import torch
import numpy as np
import torch.nn.functional as F
import torch.npu
import os
def get_mask(tscale):
mask = np.zeros([tscale, tscale], np.float32)
for i in range(tscale):
for j in range(i, tscale):
mask[i, j] = 1
return torch.Tensor(mask)
def bmn_loss_func(pred_bm, pred_start, pred_end, gt_iou_map, gt_start, gt_end, bm_mask):
pred_bm_reg = pred_bm[:, 0].contiguous()
pred_bm_cls = pred_bm[:, 1].contiguous()
gt_iou_map = gt_iou_map * bm_mask
pem_reg_loss = pem_reg_loss_func(pred_bm_reg, gt_iou_map, bm_mask)
pem_cls_loss = pem_cls_loss_func(pred_bm_cls, gt_iou_map, bm_mask)
tem_loss = tem_loss_func(pred_start, pred_end, gt_start, gt_end)
loss = tem_loss + 10 * pem_reg_loss + pem_cls_loss
return loss, tem_loss, pem_reg_loss, pem_cls_loss
def tem_loss_func(pred_start, pred_end, gt_start, gt_end):
def bi_loss(pred_score, gt_label):
pred_score = pred_score.view(-1)
gt_label = gt_label.view(-1)
pmask = (gt_label > 0.5).float()
num_entries = len(pmask)
num_positive = torch.sum(pmask)
ratio = num_entries / num_positive
coef_0 = 0.5 * ratio / (ratio - 1)
coef_1 = 0.5 * ratio
epsilon = 0.000001
loss_pos = coef_1 * torch.log(pred_score + epsilon) * pmask
loss_neg = coef_0 * torch.log(1.0 - pred_score + epsilon) * (1.0 - pmask)
loss = -1 * torch.mean(loss_pos + loss_neg)
return loss
loss_start = bi_loss(pred_start, gt_start)
loss_end = bi_loss(pred_end, gt_end)
loss = loss_start + loss_end
if torch.isnan(loss):
return torch.tensor([0.0]).npu()
else:
return loss
def pem_reg_loss_func(pred_score, gt_iou_map, mask):
u_hmask = (gt_iou_map > 0.7).float()
u_mmask = ((gt_iou_map <= 0.7) & (gt_iou_map > 0.3)).float()
u_lmask = ((gt_iou_map <= 0.3) & (gt_iou_map > 0.)).float()
u_lmask = u_lmask * mask
num_h = torch.sum(u_hmask)
num_m = torch.sum(u_mmask)
num_l = torch.sum(u_lmask)
r_m = num_h / num_m
u_smmask = torch.Tensor(np.random.rand(*gt_iou_map.shape)).npu()
u_smmask = u_mmask * u_smmask
u_smmask = (u_smmask > (1. - r_m)).float()
r_l = num_h / num_l
u_slmask = torch.Tensor(np.random.rand(*gt_iou_map.shape)).npu()
u_slmask = u_lmask * u_slmask
u_slmask = (u_slmask > (1. - r_l)).float()
weights = u_hmask + u_smmask + u_slmask
loss = F.mse_loss(pred_score * weights, gt_iou_map * weights)
loss = 0.5 * torch.sum(loss * torch.ones(*weights.shape).npu()) / torch.sum(weights)
if torch.isnan(loss):
return torch.tensor([0.0]).npu()
else:
return loss
def pem_cls_loss_func(pred_score, gt_iou_map, mask):
pmask = (gt_iou_map > 0.9).float()
nmask = (gt_iou_map <= 0.9).float()
nmask = nmask * mask
num_positive = torch.sum(pmask)
num_entries = num_positive + torch.sum(nmask)
ratio = num_entries / num_positive
coef_0 = 0.5 * ratio / (ratio - 1)
coef_1 = 0.5 * ratio
epsilon = 0.000001
loss_pos = coef_1 * torch.log(pred_score + epsilon) * pmask
loss_neg = coef_0 * torch.log(1.0 - pred_score + epsilon) * nmask
loss = -1 * torch.sum(loss_pos + loss_neg) / num_entries
if torch.isnan(loss):
return torch.tensor([0.0]).npu()
else:
return loss