deep-head-pose训练是报错：nll_loss_forward_reduce_cuda_kernel_2d: block: [0,0,0], thread: [11,0,0] Asserti

deep-head-pose，训练是报错：nll_loss_forward_reduce_cuda_kernel_2d: block: [0,0,0], thread: [11,0,0] Asserti

源码地址：https://github.com/natanielruiz/deep-head-pose

由于工作时间写的，简化，deep-head-pose在训练模型的时候，基于300W_LP数据集，数据里面是不干净的，导致lable中有-1，理论上headpose，给的标签是0-65。

修改dataset.py ,找到生成标签的地方，修改如下：
# Bin values
bins = np.array(range(-99, 99, 3))
yaw = max(yaw, -99)
pitch = max(pitch, -99)
roll = max(roll, -99)
binned_pose = np.digitize([yaw, pitch, roll], bins)-1