复现深度学习源码时遇到的问题及解决方案

（1）问题：在windows下复现PointRCNN时出现

error LNK2001: 无法解析的外部符号 "public: long * __cdecl at::TensorBase::data_ptr<long>(void)const " (??$data_ptr@J@TensorBase@at@@QEBAPEAJXZ)

build\lib.win-amd64-cpython-39\iou3d_cuda.cp39-win_amd64.pyd : fatal error LNK1120: 1 个无法解析的外部命令

解决：

第一步：在对应头文件里加入：#include <cstdint>

第二步：将long替换成int64_t，将unsigned long long 替换成uint64_t

（2）问题：在windows下复现PointRCNN时出现

src/iou3d.cpp(102): error C2131: 表达式的计算结果不是常数

解决：

将uint64_t remv_cpu[col_blocks]  替换成

uint64_t *remv_cpu = new uint64_t [col_blocks];

（3） 问题：error C3861: “AT_CHECK”: 找不到标识符

                     error C3861: “THCState_getCurrentStream”: 找不到标识符

解决：将AT_CHECK替换成TORCH_CHECK 

将THCState_getCurrentStream(state)替换成c10::cuda::getCurrentCUDAStream()

(4)问题： while cur_gt[k].sum() == 0:

解决：参考OpenPCDet/pcdet/models/roi_heads/target_assigner/proposal_target_layer.py 

在proposal_target_layer.py文件中，

将 while cur_gt[k].sum() == 0: 替换成while k>=0 and cur_gt[k].sum() == 0:

在加上cur_gt = cur_gt.new_zeros((1, cur_gt.shape[1])) if len(cur_gt) == 0 else cur_gt

(5)

import tensorflow as tf 替换成
import tensorflow._api.v2.compat.v1 as tf
tf.disable_v2_behavior()

(6) h5py错误

  File "h5py\_objects.pyx", line 55, in h5py._objects.with_phil.wrapper
  File "h5py\h5f.pyx", line 106, in h5py.h5f.open

方案：

将h5_fout = h5py.File(h5_filename)替换成 h5_fout = h5py.File(str(h5_filename),'w')

参考：Missing file error message misleadingly turned to lowercase · Issue #854 · h5py/h5py · GitHub

(6)tensorflow 出现问题：AttributeError: module 'tf' has no attribute 'contrib'

解决：

将initializer = tf.contrib.layers.xavier_initializer()

替换成
initializer = tf.compat.v1.keras.initializers.VarianceScaling(scale=1.0, mode="fan_avg", distribution="uniform")