class ImageNetCatewisedTrainer(MiniBatchTrainer): def _finish_init(self): assert len(self.num_caterange_list) == len(self.num_batch) and self.num_caterange_list[-1] == 1000 self.num_batch_list = self.num_batch[1:] self.num_batch = self.num_batch[0] init_output = self.num_caterange_list[0] self.num_caterange_list = self.num_caterange_list[1:] fc = self.init_model[-2] fc['outputSize'] = init_output self.learning_rate_list = self.learning_rate[1:] self.learning_rate = self.learning_rate[0] self.set_category_range(init_output) self.net = FastNet(self.learning_rate, self.image_shape, init_model=self.init_model) MiniBatchTrainer._finish_init(self) def set_category_range(self, r): dp = data.get_by_name(self.data_provider) self.train_dp = dp(self.data_dir, self.train_range, category_range=range(r)) self.test_dp = dp(self.data_dir, self.test_range, category_range=range(r)) def train(self): MiniBatchTrainer.train(self) for i, cate in enumerate(self.num_caterange_list): self.set_category_range(cate) self.curr_batch = self.curr_epoch = 0 self.num_batch = self.num_batch_list[i] model = self.checkpoint_dumper.get_checkpoint() layers = model['layers'] fc = layers[-2] fc['weight'] = None fc['bias'] = None fc['weightIncr'] = None fc['biasIncr'] = None # for l in layers: # if l['type'] == 'fc': # l['weight'] = None # l['bias'] = None # l['weightIncr'] = None # l['biasIncr'] = None # fc = layers[-2] fc['outputSize'] = cate self.learning_rate = self.learning_rate_list[i] self.net = FastNet(self.learning_rate, self.image_shape, init_model=model) self.net.clear_weight_incr() MiniBatchTrainer.train(self)
class ImageNetCateGroupTrainer(MiniBatchTrainer): def _finish_init(self): self.num_batch_list = self.num_batch[1:] self.num_batch = self.num_batch[0] self.learning_rate_list = self.learning_rate[1:] self.learning_rate = self.learning_rate[0] layers = self.init_model fc = layers[-2] fc['outputSize'] = self.num_group_list[0] self.num_group_list = self.num_group_list[1:] self.set_num_group(fc['outputSize']) self.net = FastNet(self.learning_rate, self.image_shape, init_model=self.init_model) MiniBatchTrainer._finish_init(self) def set_num_group(self, n): dp = data.get_by_name(self.data_provider) self.train_dp = dp(self.data_dir, self.train_range, n) self.test_dp = dp(self.data_dir, self.test_range, n) def train(self): MiniBatchTrainer.train(self) for i, group in enumerate(self.num_group_list): self.set_num_group(group) self.curr_batch = self.curr_epoch = 0 self.num_batch = self.num_batch_list[i] model = self.checkpoint_dumper.get_checkpoint() layers = model['layers'] fc = layers[-2] fc['outputSize'] = group fc['weight'] = None fc['bias'] = None fc['weightIncr'] = None fc['biasIncr'] = None self.learning_rate = self.learning_rate_list[i] self.net = FastNet(self.learning_rate, self.image_shape, init_model=model) self.net.clear_weight_incr() MiniBatchTrainer.train(self)