代码链接：https://github.com/mmunhappy/ICASSP2025-PDM

代码迁移

启动脚本里加入torch_npu

python train_pdm_os.py --dataset regdb --gpu 0 --protos 10 --trial 1

vim train_pdm_os.py

from __future__ import print_function

import torch
import torch_npu
from torch_npu.contrib import transfer_to_npu
torch.npu.set_compile_mode(jit_compile=False)
import argparse
import time
import torch.nn as nn
import torch.optim as optim
import torch.backends.cudnn as cudnn
from torch.autograd import Variable
import torch.utils.data as data
import torchvision.transforms as transforms

from dcs_loss import DCSLoss
from data_loader import SYSUData, RegDBData, LLCMData, TestData
from data_manager import *
from eval_metrics import eval_sysu, eval_regdb, eval_llcm
from model_pdm_os import embed_net
from utils import *
from loss import OriTripletLoss, CPMLoss
from tensorboardX import SummaryWriter
from random_erasing import RandomErasing

报错1

Self dtype DT_FLOAT or mat2 dtype DT_FLOAT not support under keep type mode.
分析定位
社区查看算子说明 aclnnMatmul：https://www.hiascend.com/document/detail/zh/CANNCommunityEdition/800alpha003/apiref/aolapi/context/aclnnMatmul.md

true

确认数据类型
打印结果 x.clone().type:torch.float32 x.clone().shape:torch.Size([32, 3, 384, 192])
导出plog日志 export ASCEND_WORK_PATH=/path export ASCEND_GLOBAL_LOG_LEVEL=1 export ASCEND_HOST_LOG_FILE_NUM=1000
原因

当aclnnMatmul参数cubeMathType为0时，输入数据类型不支持FLOAT32
torch.fft.fft2调用aclnnMatmu时，cubeMathType写死为0
解决方案

torch.fft.fft2除GPU外只能使用fp32

torch.fft.fft2使用cpu计算

cubeMathType固定为0原因

报错2
op type TransData is not found in this op store.

分析定位
dump组图 export DUMP_GE_GRAPH=2 export DUMP_GRAPH_LEVEL=2 export ASCEND_GLOBAL_LOG_LEVEL=0 export ASCEND_SLOG_PRINT_TO_STDOUT=1