.put(this_val_loss)
that_val_loss = gpu_recv_queue.get()
this_val_loss = (this_val_loss + that_val_loss) / 2.
if private_config['flag_verbose']:
print('epoch %i: validation loss %f ' %
(epoch, this_val_loss))
print('epoch %i: validation error %f %%' %
(epoch, this_val_error * 100.))
val_record.append([this_val_error, this_val_loss])
if private_config['flag_save']:
np.save(config['weights_dir'] + 'val_record.npy', val_record)
DropoutLayer.SetDropoutOn()
############################################
# Adapt Learning Rate
step_idx = adjust_learning_rate(config, epoch, step_idx,
val_record, learning_rate)
# Save Weights, only one of them will do
if private_config['flag_save']:
if epoch % config['snapshot_freq'] == 0:
save_weights(layers, config['weights_dir'], epoch)
np.save(config['weights_dir'] + 'lr_' + str(epoch) + '.npy',
learning_rate.get_value())
save_momentums(vels, config['weights_dir'], epoch)
print('Optimization complete.')
## BUILD NETWORK ##
model = AlexNet(config)
layers = model.layers
batch_size = model.batch_size
## LOAD DATASET.
with Timer('loading training data'):
train_set = load_lmdb(train_file, img_mean)
with Timer('loading test data'):
test_set = load_lmdb(test_file, img_mean)
## COMPILE FUNCTIONS ##
(train_model, validate_model, train_error, learning_rate, shared_x, shared_y,
rand_arr, vels) = compile_models(model, config)
######################### TRAIN MODEL ################################
print '... training'
# Start Training Loop
epoch = 0
step_idx = 0
val_record = []
while epoch < config['n_epochs']:
epoch = epoch + 1
if config['resume_train'] and epoch == 1:
load_epoch = config['load_epoch']
load_weights(layers, config['weights_dir'], load_epoch)