pytorch loss 出现nan,原因之一,计算图中存在torch.sqrt或者 **0.5,以及如何解决

本文探讨了在PyTorch中使用sqrt函数可能导致的梯度为inf问题,尤其是在loss函数计算中。通过两个解决方案避免了该问题:一是避免直接开方运算;二是向平方和中添加极小值防止分母为零。提供了测试代码示例。

摘要生成于 C知道 ,由 DeepSeek-R1 满血版支持, 前往体验 >

今天写一个loss函数

dist=torch.sqrt(x*x+y*y)
loss=soomthL1loss(dist,gt_dist)

我随便写的几句示意代码,这样会导致在第一个iteration之后出现nan,第一次iteration之内,还是可以看到loss不为nan的。

解决办法:

     1、不开方,因为开方的求导会出现在分母上,因此需要避免分母为0!

     2、torch.sqrt(x*x+0.000001)增加一个 很小的 “一瞥西漏”

给大家个测试代码:

import torch
a = torch.zeros(1,requireds_grad = True)
b = torch.sqrt(a)
b.backward()
print(a.grad)
#得到tensor([inf]),看到inf就知道,一般来说没办法传递了,为什么是一般来说,因为用过darknet的yolo的话,里面出现inf还是可以训练的(可能我记错了)


#修改下:
import torch
a = torch.zeros(1,requireds_grad = True)
b = torch.sqrt(a+0.001)
b.backward()
print(a.grad)
#tensor([15.814])

 

pytorch部分代码如下:train_loss, train_acc = train(model_ft, DEVICE, train_loader, optimizer, epoch,model_ema) for batch_idx, (data, target) in enumerate(train_loader): data, target = data.to(device, non_blocking=True), Variable(target).to(device,non_blocking=True) # 3、将数据输入mixup_fn生成mixup数据 samples, targets = mixup_fn(data, target) # 4、将上一步生成的数据输入model,输出预测结果,再计算loss output = model(samples) # 5、梯度清零(将loss关于weight的导数变成0) optimizer.zero_grad() # 6、若使用混合精度 if use_amp: with torch.cuda.amp.autocast(): # 开启混合精度 loss = torch.nan_to_num(criterion_train(output, targets)) # 计算loss scaler.scale(loss).backward() # 梯度放大 torch.nn.utils.clip_grad_norm_(model.parameters(), CLIP_GRAD) if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks or global_forward_hooks or global_forward_pre_hooks): return forward_call(*input, **kwargs) class LDAMLoss(nn.Module): def init(self, cls_num_list, max_m=0.5, weight=None, s=30): super(LDAMLoss, self).init() m_list = 1.0 / np.sqrt(np.sqrt(cls_num_list)) m_list = m_list * (max_m / np.max(m_list)) m_list = torch.cuda.FloatTensor(m_list) self.m_list = m_list assert s > 0 self.s = s self.weight = weight def forward(self, x, target): index = torch.zeros_like(x, dtype=torch.uint8) target = torch.clamp(target, 0, index.size(1) - 1) index.scatter(1, target.data.view(-1, 1).type(torch.int64), 1) index_float = index.type(torch.cuda.FloatTensor) batch_m = torch.matmul(self.m_list[None, :], index_float.transpose(0,1)) batch_m = batch_m.view((-1, 1)) x_m = x - batch_m output = torch.where(index, x_m, x) return F.cross_entropy(self.s*output, target, weight=self.weight) 报错:RuntimeError: Expected index [112, 1] to be smaller than self [16, 7] apart from dimension 1 帮我看看如何修改源代码
06-10
pytorch部分代码如下:train_loss, train_acc = train(model_ft, DEVICE, train_loader, optimizer, epoch,model_ema) for batch_idx, (data, target) in enumerate(train_loader): data, target = data.to(device, non_blocking=True), Variable(target).to(device,non_blocking=True) samples, targets = mixup_fn(data, target) output = model(samples) optimizer.zero_grad() if use_amp: with torch.cuda.amp.autocast(): loss = torch.nan_to_num(criterion_train(output, targets)) scaler.scale(loss).backward() torch.nn.utils.clip_grad_norm_(model.parameters(), CLIP_GRAD) if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks or global_forward_hooks or global_forward_pre_hooks): return forward_call(*input, **kwargs) class LDAMLoss(nn.Module): def init(self, cls_num_list, max_m=0.5, weight=None, s=30): super(LDAMLoss, self).init() m_list = 1.0 / np.sqrt(np.sqrt(cls_num_list)) m_list = m_list * (max_m / np.max(m_list)) m_list = torch.cuda.FloatTensor(m_list) self.m_list = m_list assert s > 0 self.s = s self.weight = weight def forward(self, x, target): index = torch.zeros_like(x, dtype=torch.uint8) target = torch.clamp(target, 0, index.size(1) - 1) index.scatter_(1, target.unsqueeze(1).type(torch.int64), 1) index = index[:, :x.size(1)] index_float = index.type(torch.cuda.FloatTensor) batch_m = torch.matmul(self.m_list[None, :], index_float.transpose(0,1)) batch_m = batch_m.view((-1, 1)) x_m = x - batch_m output = torch.where(index, x_m, x) return F.cross_entropy(self.s*output, target, weight=self.weight) 报错: File "/home/adminis/hpy/ConvNextV2_Demo/train+ca.py", line 46, in train loss = torch.nan_to_num(criterion_train(output, targets)) # 计算loss File "/home/adminis/anaconda3/envs/wln/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl return forward_call(*input, **kwargs) File "/home/adminis/hpy/ConvNextV2_Demo/models/utils.py", line 622, in forward index.scatter_(1, target.unsqueeze(1).type(torch.int64), 1) # target.data.view(-1, 1). RuntimeError: Index tensor must have the same number of dimensions as self tensor 帮我看看如何修改源代码
06-10
pytorch部分代码如下:class LDAMLoss(nn.Module): def init(self, cls_num_list, max_m=0.5, weight=None, s=30): super(LDAMLoss, self).init() m_list = 1.0 / np.sqrt(np.sqrt(cls_num_list)) m_list = m_list * (max_m / np.max(m_list)) m_list = torch.cuda.FloatTensor(m_list) self.m_list = m_list assert s > 0 self.s = s # self.weight = weight if weight is not None: weight = torch.FloatTensor(weight).cuda() self.weight = weight self.cls_num_list = cls_num_list def forward(self, x, target): index = torch.zeros_like(x, dtype=torch.uint8) index_float = index.type(torch.cuda.FloatTensor) batch_m = torch.matmul(self.m_list[None, :], index_float.transpose(0,1)) # 0,1 batch_m = batch_m.view((x.size(0), 1)) # size=(batch_size, 1) (-1,1) x_m = x - batch_m output = torch.where(index, x_m, x) # return F.cross_entropy(self.s*output, target, weight=self.weight) if self.weight is not None: output = output * self.weight[None, :] target = torch.flatten(target) # 将 target 转换成 1D Tensor logit = output * self.s return F.cross_entropy(logit, target, weight=self.weight) for batch_idx, (data, target) in enumerate(train_loader): data, target = data.to(device, non_blocking=True), Variable(target).to(device,non_blocking=True) # 3、将数据输入mixup_fn生成mixup数据 samples, targets = mixup_fn(data, target) # 4、将上一步生成的数据输入model,输出预测结果,再计算loss output = model(samples) # 5、梯度清零(将loss关于weight的导数变成0) optimizer.zero_grad() loss = criterion_train(output, targets) # 6、若使用混合精度 if use_amp: with torch.cuda.amp.autocast(): # 开启混合精度 # loss = torch.nan_to_num(criterion_train(output, target_a, target_b, lam)) # 计算loss # loss = lam * criterion_train(output, target_a) + (1 - lam) * criterion_train(output, target_b) # 计算 mixup 后的损失函数 scaler.scale(loss).backward() # 梯度放大 torch.nn.utils.clip_grad_norm_(model.parameters(), CLIP_GRAD) # 梯度裁剪,防止梯度爆炸 scaler.step(optimizer) # 更新下一次迭代的scaler scaler.update() # 否则,直接反向传播求梯度 else: # loss = criterion_train(output, targets) loss.backward() torch.nn.utils.clip_grad_norm_(model.parameters(), CLIP_GRAD) optimizer.step() 报错:) File "/home/adminis/hpy/ConvNextV2_Demo/models/losses.py", line 48, in forward output = torch.where(index, x_m, x) RuntimeError: expected scalar type float but found c10::Half
05-30
评论 3
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值