Created by: ShadowSkyLiu

训练网络最后一层为负采样nce，预测替换成input * nce_w^T + nce_b，保存和load模型的时候出现问题。

1. 训练保存模型参数指定filename，load_persistables之后维度不对 训练保存： with open(model_name, "wb") as f: f.write(main_program.desc.serialize_to_string()) fluid.io.save_persistables(exe, save_dirname, main_program, filename=save_path) 预测加载： print("nce_w shape before", fluid.global_scope().find_var("nce_w").get_tensor().shape()) fluid.io.load_persistables(exe, save_dirname, main_program, filename=params_filename) print("nce_w shape after", fluid.global_scope().find_var("nce_w").get_tensor().shape()) ... 打印结果： nce_w shape before [21391L, 31L] nce_w shape after [1L]

2. 训练保存模型参数不指定filename，显示加载之后数据维度没有问题 训练保存： with open(model_name, "wb") as f: f.write(main_program.desc.serialize_to_string()) fluid.io.save_persistables(exe, save_dirname, main_program) #, filename=save_path) 预测加载： print("nce_w shape before", fluid.global_scope().find_var("nce_w").get_tensor().shape()) fluid.io.load_persistables(exe, save_dirname, main_program) #, filename=params_filename) print("nce_w shape after", fluid.global_scope().find_var("nce_w").get_tensor().shape()) ... 打印结果： nce_w shape before [21391L, 31L] nce_w shape after [21391L, 31L]

ps:附加网络最后一层定义： def net(...): ... if is_train: cost = fluid.layers.nce( input=fc, label=next_id, num_total_classes=nid_dict_size, param_attr=fluid.ParamAttr(name="nce_w"), bias_attr=fluid.ParamAttr(name="nce_b"), num_neg_samples=5) return cost, fc else:
w_var = fluid.layers.create_global_var( shape=[nid_dict_size, 31], value=1.0, dtype='float32', persistable=True, name='nce_w') b_var = fluid.layers.create_global_var( shape=[nid_dict_size, 1], value=1.0, dtype='float32', persistable=True, name='nce_b') mul_var = fluid.layers.matmul(fc, w_var, transpose_y=True) sum_var = fluid.layers.sums(mul_var, b_var) prediction_layer = fluid.layers.softmax(sum_var) return prediction_layer, fc