class Trainer:
CHECKPOINT_REGEX = None
def __init__(self, test_id, data_dir, data_provider, checkpoint_dir, train_range, test_range, test_freq, save_freq, batch_size, num_epoch, image_size,
image_color, learning_rate, auto_init=False, init_model=None, adjust_freq=1, factor=1.0):
self.test_id = test_id
self.data_dir = data_dir
self.data_provider = data_provider
self.checkpoint_dir = checkpoint_dir
self.train_range = train_range
self.test_range = test_range
self.test_freq = test_freq
self.save_freq = save_freq
self.batch_size = batch_size
self.num_epoch = num_epoch
self.image_size = image_size
self.image_color = image_color
self.learning_rate = learning_rate
# doesn't matter anymore
self.n_out = 10
self.factor = factor
self.adjust_freq = adjust_freq
self.regex = re.compile('^test%d-(\d+)\.(\d+)$' % self.test_id)
self.init_data_provider()
self.image_shape = (self.batch_size, self.image_color, self.image_size, self.image_size)
if init_model is not None and 'model_state' in init_model:
self.train_outputs = init_model['model_state']['train_outputs']
self.test_outputs = init_model['model_state']['test_outputs']
else:
self.train_outputs = []
self.test_outputs = []
self.curr_minibatch = self.num_batch = self.curr_epoch = self.curr_batch = 0
self.net = FastNet(self.learning_rate, self.image_shape, self.n_out, init_model=init_model)
self.train_data = None
self.test_data = None
self.num_train_minibatch = 0
self.num_test_minibatch = 0
self.checkpoint_file = ''
self.train_dumper = None #DataDumper('/scratch1/imagenet-pickle/train-data.pickle')
self.test_dumper = None #DataDumper('/scratch1/imagenet-pickle/test-data.pickle')
self.input = None
def init_data_provider(self):
dp = DataProvider.get_by_name(self.data_provider)
self.train_dp = dp(self.data_dir, self.train_range)
self.test_dp = dp(self.data_dir, self.test_range)
def get_next_minibatch(self, i, train=TRAIN):
if train == TRAIN:
data = self.train_data
else:
data = self.test_data
batch_data = data.data
batch_label = data.labels
batch_size = self.batch_size
mini_data = batch_data[:, i * batch_size: (i + 1) * batch_size]
locked_data = driver.pagelocked_empty(mini_data.shape, mini_data.dtype, order='C',
mem_flags=driver.host_alloc_flags.PORTABLE)
locked_data[:] = mini_data
if self.input is not None and locked_data.shape == self.input.shape:
self.input.set(locked_data)
else:
self.input = gpuarray.to_gpu(locked_data)
label = batch_label[i * batch_size : (i + 1) * batch_size]
#label = gpuarray.to_gpu(label)
#label = gpuarray.to_gpu(np.require(batch_label[i * batch_size : (i + 1) * batch_size], dtype =
# np.float, requirements = 'C'))
return self.input, label
def save_checkpoint(self):
model = {}
model['batchnum'] = self.train_dp.get_batch_num()
model['epoch'] = self.num_epoch + 1
model['layers'] = self.net.get_dumped_layers()
model['train_outputs'] = self.train_outputs
model['test_outputs'] = self.test_outputs
dic = {'model_state': model, 'op':None}
self.print_net_summary()
if not os.path.exists(self.checkpoint_dir):
os.system('mkdir -p \'%s\'' % self.checkpoint_dir)
saved_filename = [f for f in os.listdir(self.checkpoint_dir) if self.regex.match(f)]
for f in saved_filename:
os.remove(os.path.join(self.checkpoint_dir, f))
checkpoint_filename = "test%d-%d.%d" % (self.test_id, self.curr_epoch, self.curr_batch)
checkpoint_file_path = os.path.join(self.checkpoint_dir, checkpoint_filename)
self.checkpoint_file = checkpoint_file_path
print >> sys.stderr, checkpoint_file_path
with open(checkpoint_file_path, 'w') as f:
cPickle.dump(dic, f, protocol=-1)
util.log('save file finished')
def get_test_error(self):
start = time.time()
self.test_data = self.test_dp.get_next_batch()
self.num_test_minibatch = divup(self.test_data.data.shape[1], self.batch_size)
for i in range(self.num_test_minibatch):
input, label = self.get_next_minibatch(i, TEST)
self.net.train_batch(input, label, TEST)
self._capture_test_data()
cost , correct, numCase, = self.net.get_batch_information()
self.test_outputs += [({'logprob': [cost, 1 - correct]}, numCase, time.time() - start)]
print >> sys.stderr, '[%d] error: %f logreg: %f time: %f' % (self.test_data.batchnum, 1 - correct, cost, time.time() - start)
def print_net_summary(self):
print >> sys.stderr, '--------------------------------------------------------------'
for s in self.net.get_summary():
name = s[0]
values = s[1]
print >> sys.stderr, "Layer '%s' weight: %e [%e]" % (name, values[0], values[1])
print >> sys.stderr, "Layer '%s' bias: %e [%e]" % (name, values[2], values[3])
def should_continue_training(self):
return self.curr_epoch <= self.num_epoch
def check_test_data(self):
return self.num_batch % self.test_freq == 0
def check_save_checkpoint(self):
return self.num_batch % self.save_freq == 0
def check_adjust_lr(self):
return self.num_batch % self.adjust_freq == 0
def _finished_training(self):
if self.train_dumper is not None:
self.train_dumper.flush()
if self.test_dumper is not None:
self.test_dumper.flush()
def _capture_training_data(self):
if not self.train_dumper:
return
self.train_dumper.add({'labels' : self.net.label.get(),
'fc' : self.net.outputs[-3].get().transpose() })
def _capture_test_data(self):
if not self.test_dumper:
return
self.test_dumper.add({'labels' : self.net.label.get(),
'fc' : self.net.outputs[-3].get().transpose() })
def train(self):
self.print_net_summary()
util.log('Starting training...')
while self.should_continue_training():
self.train_data = self.train_dp.get_next_batch() # self.train_dp.wait()
self.curr_epoch = self.train_data.epoch
self.curr_batch = self.train_data.batchnum
start = time.time()
self.num_train_minibatch = divup(self.train_data.data.shape[1], self.batch_size)
t = 0
for i in range(self.num_train_minibatch):
input, label = self.get_next_minibatch(i)
stime = time.time()
self.net.train_batch(input, label)
self._capture_training_data()
t += time.time() - stime
self.curr_minibatch += 1
cost , correct, numCase = self.net.get_batch_information()
self.train_outputs += [({'logprob': [cost, 1 - correct]}, numCase, time.time() - start)]
print >> sys.stderr, '%d.%d: error: %f logreg: %f time: %f' % (self.curr_epoch, self.curr_batch, 1 - correct, cost, time.time() - start)
self.num_batch += 1
if self.check_test_data():
print >> sys.stderr, '---- test ----'
self.get_test_error()
print >> sys.stderr, '------------'
if self.factor != 1.0 and self.check_adjust_lr():
print >> sys.stderr, '---- adjust learning rate ----'
self.net.adjust_learning_rate(self.factor)
print >> sys.stderr, '--------'
if self.check_save_checkpoint():
print >> sys.stderr, '---- save checkpoint ----'
self.save_checkpoint()
print >> sys.stderr, '------------'
wait_time = time.time()
#print 'waitting', time.time() - wait_time, 'secs to load'
#print 'time to train a batch file is', time.time() - start)
self.get_test_error()
self.save_checkpoint()
self.report()
self._finished_training()
def predict(self, save_layers = None, filename = None):
self.net.save_layerouput(save_layers)
self.print_net_summary()
util.log('Starting predict...')
save_output = []
while self.curr_epoch < 2:
start = time.time()
self.test_data = self.test_dp.get_next_batch()
self.curr_epoch = self.test_data.epoch
self.curr_batch = self.test_data.batchnum
self.num_test_minibatch = divup(self.test_data.data.shape[1], self.batch_size)
for i in range(self.num_test_minibatch):
input, label = self.get_next_minibatch(i, TEST)
self.net.train_batch(input, label, TEST)
cost , correct, numCase = self.net.get_batch_information()
print >> sys.stderr, '%d.%d: error: %f logreg: %f time: %f' % (self.curr_epoch, self.curr_batch, 1 - correct, cost, time.time() - start)
if save_layers is not None:
save_output.extend(self.net.get_save_output())
if save_layers is not None:
if filename is not None:
with open(filename, 'w') as f:
cPickle.dump(save_output, f, protocol = -1)
util.log('save layer output finished')
def report(self):
rep = self.net.get_report()
if rep is not None:
print rep
