class ImageNetLayerwisedTrainer(AutoStopTrainer):
def __init__(self, test_id, data_dir, provider, checkpoint_dir, train_range, test_range, test_freq,
save_freq, batch_size, num_epoch, image_size, image_color, learning_rate, params):
self.origin_test_range = test_range
if len(test_range) != 1:
test_range = [test_range[0]]
AutoStopTrainer.__init__(self, test_id, data_dir, provider, checkpoint_dir, train_range, test_range, test_freq,
save_freq, batch_size, num_epoch, image_size, image_color, learning_rate, False)
self.conv_params = []
self.fc_params = []
self.softmax_param = None
self.params = params
conv = True
for ld in self.params:
if ld['type'] in ['conv', 'rnorm', 'pool', 'neuron'] and conv:
self.conv_params.append(ld)
elif ld['type'] == 'fc' or (not conv and ld['type'] == 'neuron'):
self.fc_params.append(ld)
conv = False
else:
self.softmax_param = ld
self.conv_stack = FastNet.split_conv_to_stack(self.conv_params)
self.fc_stack = FastNet.split_fc_to_stack(self.fc_params)
pprint.pprint(self.conv_stack)
pprint.pprint(self.fc_stack)
self.fakefc_param = self.fc_stack[-1][0]
def report(self):
pass
def init_data_provider(self):
self.train_dp = ImageNetDataProvider(self.data_dir, self.train_range)
self.test_dp = ImageNetDataProvider(self.data_dir, self.test_range)
def train(self):
# train conv stack layer by layer
for i, stack in enumerate(self.conv_stack):
if self.checkpoint_file != '':
model = load(self.checkpoint_file)
self.net = FastNet(self.learning_rate, self.image_shape, self.n_out, initModel=model)
# delete softmax layer
self.net.del_layer()
self.net.del_layer()
# for i in range(len(self.fc_params)):
# self.net.del_layer()
self.net.disable_bprop()
layerParam = stack + [self.fakefc_param, self.softmax_param]
self.net.append_layers_from_dict(layerParam)
self.init_data_provider()
self.scheduler.reset()
self.scheduler.set_level(i)
self.test_outputs = []
self.train_output = []
AutoStopTrainer.train(self)
# train fc layer
for i, stack in enumerate(self.fc_stack):
model = load(self.checkpoint_file)
self.net = FastNet(self.learning_rate, self.image_shape, self.n_out, initModel=model)
self.net.del_layer()
self.net.del_layer()
self.net.disable_bprop()
if i == len(self.fc_stack) - 1:
layerParam = stack + [self.softmax_param]
else:
layerParam = stack + [self.fakefc_param, self.softmax_param]
self.net.append_layers_from_dict(layerParam)
self.init_data_provider()
self.scheduler.reset()
self.scheduler.set_level(i)
self.test_outputs = []
self.train_output = []
AutoStopTrainer.train(self)
model = load(self.checkpoint_file)
self.test_id += 1
self.net = FastNet(self.learning_rate, self.image_shape, self.n_out, initModel=model)
self.test_range = self.origin_test_range
self.init_data_provider()
self.scheduler = Scheduler(self)
self.num_epoch /= 2
AutoStopTrainer.train(self)
