import torch from trainer import Trainer from network import WideDeep from criteo_loader import getTestData, getTrainData import torch.utils.data as Data import torch_mlu import torch_mlu.core.mlu_model as ct import torch_mlu.core.mlu_quantize as mlu_quantize import torchvision.models as models import platform  # 计时 import time # ct.set_core_number(1) ct.set_core_version("MLU270") torch.set_grad_enabled(False) import numpy as np mean = [0.485, 0.456, 0.406] # std = [1 / 255] std = [0.229, 0.224, 0.225] widedeep_config = \    {        'deep_dropout': 0,        ' _dim': 8,  # 用于控制稀疏特征经过 ding层后的稠密特征大小        'hidden_ s': [256, 128, 64],        'num_epoch': 10,        'batch_size': 32,        'lr': 1e-3,        'l2_regularization': 1e-4,        'device_id': 0,        'use_cuda': False,        'train_file': '../Data/criteo/processed_data/train_set.csv',        'fea_file': '../Data/criteo/processed_data/fea_col.npy',        'validate_file': '../Data/criteo/processed_data/val_set.csv',        'test_file': '../Data/criteo/processed_data/test_set.csv',        'model_name': '../TrainedModels/WideDeep.model'    } if __name__ == "__main__":    ####################################################################################    # WideDeep 模型    ####################################################################################    training_data, training_label, dense_features_col, sparse_features_col = getTrainData(widedeep_config['train_file'],                                                                                          widedeep_config['fea_file'])    # train_dataset = Data.TensorDataset(torch.tensor(training_data).float(), torch.tensor(training_label).float())    test_data = getTestData(widedeep_config['test_file'])    test_dataset = Data.TensorDataset(torch.tensor(test_data).float())    model = WideDeep(widedeep_config, dense_features_cols=dense_features_col,                        sparse_features_cols=sparse_features_col).cpu()    model_path = "old.pth"    # wideDeep.load_state_dict(torch.load(model_path), False)    # 加载pth预训练模型    pretrained_dict = torch.load(model_path, map_location="cpu")    model_dict = model.state_dict()    # 重新制作预训练的权重，主要是减去参数不匹配的层    pretrained_dict = {k: v for k, v in pretrained_dict.items() if (k in model_dict and 'fc' not in k)}    # 更新权重    model_dict.update(pretrained_dict)    model.load_state_dict(model_dict)    # # 有两个层参数不一样，不能用，切掉    # ckpt = torch.load(model_path)    # # ckpt = torch.load(model_path, map_location='cpu')    # ckpt.pop("0.bias")    # ckpt.pop("0.weight")    # wideDeep.load_state_dict(ckpt, False)    ####################################################################################    # 模型训练阶段    ####################################################################################    # # 实例化模型训练器    # trainer = Trainer(model=wideDeep, config=widedeep_config)    # 训练    # trainer.train(train_dataset)    # 保存模型    # torch.save(wideDeep.state_dict(), "./checkpoint.pth")    # 加载保存的权重    # 模型量化    quantized_model = torch_mlu.core.mlu_quantize.quantize_dynamic_mlu(        model, dtype='int8', gen_quant = True)    #在CPU上进行推理，生成量化值    input_tensor = torch.tensor(test_data).float()    quantized_model(input_tensor)    #保存量化模型    model_path = "quantize_widedeep.pth"    torch.save(quantized_model.state_dict(), model_path)    print("\nSuccessfully Save quantized net\n")    net = mlu_quantize.quantize_dynamic_mlu(model)    net.load_state_dict(torch.load(model_path))    net_mlu = net.to(ct.mlu_device())    print("\n\nSuccessfuuly Load Quantized Model\n\n")    # 转为LongTensor    # input_tensor = input_tensor.cpu()    # input_tensor = torch.LongTensor(input_tensor.numpy())    print("input_tensor.to(ct.mlu_device()) = ", type(input_tensor.to(ct.mlu_device())))    output = net_mlu(input_tensor.to(ct.mlu_device()))    # print("\nSuccessfully Predict quantized net\n")    # net_quantization = mlu_quantize.quantize_dynamic_mlu(    #     wideDeep, {'mean': mean, 'std': std, 'firstconv': True}, dtype='int8', gen_quant=True)    # net_quantization.eval()    # torch.save(net_quantization.state_dict(), 'test_quantization.pth')    #    # net = mlu_quantize.quantize_dynamic_mlu(wideDeep)    # net.load_state_dict(torch.load('test_quantization.pth'))    # net_mlu = net.to(ct.mlu_device())    ####################################################################################    # 模型测试阶段    ####################################################################################    # wideDeep.eval()    # if widedeep_config['use_cuda']:    #     wideDeep.loadModel(map_location=lambda storage, loc: storage.cuda(widedeep_config['device_id']))    #     resNet = wideDeep.cuda()    # else:    #     wideDeep.loadModel(map_location=torch.device('cpu'))    # 融合，生成静态图    # ct.set_core_number(1)    # ct.save_as_cambricon("int8_accuracy")    #    # trace_input = torch.tensor(test_data).float()    #    # # 从Float Tensor 转成Long Tensor    # # trace_input = torch.LongTensor(trace_input.numpy())    #    # trace_input = trace_input.to(ct.mlu_device())    #    # print("\nSuccessfully Trace input\n\n")    # quantized_net = torch.jit.trace(net_mlu, trace_input, check_trace=False)    #    # print("\n\nSuccessfully RongHe\n\n")    # y_pred_probs = quantized_net(trace_input)    print("\n\nSuccessfully output\n\n")    print("Output = ", output, "\t", type(output))    # y_pred = torch.where(y_pred_probs > 0.5, torch.ones_like(y_pred_probs), torch.zeros_like(y_pred_probs))    # print("Test Data CTR Predict...\n ", y_pred.view(-1))    index = torch.argmax(output)    print("Predict = ", output[index])