584 lines
25 KiB
Python
584 lines
25 KiB
Python
|
#!/usr/bin/env python
|
|||
|
|
# -*- coding: utf-8 -*-
|
|||
|
|
"""
|
|||
|
|
@project:
|
|||
|
|
@File : farsegloss
|
|||
|
|
@Author : qiqq
|
|||
|
|
@create_time : 2022/11/25 14:54
|
|||
|
|
farseg论文里的F-A模块对于loss的优化
|
|||
|
|
"""
|
|||
|
|
|
|||
|
|
import torch
|
|||
|
|
import torch.nn.functional as F
|
|||
|
|
import math
|
|||
|
|
from torch import nn
|
|||
|
|
|
|||
|
|
|
|||
|
|
def softmax_focalloss(y_pred, y_true, ignore_index=255, gamma=2.0, normalize=False):
|
|||
|
|
"""
|
|||
|
|
|
|||
|
|
Args:
|
|||
|
|
y_pred: [N, #class, H, W]
|
|||
|
|
y_true: [N, H, W] from 0 to #class
|
|||
|
|
gamma: scalar
|
|||
|
|
|
|||
|
|
Returns:
|
|||
|
|
|
|||
|
|
"""
|
|||
|
|
losses = F.cross_entropy(y_pred, y_true, ignore_index=ignore_index, reduction='none')
|
|||
|
|
with torch.no_grad():##.你tm的为啥torch.no_grad()这还怎么在训练的时候用 ？？难道是为了节省显存？？我不理解
|
|||
|
|
p = y_pred.softmax(dim=1)
|
|||
|
|
modulating_factor = (1 - p).pow(gamma) #权重
|
|||
|
|
valid_mask = ~ y_true.eq(ignore_index)
|
|||
|
|
masked_y_true = torch.where(valid_mask, y_true, torch.zeros_like(y_true)) # 我觉得这个可能还是为了去选择那些不是ignorindex的去参与计算
|
|||
|
|
#不对把，你tm把255的也弄成0了，tm这不就和背景一样了吗
|
|||
|
|
'''
|
|||
|
|
torch.where(condition，a，b)其中
|
|||
|
|
输入参数condition：条件限制，如果满足条件，则选择a，否则选择b作为输出
|
|||
|
|
'''
|
|||
|
|
#这个是干啥的..
|
|||
|
|
modulating_factor = torch.gather(modulating_factor, dim=1, index=masked_y_true.unsqueeze(dim=1)).squeeze_(dim=1)
|
|||
|
|
'''
|
|||
|
|
torch.gather:
|
|||
|
|
index实际上是索引，具体是行还是列的索引要看前面dim 的指定,index的大小就是输出的大小
|
|||
|
|
举个例子：
|
|||
|
|
输入:[[1,2,3],[4,5,6] index [[0 ,1],[2 ,0]] dim=1 输出：[[1 ,2],[6,4]]
|
|||
|
|
gather在one-hot为输出的多分类问题中，可以把最大值坐标作为index传进去，然后提取到每一行的正确预测结果，这也是gather可能的一个作用。
|
|||
|
|
|
|||
|
|
|
|||
|
|
在这里dim1就是类别方向比如batch,channel,h,w dim1就是channel方向
|
|||
|
|
torch.gather相当于以类别为索引把modulating_factor对应的取出来
|
|||
|
|
这里的
|
|||
|
|
'''
|
|||
|
|
scale = 1.
|
|||
|
|
if normalize:
|
|||
|
|
scale = losses.sum() / (losses * modulating_factor).sum()
|
|||
|
|
losses = scale * (losses * modulating_factor).sum() / (valid_mask.sum() + p.size(0)) # focalloss里的对于困难样本的加权
|
|||
|
|
#我不太明白为什么这里/的时候嗨哟啊+batch 这个valid_mask.sum()不久包含了batch在内的所有的样本点吗
|
|||
|
|
|
|||
|
|
return losses
|
|||
|
|
|
|||
|
|
|
|||
|
|
def cosine_annealing(lower_bound, upper_bound, _t, _t_max):
|
|||
|
|
'''
|
|||
|
|
|
|||
|
|
'''
|
|||
|
|
return upper_bound + 0.5 * (lower_bound - upper_bound) * (math.cos(math.pi * _t / _t_max) + 1)
|
|||
|
|
|
|||
|
|
|
|||
|
|
def poly_annealing(lower_bound, upper_bound, _t, _t_max):
|
|||
|
|
factor = (1 - _t / _t_max) ** 0.9
|
|||
|
|
return upper_bound + factor * (lower_bound - upper_bound)
|
|||
|
|
|
|||
|
|
|
|||
|
|
def linear_annealing(lower_bound, upper_bound, _t, _t_max):
|
|||
|
|
factor = 1 - _t / _t_max
|
|||
|
|
return upper_bound + factor * (lower_bound - upper_bound)
|
|||
|
|
|
|||
|
|
|
|||
|
|
def annealing_softmax_focalloss(y_pred, y_true, t, t_max, ignore_index=255, gamma=2.0,
|
|||
|
|
annealing_function=cosine_annealing):
|
|||
|
|
losses = F.cross_entropy(y_pred, y_true, ignore_index=ignore_index, reduction='none')
|
|||
|
|
with torch.no_grad():
|
|||
|
|
p = y_pred.softmax(dim=1)
|
|||
|
|
modulating_factor = (1 - p).pow(gamma)
|
|||
|
|
valid_mask = ~ y_true.eq(ignore_index) # 没有的话这个可以忽略
|
|||
|
|
masked_y_true = torch.where(valid_mask, y_true, torch.zeros_like(y_true))
|
|||
|
|
modulating_factor = torch.gather(modulating_factor, dim=1, index=masked_y_true.unsqueeze(dim=1)).squeeze_(dim=1)
|
|||
|
|
# 截止到这里还是和focalloss那边差不多
|
|||
|
|
normalizer = losses.sum() / (losses * modulating_factor).sum() # 大概就是论文种那个z 其实这个/换成h*w就成了mean loss了
|
|||
|
|
scales = modulating_factor * normalizer # modulating_factor对应原文的那个（1-pi）r 这个scales对应原文的（1/z）*（1-pi）r
|
|||
|
|
if t > t_max: # 大概的意思是整个训练过程比如1000轮 t_max=500轮 就是500以内我的factor就是不断呈现一个余弦函数变化，然后500-100轮 我的factor就固定了
|
|||
|
|
scale = scales
|
|||
|
|
else:
|
|||
|
|
scale = annealing_function(1, scales, t,
|
|||
|
|
t_max) # 比起一般的focalloss从头到位一直是modulating_factor我这里是动态加权从1开始一直到modulating_factor
|
|||
|
|
losses = (losses * scale).sum() / (valid_mask.sum() + p.size(0))
|
|||
|
|
return losses
|
|||
|
|
|
|||
|
|
|
|||
|
|
class OhemCELoss(nn.Module):
|
|||
|
|
"""
|
|||
|
|
Online hard example mining cross-entropy loss:在线难样本挖掘
|
|||
|
|
if loss[self.n_min] > self.thresh: 最少考虑 n_min 个损失最大的 pixel，
|
|||
|
|
如果前 n_min 个损失中最小的那个的损失仍然大于设定的阈值，
|
|||
|
|
那么取实际所有大于该阈值的元素计算损失:loss=loss[loss>thresh]。
|
|||
|
|
否则，计算前 n_min 个损失:loss = loss[:self.n_min]
|
|||
|
|
"""
|
|||
|
|
|
|||
|
|
def __init__(self, thresh, n_min, ignore_lb=255, *args, **kwargs):
|
|||
|
|
super(OhemCELoss, self).__init__()
|
|||
|
|
self.thresh = -torch.log(torch.tensor(thresh, dtype=torch.float)).cuda() # 将输入的概率 转换为loss值
|
|||
|
|
self.n_min = n_min
|
|||
|
|
self.ignore_lb = ignore_lb
|
|||
|
|
self.criteria = nn.CrossEntropyLoss(ignore_index=ignore_lb, reduction='none') # 交叉熵
|
|||
|
|
|
|||
|
|
def forward(self, logits, labels):
|
|||
|
|
N, C, H, W = logits.size()
|
|||
|
|
loss = self.criteria(logits, labels).view(-1)
|
|||
|
|
loss, _ = torch.sort(loss, descending=True) # 排序
|
|||
|
|
if loss[self.n_min] > self.thresh: # 当loss大于阈值(由输入概率转换成loss阈值)的像素数量比n_min多时，取所以大于阈值的loss值
|
|||
|
|
loss = loss[loss > self.thresh]
|
|||
|
|
else:
|
|||
|
|
loss = loss[:self.n_min]
|
|||
|
|
return torch.mean(loss)
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
def setup_seed(seed=0):
|
|||
|
|
import torch
|
|||
|
|
import os
|
|||
|
|
import numpy as np
|
|||
|
|
import random
|
|||
|
|
torch.manual_seed(seed) # 为CPU设置随机种子
|
|||
|
|
np.random.seed(seed) # Numpy module.
|
|||
|
|
random.seed(seed) # Python random module.
|
|||
|
|
############################################################################
|
|||
|
|
#2022.12.8我感觉废了
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
#########################以下作废，准备重写
|
|||
|
|
|
|||
|
|
|
|||
|
|
####从此处开始往下到main 有关自己损失函数优化的全部作废......
|
|||
|
|
def sampleloss(y_pred, y_true, ignore_index=255, gamma=2.0, normalize=False):
|
|||
|
|
#
|
|||
|
|
# 在定义标签的时候背景必须是0
|
|||
|
|
# 想法：loss分成两部分：
|
|||
|
|
# 所有前景样本的loss+部分困难背景样本的loss,然后背景像素的调制系数在参照farseg
|
|||
|
|
# 其中困难的定义是prob<0.5 (是个超参数，暂定)
|
|||
|
|
# y_pred: [N, #class, H, W]
|
|||
|
|
# y_true: [N, H, W] from 0 to #class
|
|||
|
|
#
|
|||
|
|
# :return:
|
|||
|
|
#
|
|||
|
|
pass
|
|||
|
|
|
|||
|
|
# #1.首先把所有前景样本取出来
|
|||
|
|
# losses = F.cross_entropy(y_pred, y_true, ignore_index=ignore_index, reduction='none')
|
|||
|
|
# with torch.no_grad():
|
|||
|
|
# p = y_pred.softmax(dim=1)
|
|||
|
|
# modulating_factor = (1 - p).pow(gamma) #
|
|||
|
|
# valid_mask = ~ y_true.eq(ignore_index)
|
|||
|
|
# valid_mask = ~ y_true.eq(ignore_index)
|
|||
|
|
# masked_y_true = torch.where(valid_mask, y_true, torch.zeros_like(y_true)) #
|
|||
|
|
# modulating_factor = torch.gather(modulating_factor, dim=1, index=masked_y_true.unsqueeze(dim=1)).squeeze_(dim=1)
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
#2.把符合条件的背景样本取出来
|
|||
|
|
|
|||
|
|
|
|||
|
|
def sample_foreground(y_pred, y_true, ignore_index=255,backgroud=0, gamma=2.0, normalize=False):
|
|||
|
|
# 把所有前景样本取出来。形成专门的前景loss
|
|||
|
|
losses = F.cross_entropy(y_pred, y_true, ignore_index=ignore_index, reduction='none') #这一部分的loss是把前景和背景的loss都给计算出来了
|
|||
|
|
with torch.no_grad():
|
|||
|
|
p = y_pred.softmax(dim=1)
|
|||
|
|
mask1 = torch.ones_like(y_true) #
|
|||
|
|
# valid_mask1 = ~ y_true.eq(ignore_index) #ignor的像素找出来
|
|||
|
|
# valid_mask2= ~ y_true.eq(backgroud) #背景像素找出来
|
|||
|
|
# # valid_mask=valid_mask1.eq(valid_mask2)
|
|||
|
|
valid_mask=~ y_true.eq(ignore_index) & ~ y_true.eq(backgroud) #背景和ignor的像素的位置设置为false
|
|||
|
|
|
|||
|
|
foregroundcount= valid_mask[valid_mask==True].shape[0] #统计前景像素的个数
|
|||
|
|
|
|||
|
|
|
|||
|
|
foregmask = torch.where(valid_mask, mask1, torch.zeros_like(y_true)) #
|
|||
|
|
|
|||
|
|
|
|||
|
|
losses = (losses * foregmask).sum() / (foregroundcount ) # #这一部分的loss是把前景和背景的loss都给计算出来了。。我们只要前景的所以背景的就是
|
|||
|
|
# 我不太明白为什么这里原来/的时候要 p.size(0)（+batch ）这个valid_mask.sum()不久包含了batch在内的所有的样本点吗...所以我把原来的+ p.size(0)去掉了
|
|||
|
|
|
|||
|
|
|
|||
|
|
# foregmask是前景的就是1 不是前景的就是0，这样就把原来loss所有的loss中的前景的loss提取出来了
|
|||
|
|
|
|||
|
|
return losses
|
|||
|
|
|
|||
|
|
def sample_hardbackgroundversion1(y_pred, y_true, ignore_index=255,threshold=0.5, backgroud=0,gamma=2.0, normalize=False):
|
|||
|
|
|
|||
|
|
# 采样困难的背景像素。设置小于阈值的就算是困难的 .整个训练过程中一直保持0.5不变
|
|||
|
|
|
|||
|
|
|
|||
|
|
# 把所有符合阈值的背景像素。形成专门的背景loss
|
|||
|
|
losses = F.cross_entropy(y_pred, y_true, ignore_index=ignore_index,
|
|||
|
|
reduction='none') # 这一部分的loss是把前景和背景的loss都给计算出来了
|
|||
|
|
|
|||
|
|
with torch.no_grad():
|
|||
|
|
|
|||
|
|
p = y_pred.softmax(dim=1)
|
|||
|
|
mask = torch.ones_like(y_true)
|
|||
|
|
valid_mask = ~ y_true.eq(ignore_index) & y_true.eq(backgroud) # 背景和ignor的像素的位置设置为false，背景的设置为true
|
|||
|
|
threshold_mask =p[:,0,:,:]<threshold #，在预测的背景里边大于阈值的就属于容易，所以不要，小于阈值的是困难
|
|||
|
|
|
|||
|
|
backgroundmask = torch.where(valid_mask, mask, torch.zeros_like(y_true)) #
|
|||
|
|
backgroundthresholdmask = torch.where(threshold_mask, backgroundmask, torch.zeros_like(y_true)) #
|
|||
|
|
|
|||
|
|
# backgroundloss = losses * backgroundmask
|
|||
|
|
background_thresholdloss = losses * backgroundthresholdmask
|
|||
|
|
count=backgroundthresholdmask[backgroundthresholdmask==1].shape[0]
|
|||
|
|
|
|||
|
|
finalloss =background_thresholdloss.sum()/count
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
# backgroundmask 不是背景就是1，不是背景的（前景和ignor的）就是0，这样就把原来所有的loss中的背景的loss提取出来了
|
|||
|
|
|
|||
|
|
return finalloss
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
def sample_hardbackgroundversion2(y_pred, y_true, t, t_max,ignore_index=255,backgroud=0, annealing_function=cosine_annealing):
|
|||
|
|
'''
|
|||
|
|
采样困难的背景像素。设置小于阈值的就算是困难的
|
|||
|
|
想仿照farseg做一个动态阈值
|
|||
|
|
'''
|
|||
|
|
# 把所有符合阈值的背景像素。形成专门的背景loss
|
|||
|
|
|
|||
|
|
losses = F.cross_entropy(y_pred, y_true, ignore_index=ignore_index,
|
|||
|
|
reduction='none') # 这一部分的loss是把前景和背景的loss都给计算出来了
|
|||
|
|
|
|||
|
|
with torch.no_grad():
|
|||
|
|
|
|||
|
|
p = y_pred.softmax(dim=1)
|
|||
|
|
mask1 = torch.ones_like(y_true) #
|
|||
|
|
|
|||
|
|
valid_mask = ~ y_true.eq(ignore_index) & y_true.eq(backgroud) # 背景和ignor的像素的位置设置为false
|
|||
|
|
if t > t_max: # 大概的意思是整个训练过程比如1000轮 t_max=500轮 就是500以内我的factor就是不断呈现一个余弦函数变化，然后500-100轮 我的factor就固定了
|
|||
|
|
threshold = 0.5
|
|||
|
|
else:
|
|||
|
|
threshold = annealing_function(0.8, 0.5, t,
|
|||
|
|
t_max)
|
|||
|
|
threshold_mask = p[:, 0, :, :] < threshold
|
|||
|
|
|
|||
|
|
backgroundmask = torch.where(valid_mask, mask1, torch.zeros_like(y_true)) #
|
|||
|
|
backgroundthresholdmask = torch.where(threshold_mask, backgroundmask, torch.zeros_like(y_true)) #
|
|||
|
|
|
|||
|
|
# backgroundloss = losses * backgroundmask
|
|||
|
|
background_thresholdloss = losses * backgroundthresholdmask
|
|||
|
|
count = backgroundthresholdmask[backgroundthresholdmask == 1].shape[0]
|
|||
|
|
|
|||
|
|
finalloss = background_thresholdloss.sum() / count
|
|||
|
|
|
|||
|
|
return finalloss
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
def foreground_sample_hardbackgroundversion1(y_pred, y_true, ignore_index=255,threshold=0.5, backgroud=0,gamma=2.0, normalize=False):
|
|||
|
|
'''我还就不信了。分开写不行合起来：在做判断规则的时候直接一部到位，前景的和符合条件的背景的'''
|
|||
|
|
# 把所有前景样本取出来。形成专门的前景loss
|
|||
|
|
losses = F.cross_entropy(y_pred, y_true, ignore_index=ignore_index,
|
|||
|
|
reduction='none') # 这一部分的loss是把前景和背景的loss都给计算出来了
|
|||
|
|
with torch.no_grad():
|
|||
|
|
p = y_pred.softmax(dim=1)
|
|||
|
|
mask1 = torch.ones_like(y_true) #
|
|||
|
|
#1.前景的mask
|
|||
|
|
fore_mask = ~ y_true.eq(ignore_index) & ~ y_true.eq(backgroud) #
|
|||
|
|
foregmask = torch.where(fore_mask, mask1, torch.zeros_like(y_true)) #
|
|||
|
|
|
|||
|
|
#2.符合条件的背景的mask
|
|||
|
|
|
|||
|
|
#背景mask
|
|||
|
|
background_mask = ~ y_true.eq(ignore_index) & y_true.eq(backgroud)
|
|||
|
|
backgroundmask = torch.where(background_mask, mask1, torch.zeros_like(y_true)) #
|
|||
|
|
|
|||
|
|
#背景中符合条件的mask
|
|||
|
|
threshold_mask = p[:, 0, :, :] < threshold # ，
|
|||
|
|
backgroundthresholdmask = torch.where(threshold_mask, backgroundmask, torch.zeros_like(y_true)) #
|
|||
|
|
#
|
|||
|
|
|
|||
|
|
#3.前景的mask和符合条件的背景的mask
|
|||
|
|
totmask=foregmask.eq(1) | backgroundthresholdmask.eq(1)
|
|||
|
|
total_mask = torch.where(totmask, mask1, torch.zeros_like(y_true))
|
|||
|
|
valcount= total_mask[total_mask==1].shape[0]
|
|||
|
|
|
|||
|
|
finaloss = (losses * total_mask).sum() / (valcount)
|
|||
|
|
|
|||
|
|
return finaloss
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
def foreground_sample_hardbackgroundversion2(y_pred, y_true, t, t_max,ignore_index=255,backgroud=0, annealing_function=cosine_annealing):
|
|||
|
|
'''我还就不信了。分开写不行合起来：在做判断规则的时候直接一部到位，前景的和符合条件的背景的'''
|
|||
|
|
# 把所有前景样本取出来。形成专门的前景loss
|
|||
|
|
losses = F.cross_entropy(y_pred, y_true, ignore_index=ignore_index,
|
|||
|
|
reduction='none') # 这一部分的loss是把前景和背景的loss都给计算出来了
|
|||
|
|
with torch.no_grad():
|
|||
|
|
p = y_pred.softmax(dim=1)
|
|||
|
|
mask1 = torch.ones_like(y_true) #
|
|||
|
|
#1.前景的mask
|
|||
|
|
fore_mask = ~ y_true.eq(ignore_index) & ~ y_true.eq(backgroud) #
|
|||
|
|
foregmask = torch.where(fore_mask, mask1, torch.zeros_like(y_true)) #
|
|||
|
|
|
|||
|
|
#2.符合条件的背景的mask
|
|||
|
|
|
|||
|
|
#背景mask
|
|||
|
|
background_mask = ~ y_true.eq(ignore_index) & y_true.eq(backgroud)
|
|||
|
|
backgroundmask = torch.where(background_mask, mask1, torch.zeros_like(y_true)) #
|
|||
|
|
|
|||
|
|
#背景中符合条件的mask
|
|||
|
|
if t > t_max: # 大概的意思是整个训练过程比如1000轮 t_max=500轮 就是500以内我的factor就是不断呈现一个余弦函数变化，然后500-100轮 我的factor就固定了
|
|||
|
|
threshold = 0.5
|
|||
|
|
else:
|
|||
|
|
threshold = annealing_function(0.99, 0.5, t,
|
|||
|
|
t_max)
|
|||
|
|
|
|||
|
|
threshold_mask = p[:, 0, :, :] < threshold # ，
|
|||
|
|
backgroundthresholdmask = torch.where(threshold_mask, backgroundmask, torch.zeros_like(y_true)) #
|
|||
|
|
#
|
|||
|
|
|
|||
|
|
#3.前景的mask和符合条件的背景的mask
|
|||
|
|
totmask=foregmask.eq(1) | backgroundthresholdmask.eq(1)
|
|||
|
|
total_mask = torch.where(totmask, mask1, torch.zeros_like(y_true))
|
|||
|
|
valcount= total_mask[total_mask==1].shape[0]
|
|||
|
|
|
|||
|
|
finaloss = (losses * total_mask).sum() / (valcount)
|
|||
|
|
|
|||
|
|
return finaloss
|
|||
|
|
|
|||
|
|
|
|||
|
|
###########################################################
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
##关于损失函数优化的###########重写
|
|||
|
|
'''
|
|||
|
|
1.第一版本的思想
|
|||
|
|
对于前背景不平衡的问题：
|
|||
|
|
前景像素全部参与计算 ，背景只选择困难的参与计算，
|
|||
|
|
背景的困难样本选择：设置一个阈值 比如0.5 预测的背景概率小于0.5的就算是困难的
|
|||
|
|
阈值设置为动态阈值用一个cos函数 前x个epoch 从0.99到0.5，x epoch后固定为0.5（有点farseg的思想）
|
|||
|
|
|
|||
|
|
2.第二版本的思想
|
|||
|
|
关于困难样本挖掘的问题：
|
|||
|
|
基本和focal loss差不多
|
|||
|
|
1.先根据第一版的把前景样本和符合阈值的背景样本采出来
|
|||
|
|
2.然后用focal loss里边的关于困难样本挖掘的（1-p）r对所有的进行加权,其中加权方式也可以按照动态加权
|
|||
|
|
'''
|
|||
|
|
|
|||
|
|
|
|||
|
|
def fb_loss_version1(y_pred, y_true, t, t_max,ignore_index=255,backgroud=0, gamma=2.0, annealing_function=cosine_annealing):
|
|||
|
|
'''这个基本代码没什么问题了，可以不动了'''
|
|||
|
|
losses = F.cross_entropy(y_pred, y_true, ignore_index=ignore_index,
|
|||
|
|
reduction='none') # 这一部分的loss是把前景和背景的loss都给计算出来了
|
|||
|
|
with torch.no_grad():
|
|||
|
|
p = y_pred.softmax(dim=1)
|
|||
|
|
mask = torch.ones_like(y_true) # #通用的mask
|
|||
|
|
|
|||
|
|
fore_mask = ~ y_true.eq(ignore_index) & ~ y_true.eq(backgroud) # 形成前景的mask（boole的） #非背景和ignore的就是前景像素
|
|||
|
|
foregmask = torch.where(fore_mask, mask, torch.zeros_like(y_true)) # #形成前景的mask（0 1形式的
|
|||
|
|
|
|||
|
|
# 2.背景中符合阈值的
|
|||
|
|
|
|||
|
|
background_mask = ~ y_true.eq(ignore_index) & y_true.eq(backgroud) #背景的mask boole类型
|
|||
|
|
backgroundmask = torch.where(background_mask, mask, torch.zeros_like(y_true)) # #背景的mask int (0,1)类型
|
|||
|
|
|
|||
|
|
# 背景中符合条件的mask
|
|||
|
|
if t > t_max: # 大概的意思是整个训练过程比如1000轮 t_max=500轮 就是500以内我的factor就是不断呈现一个余弦函数变化，然后500-100轮 我的factor就固定了
|
|||
|
|
threshold = 0.5
|
|||
|
|
else:
|
|||
|
|
threshold = annealing_function(0.99, 0.5, t,
|
|||
|
|
t_max)
|
|||
|
|
|
|||
|
|
threshold_mask = p[:, 0, :, :] < threshold # ，像素预测为背景的概率小于阈值的则为背景困难的 #boole类型的
|
|||
|
|
backgroundthresholdmask = torch.where(threshold_mask, backgroundmask, torch.zeros_like(y_true)) # int 0 1 类型的 #最终的符合阈值的background的样本
|
|||
|
|
#
|
|||
|
|
|
|||
|
|
# 3.前景的mask和符合条件的背景的mask（是前景或者是符合要求的背景） #如果摊上了所有背景样本都符合阈值的情况那就是原来的loss
|
|||
|
|
totmask = foregmask.eq(1) | backgroundthresholdmask.eq(1)
|
|||
|
|
total_mask = torch.where(totmask, mask, torch.zeros_like(y_true))
|
|||
|
|
valcount = total_mask[total_mask == 1].shape[0]
|
|||
|
|
|
|||
|
|
finaloss = (losses * total_mask).sum() / (valcount)
|
|||
|
|
|
|||
|
|
return finaloss
|
|||
|
|
|
|||
|
|
|
|||
|
|
def fb_loss_version2(y_pred, y_true, t, t_max,ignore_index=255,backgroud=0, gamma=2.0, annealing_function=cosine_annealing):
|
|||
|
|
losses = F.cross_entropy(y_pred, y_true, ignore_index=ignore_index,
|
|||
|
|
reduction='none') # 这一部分的loss是把前景和背景的loss都给计算出来了
|
|||
|
|
with torch.no_grad():
|
|||
|
|
p = y_pred.softmax(dim=1)
|
|||
|
|
mask = torch.ones_like(y_true) # #通用的mask
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
fore_mask = ~ y_true.eq(ignore_index) & ~ y_true.eq(backgroud) # 形成前景的mask（boole的） #非背景和ignore的就是前景像素
|
|||
|
|
foregmask = torch.where(fore_mask, mask, torch.zeros_like(y_true)) # #形成前景的mask（0 1形式的
|
|||
|
|
|
|||
|
|
# 2.背景中符合阈值的
|
|||
|
|
|
|||
|
|
background_mask = ~ y_true.eq(ignore_index) & y_true.eq(backgroud) # 背景的mask boole类型
|
|||
|
|
backgroundmask = torch.where(background_mask, mask, torch.zeros_like(y_true)) # #背景的mask int (0,1)类型
|
|||
|
|
|
|||
|
|
# 背景中符合条件的mask
|
|||
|
|
if t > t_max: # 大概的意思是整个训练过程比如1000轮 t_max=500轮 就是500以内我的factor就是不断呈现一个余弦函数变化，然后500-100轮 我的factor就固定了
|
|||
|
|
threshold = 0.5
|
|||
|
|
else:
|
|||
|
|
threshold = annealing_function(0.99, 0.5, t,
|
|||
|
|
t_max)
|
|||
|
|
|
|||
|
|
threshold_mask = p[:, 0, :, :] < threshold # ，像素预测为背景的概率小于阈值的则为背景困难的 #boole类型的
|
|||
|
|
backgroundthresholdmask = torch.where(threshold_mask, backgroundmask,
|
|||
|
|
torch.zeros_like(y_true)) # int 0 1 类型的 #最终的符合阈值的background的样本
|
|||
|
|
#
|
|||
|
|
|
|||
|
|
# 3.前景的mask和符合条件的背景的mask（是前景或者是符合要求的背景） #如果摊上了所有背景样本都符合阈值的情况那就是原来的loss
|
|||
|
|
totmask = foregmask.eq(1) | backgroundthresholdmask.eq(1)
|
|||
|
|
total_mask = torch.where(totmask, mask, torch.zeros_like(y_true))
|
|||
|
|
valcount = total_mask[total_mask == 1].shape[0]
|
|||
|
|
|
|||
|
|
|
|||
|
|
#在version1的基础上
|
|||
|
|
|
|||
|
|
modulating_factor = (1 - p).pow(gamma) #
|
|||
|
|
valid_mask = ~ y_true.eq(ignore_index)
|
|||
|
|
masked_y_true = torch.where(valid_mask, y_true, torch.zeros_like(y_true))
|
|||
|
|
'''
|
|||
|
|
torch.where(condition，a，b)其中
|
|||
|
|
输入参数condition：条件限制，如果满足条件，则选择a，否则选择b作为输出
|
|||
|
|
'''
|
|||
|
|
modulating_factor = torch.gather(modulating_factor, dim=1, index=masked_y_true.unsqueeze(dim=1)).squeeze_(dim=1)
|
|||
|
|
'''
|
|||
|
|
modulating_factor=(1 - p).pow(gamma)是个全通道的调制系数 和ypred的尺寸一样 但是我们想要的是一个单通道的
|
|||
|
|
也就是每个pix的一个权重，也就是仙子我们的每个有效的位置的pix的loss已经有了 就差权重了
|
|||
|
|
torch.gather(modulating_factor, dim=1, index=masked_y_true.unsqueeze(dim=1)).squeeze_(dim=1)
|
|||
|
|
就是根据gt的index索引，从相应的dim维度去取对应的factor的值，比如某个位置 gt是2 那就从dim上的第二个位置取他的factor
|
|||
|
|
'''
|
|||
|
|
|
|||
|
|
finaloss = (modulating_factor*(losses * total_mask)).sum() / (valcount)
|
|||
|
|
|
|||
|
|
return finaloss
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
if __name__ == '__main__':
|
|||
|
|
setup_seed()
|
|||
|
|
# input = torch.randn(2, 3, 4, 4, dtype=torch.float32)
|
|||
|
|
input = torch.randn(1, 3, 4, 4, dtype=torch.float32)
|
|||
|
|
# target = torch.randint(3, (1, 4, 4), dtype=torch.int64)
|
|||
|
|
# target = torch.randint(3, (2, 4, 4), dtype=torch.int64)
|
|||
|
|
target = torch.tensor(([[[2, 1, 1, 2],
|
|||
|
|
[2, 255, 2, 2],
|
|||
|
|
[2, 0, 255, 2],
|
|||
|
|
[1, 1, 0, 0]]]))
|
|||
|
|
|
|||
|
|
|
|||
|
|
# target = torch.tensor(([[[1, 1, 2, 1],
|
|||
|
|
# [2, 0, 1, 0],
|
|||
|
|
# [2, 0, 255, 2],
|
|||
|
|
# [2, 2, 0, 1]],
|
|||
|
|
# [[2, 2, 1, 0],
|
|||
|
|
# [1, 0, 0, 0],
|
|||
|
|
# [0, 255, 0, 1],
|
|||
|
|
# [1, 2, 1, 1]]]))
|
|||
|
|
|
|||
|
|
print("---------pred----------")
|
|||
|
|
print(input)
|
|||
|
|
print("---------target----------")
|
|||
|
|
print(target)
|
|||
|
|
print("__________softmax___________")
|
|||
|
|
soft = F.softmax(input, dim=1)
|
|||
|
|
print(soft)
|
|||
|
|
print("__________log___________")
|
|||
|
|
log = torch.log(soft)
|
|||
|
|
print(log)
|
|||
|
|
|
|||
|
|
# backloss =foreground_sample_hardbackgroundversion1(input,target)
|
|||
|
|
backloss =softmax_focalloss(input,target)
|
|||
|
|
|
|||
|
|
print(backloss)
|
|||
|
|
#
|
|||
|
|
|
|||
|
|
#
|
|||
|
|
#
|
|||
|
|
# if __name__ == '__main__':
|
|||
|
|
#
|
|||
|
|
# setup_seed()
|
|||
|
|
#
|
|||
|
|
# input = torch.randn(1, 3, 4, 4, dtype=torch.float32)
|
|||
|
|
# target = torch.randint(3, (1, 4, 4), dtype=torch.int64)
|
|||
|
|
#
|
|||
|
|
# print("---------pred----------")
|
|||
|
|
# print(input)
|
|||
|
|
#
|
|||
|
|
# print("---------target----------")
|
|||
|
|
# print(target)
|
|||
|
|
#
|
|||
|
|
# print("__________softmax___________")
|
|||
|
|
# soft = F.softmax(input, dim=1)
|
|||
|
|
# print(soft)
|
|||
|
|
# print("__________log___________")
|
|||
|
|
# log = torch.log(soft)
|
|||
|
|
# print(log)
|
|||
|
|
#
|
|||
|
|
# print("-----------------------------")
|
|||
|
|
# # softmax_focalloss=softmax_focalloss(input,target)
|
|||
|
|
#
|
|||
|
|
#
|
|||
|
|
# # tmx = 10000
|
|||
|
|
# # for i in range(tmx):
|
|||
|
|
# # loss = annealing_softmax_focalloss(input, target, t=i, t_max=tmx)
|
|||
|
|
# # # softmax_focalloss=softmax_focalloss(input,target)
|
|||
|
|
#
|
|||
|
|
#
|
|||
|
|
# # import matplotlib.pyplot as plt
|
|||
|
|
# #
|
|||
|
|
# # _t_max=300
|
|||
|
|
# # ylist= []
|
|||
|
|
# # xlist=[]
|
|||
|
|
# # for i in range(_t_max):
|
|||
|
|
# # y=cosine_annealing(0.1,0.5,_t=i,_t_max=_t_max)
|
|||
|
|
# # ylist.append(y)
|
|||
|
|
# # xlist.append(i)
|
|||
|
|
# #
|
|||
|
|
# # plt.figure()
|
|||
|
|
# # plt.plot(xlist,ylist)
|
|||
|
|
# # plt.show()
|
|||
|
|
# #
|
|||
|
|
#
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
|
|||
|
|
#
|
|||
|
|
# if __name__ == '__main__':
|
|||
|
|
#
|
|||
|
|
# target = torch.tensor(([[[2, 1, 1, 2],
|
|||
|
|
# [2, 255, 2, 2],
|
|||
|
|
# [2, 0, 255, 2],
|
|||
|
|
# [1, 1, 0, 0]]]))
|
|||
|
|
#
|
|||
|
|
# epoch =300
|
|||
|
|
# for i in range(epoch):
|
|||
|
|
# # input = torch.randn(2, 3, 4, 4, dtype=torch.float32)
|
|||
|
|
# input = torch.randn(1, 3, 4, 4, dtype=torch.float32)
|
|||
|
|
#
|
|||
|
|
#
|
|||
|
|
# print("---------pred----------")
|
|||
|
|
# print(input)
|
|||
|
|
# print("---------target----------")
|
|||
|
|
# print(target)
|
|||
|
|
# print("__________softmax___________")
|
|||
|
|
# soft = F.softmax(input, dim=1)
|
|||
|
|
# print(soft)
|
|||
|
|
# print("__________log___________")
|
|||
|
|
# log = torch.log(soft)
|
|||
|
|
# print(log)
|
|||
|
|
# backloss =sample_hardbackgroundversion2(input,target,t=i+1, t_max=50)
|
|||
|
|
#
|
|||
|
|
# print(backloss)
|
|||
|
|
# #
|
|||
|
|
#
|
|||
|
|
#
|