将 Captum 与 Pytorch Lightning 一起使用?
Posted
技术标签:
【中文标题】将 Captum 与 Pytorch Lightning 一起使用?【英文标题】:Using Captum with Pytorch Lightning? 【发布时间】:2021-01-23 08:07:35 【问题描述】:所以我尝试将 Captum 与 PyTorch Lightning 一起使用。将模块传递给 Captum 时遇到问题,因为它似乎对张量进行了奇怪的重塑。 例如,在下面的最小示例中,闪电代码运行起来既简单又好。 但是当我将 IntegratedGradient 与“n_step>=1”一起使用时,我遇到了一个问题。 LighningModule 的代码我会说并不重要,我想知道最底部的代码行。
有人知道如何解决这个问题吗?
from captum.attr import IntegratedGradients
from torch import nn, optim, rand, sum as tsum, reshape, device
import torch.nn.functional as F
from pytorch_lightning import seed_everything, LightningModule, Trainer
from torch.utils.data import DataLoader, Dataset
SAMPLE_DIM = 3
class CustomDataset(Dataset):
def __init__(self, samples=42):
self.dataset = rand(samples, SAMPLE_DIM).cuda().float() * 2 - 1
def __getitem__(self, index):
return (self.dataset[index], (tsum(self.dataset[index]) > 0).cuda().float())
def __len__(self):
return self.dataset.size()[0]
class OurModel(LightningModule):
def __init__(self):
super(OurModel, self).__init__()
# Network layers
self.linear = nn.Linear(SAMPLE_DIM, 2048)
self.linear2 = nn.Linear(2048, 1)
self.output = nn.Sigmoid()
# Hyper-parameters, that we will auto-tune using lightning!
self.lr = 0.001
self.batch_size = 512
def forward(self, x):
x = self.linear(x)
x = self.linear2(x)
output = self.output(x)
return reshape(output, (-1,))
def configure_optimizers(self):
return optim.Adam(self.parameters(), lr=self.lr)
def train_dataloader(self):
loader = DataLoader(CustomDataset(samples=1000), batch_size=self.batch_size, shuffle=True)
return loader
def training_step(self, batch, batch_nb):
x, y = batch
loss = F.binary_cross_entropy(self(x), y)
return 'loss': loss, 'log': 'train_loss': loss
if __name__ == '__main__':
seed_everything(42)
device = device("cuda")
model = OurModel().to(device)
trainer = Trainer(max_epochs=2, min_epochs=1, auto_lr_find=False,
progress_bar_refresh_rate=10)
trainer.fit(model)
# ok Now the Problem
test_input = CustomDataset(samples=1).__getitem__(0)[0].requires_grad_()
ig = IntegratedGradients(model)
attr, delta = ig.attribute(test_input, target=1, return_convergence_delta=True)
【问题讨论】:
【参考方案1】:解决方案是包装 forward 函数。确保进入 mode.foward() 的形状是正确的!
# Solution is this wrapper function
def modified_f(in_vec):
# Shape here is wrong
print("IN:", in_vec.size())
x = torch.reshape(in_vec, (int(in_vec.size()[0]/SAMPLE_DIM), SAMPLE_DIM))
print("x:", x.size())
res = model.forward(x)
print("res:", res.size())
res = torch.reshape(res, (res.size()[0], 1))
print("res2:", res.size())
return res
ig = IntegratedGradients(modified_f)
attr, delta = ig.attribute(test_input, return_convergence_delta=True, n_steps=STEP_AMOUNT)
【讨论】:
以上是关于将 Captum 与 Pytorch Lightning 一起使用?的主要内容,如果未能解决你的问题,请参考以下文章
学习打卡05可解释机器学习笔记之CAM+Captum代码实战
学习打卡05可解释机器学习笔记之CAM+Captum代码实战