some questions about keypoints training?

[lunalulu]

When I was training keypoints subnet, I was wondering if there was no training in segment network, according to the following code, the comment code
`
def build_keypoint_loss(saved_for_loss, heat_temp, heat_weight):
names = build_names()
saved_for_log = OrderedDict()
criterion = nn.MSELoss(size_average=True).cuda()
total_loss = 0
div1 = 1.
#div2 = 100.

for j in range(5):

    pred1 = saved_for_loss[j][:, :18, :, :] * heat_weight
    gt1 = heat_weight * heat_temp

    #pred2 = saved_for_loss[j][:, 18:, :, :]
    #gt2 = mask_all

    # Compute losses
    loss1 = criterion(pred1, gt1)/div1  # heatmap_loss
    #loss2 = criterion(pred2, gt2)/div2  # mask_loss
    total_loss += loss1
    #total_loss += loss2

    # Get value from Tensor and save for log
    saved_for_log[names[j*2]] = loss1.item() #只是保留heatmap的loss
    #saved_for_log[names[j*2+1]] = loss2.item()

saved_for_log['max_ht'] = torch.max(
    saved_for_loss[-1].data[:, :18, :, :]).item()
saved_for_log['min_ht'] = torch.min(
    saved_for_loss[-1].data[:, :18, :, :]).item()
#saved_for_log['max_mask'] = torch.max(
#    saved_for_loss[-1].data[:, 18:, :, :]).item()
#saved_for_log['min_mask'] = torch.min(
#    saved_for_loss[-1].data[:, 18:, :, :]).item()

return total_loss, saved_for_log

`

another question is i found when i training on two gpu , it's speed lower than one gpu , log is below:

thanks for your reply

[LiMeng95]

hi,

1. The task Keypoint Estimation Subnet with person segmentation mask is still in to-do list, I have not added it in keypoints subnet yet.

2. The batch-size of dataloader is linearly related to the number of GPUs. Specifically, batch-size will be multiplied by the number of GPUs you enabled in CUDA_VISIBLES_DEVICES. Related code is in multipose_keypoint_train.py#L61.
   That is more GPUs means more pre-processing such as data agumentation, GT generation and so on. All of these pre-processing are on CPUs. So, I think the training speed maybe limited by not only the GPU but the CPUs' computational ability.

[lunalulu]

@LiMeng95 thanks very much~

Another question, I trained a resnet50 as backbone, but compared to ckpt_baseline_resnet101.h5 in your project, the inference FPS went up by only 2 frames.
Do you have training with resnet50? How fast is the result compared with resnet101?Can you provide one resnet50?
thanks

[LiMeng95]

Sorry, I didn't train with resnet50. Maybe the backbone is not the main limitation of inference FPS, other parts like person detection will also take some time.