Created by: maosengshulei

  1）PaddlePaddle版本： 1.8.1  2）CPU

模型整体结构类似预GAN，模型中的判别器已加载预训练好的参数，已检查最后一层fc的参数 和 预训练好的参数一致。

但是在模型训练时检查判别器infer的结果发现每次重新执行后预测结果都不一样

代码如下：

class GTrainer():
    def __init__(self, input, config, args, feat_list):
        self.program = fluid.default_main_program().clone()
        with fluid.program_guard(self.program):
            g_model =  PointerNet(config)
            d_model = DNN()

            self.actions, self.logp = g_model.learn(input)

            self.infer_program = self.program.clone()
            self.online_reward = d_model.net(input, args, feat_list, [], rerank_flag=False)
            self.rerank_reward = d_model.net(input, args, feat_list, self.actions, rerank_flag=True)

            self.online_reward.stop_gradient = True
            self.rerank_reward.stop_gradient = True

            self.pi_loss = -1.0 * fluid.layers.reduce_mean(self.logp * self.rerank_reward)
            self.avg_reward = fluid.layers.reduce_mean(self.rerank_reward)
            vars = []
            for var in self.program.list_vars():
                if fluid.io.is_parameter(var) and (var.name.startswith("Actor")):
                    vars.append(var.name)

            print(vars)
            optimizer = fluid.optimizer.AdamOptimizer(1e-4)
            optimizer.minimize(self.pi_loss, parameter_list=vars)
            


class ListwiseRerankModel(object):

    def __init__(self, config, args, feat_list, dataset, param_path):
        self.config = config
        self.feat_list = feat_list
        self.args = args
        self.dataset = dataset
        self.param_path = param_path

    def build_model(self):
        dense_feature_user_dim = reduce(lambda x,y:x+y,[ 1 for _,feat in self.feat_list.items() if feat.value_type == 0 and feat.node_type == 'u'])
        dense_feature_thread_dim = reduce(lambda x,y:x+y,[ 1 for _,feat in self.feat_list.items() if feat.value_type == 0 and feat.node_type == 'i'])
        dense_feature_user = fluid.data(name="dense_input_user", shape=[None, 1, dense_feature_user_dim], dtype='float32')
        dense_feature_thread = fluid.data(name="dense_input_thread", shape=[None, 1, dense_feature_thread_dim], lod_level=1, dtype='float32')

        sparse_feature_user = [fluid.data(name=feat.prefix, shape=[None, 1], lod_level=1, dtype='int64')  for _, feat in self.feat_list.items() if feat.value_type == 1 and feat.node_type == 'u']
        sparse_weight_feature_user = [fluid.data(name=feat.prefix + '@index', shape=[None, 1], lod_level=1, dtype='int64')  for _, feat in self.feat_list.items() if feat.value_type == 2 and feat.node_type == 'u']
        sparse_weight_value_user = [fluid.data(name=feat.prefix + '@value', shape=[None, 1], lod_level=1, dtype='float32')  for _, feat in self.feat_list.items() if feat.value_type == 2 and feat.node_type == 'u']

        sparse_feature_thread = [fluid.data(name=feat.prefix, shape=[None, 1], lod_level=2, dtype='int64')  for _, feat in self.feat_list.items() if feat.value_type == 1 and feat.node_type == 'i']
        sparse_weight_feature_thread = [fluid.data(name=feat.prefix + '@index', shape=[None, 1], lod_level=2, dtype='int64')  for _, feat in self.feat_list.items() if feat.value_type == 2 and feat.node_type == 'i']
        sparse_weight_value_thread = [fluid.data(name=feat.prefix + '@value', shape=[None, 1], lod_level=2, dtype='float32')  for _, feat in self.feat_list.items() if feat.value_type == 2 and feat.node_type == 'i']

        seq_len = fluid.data(name='seq_len', shape=[None, 1], dtype='int64')

        data_list = [dense_feature_user] + [dense_feature_thread] + sparse_feature_user + sparse_weight_feature_user + sparse_weight_value_user + sparse_feature_thread + sparse_weight_feature_thread + sparse_weight_value_thread + [seq_len]
        
        g_trainer = GTrainer(data_list, self.config, self.args, self.feat_list)
        
        place = fluid.CPUPlace()
        exe = fluid.Executor(place)
        exe.run(fluid.default_startup_program())

        load_pretrained_params(exe, g_trainer.program, self.param_path)

        learn_data_loader = fluid.io.DataLoader.from_generator(capacity=64, feed_list=data_list, iterable=True)


        learn_data_loader.set_sample_list_generator(
            self.dataset.batch_reader(50, self.dataset.cluster_data_reader()),
            places=place)

        visualize_parameter('run_train_episode_before', 'all_fc.w_0')
        for pass_id in range(3):
            for data in learn_data_loader():
                online_reward = exe.run(g_trainer.program, feed=data, fetch_list=[g_trainer.online_reward])[0]  
                #print(data)
                print(online_reward)

训练数据没有shuffle，因而每次重新训练时训练数据都是一致的。 代码中的online_reward对应判别器对训练数据的预测结果，但是每次重新训练的结果都不一致。