class ImageNetLayerwisedTrainer(Trainer): def divide_layers_to_stack(self): self.fc_params = [] self.conv_params = [] conv = True for ld in self.init_model: if ld['type'] in ['conv', 'rnorm', 'pool', 'neuron'] and conv: # self.conv_params.append(ld) 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 def initialize_model(self): self.curr_model.extend(self.conv_stack['conv1']) self.curr_model.extend(self.conv_stack['conv2']) self.curr_model.extend(self.conv_stack['conv3']) self.curr_model.extend(self.conv_stack['conv4']) self.curr_model.extend(self.conv_stack['conv5']) self.curr_model.extend(self.fc_tmp) def _finish_init(self): self.final_num_epoch = self.num_epoch self.curr_model = [] self.divide_layers_to_stack() self.conv_stack = FastNet.split_conv_to_stack(self.conv_params) self.fc_stack = FastNet.split_fc_to_stack(self.fc_params) self.fc_tmp = [self.fc_stack['fc8'][0], self.softmax_param] del self.fc_stack['fc8'] self.stack = self.fc_stack self.initialize_model() pprint.pprint(self.stack) self.num_epoch = self.frag_epoch self.net = FastNet(self.learning_rate, self.image_shape, self.curr_model) self.container = deque() def report(self): pass def should_continue_training(self): return self.curr_epoch <= self.num_epoch def init_replaynet_data_provider(self): if self.output_method == 'disk': dp = data.get_by_name('intermediate') count = self.layer_output_dumper.get_count() self.train_dp = dp(self.layer_output_path, range(0, count), 'fc') elif self.output_method == 'memory': dp = data.get_by_name('memory') self.train_dp = dp(self.layer_output_dumper) def train_replaynet(self, stack): self.container.append(self.save_freq) self.container.append(self.test_freq) self.container.append(self.train_dp) self.container.append(self.test_dp) self.container.append(self.layer_output_dumper) self.container.append(self.net) self.save_freq = self.curr_batch + 100 self.test_freq = self.curr_batch + 100 self.curr_batch = self.curr_epoch = 0 self.init_replaynet_data_provider() model = [] model.extend(stack) model.extend(self.fc_tmp) self.layer_output_dumper = None size = self.net['fc8'].get_input_size() image_shape = (size, 1, 1, self.batch_size) self.net = FastNet(self.learning_rate, image_shape, model) self.replaynet = self.net self.num_epoch = self.replaynet_epoch Trainer.train(self) self.net = self.container.pop() self.layer_output_dumper = self.container.pop() self.layer_output_dumper.reset() self.test_dp = self.container.pop() self.train_dp = self.container.pop() self.test_freq = self.container.pop() self.save_freq = self.container.pop() def reset_trainer(self, i): if i == len(self.stack) - 1: self.num_epoch = self.final_num_epoch else: self.num_epoch = self.frag_epoch self.curr_batch = self.curr_epoch = 0 self.init_data_provider() def train(self): Trainer.train(self) for i, stack in enumerate(self.stack.values()): self.train_replaynet(stack) self.reset_trainer(i) self.net.drop_layer_from('fc8') for layer in self.replaynet: if layer.type != 'data': self.net.append_layer(layer) Trainer.train(self)
class ImageNetLayerwisedTrainer(Trainer): def divide_layers_to_stack(self): self.fc_params = [] self.conv_params = [] conv = True for ld in self.init_model: if ld['type'] in ['conv', 'rnorm', 'pool', 'neuron'] and conv: # self.conv_params.append(ld) 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 def initialize_model(self): self.curr_model.extend(self.conv_stack['conv1']) self.curr_model.extend(self.conv_stack['conv2']) self.curr_model.extend(self.conv_stack['conv3']) self.curr_model.extend(self.conv_stack['conv4']) self.curr_model.extend(self.conv_stack['conv5']) self.curr_model.extend(self.fc_tmp) def _finish_init(self): self.final_num_epoch = self.num_epoch self.curr_model = [] self.divide_layers_to_stack() self.conv_stack = FastNet.split_conv_to_stack(self.conv_params) self.fc_stack = FastNet.split_fc_to_stack(self.fc_params) self.fc_tmp = [self.fc_stack['fc8'][0], self.softmax_param] del self.fc_stack['fc8'] self.stack = self.fc_stack self.initialize_model() pprint.pprint(self.stack) self.num_epoch = self.frag_epoch self.net = FastNet(self.learning_rate, self.image_shape, self.curr_model) self.container = deque() def report(self): pass def should_continue_training(self): return self.curr_epoch <= self.num_epoch def init_replaynet_data_provider(self): if self.output_method == 'disk': dp = data.get_by_name('intermediate') count = self.train_layer_output_dumper.get_count() self.train_dp = dp(self.train_layer_output_path, range(0, count), 'fc') count = self.test_layer_output_dumper.get_count() self.test_dp = dp(self.test_layer_output_path, range(count), 'fc') elif self.output_method == 'memory': dp = data.get_by_name('memory') self.train_dp = dp(self.train_layer_output_dumper) self.test_dp = dp(self.test_layer_output_dumper) def train_replaynet(self, stack): self.container.append(self.save_freq) self.container.append(self.test_freq) self.container.append(self.train_dp) self.container.append(self.test_dp) self.container.append(self.train_layer_output_dumper) self.container.append(self.test_layer_output_dumper) self.container.append(self.net) self.save_freq = self.curr_batch + 100 self.test_freq = self.curr_batch + 100 self.curr_batch = self.curr_epoch = 0 self.init_replaynet_data_provider() model = [] model.extend(stack) model.extend(self.fc_tmp) self.train_layer_output_dumper = None self.test_layer_output_dumper = None size = self.net['fc8'].get_input_size() image_shape = (size, 1, 1, self.batch_size) self.net = FastNet(self.learning_rate, image_shape, model) self.replaynet = self.net self.num_epoch = self.replaynet_epoch Trainer.train(self) self.net = self.container.pop() self.test_layer_output_dumper = self.container.pop() self.test_layer_output_dumper.reset() self.train_layer_output_dumper = self.container.pop() self.train_layer_output_dumper.reset() self.test_dp = self.container.pop() self.train_dp = self.container.pop() self.test_freq = self.container.pop() self.save_freq = self.container.pop() def reset_trainer(self, i): if i == len(self.stack) - 1: self.num_epoch = self.final_num_epoch else: self.num_epoch = self.frag_epoch self.curr_batch = self.curr_epoch = 0 self.init_data_provider() def train(self): Trainer.train(self) for i, stack in enumerate(self.stack.values()): self.train_replaynet(stack) self.reset_trainer(i) self.net.drop_layer_from('fc8') for layer in self.replaynet: if layer.type != 'data': self.net.append_layer(layer) Trainer.train(self)