Hi everyone,
I am new to pytorch and I am writing a simple cat-dog classifier - my dataloader and trainer scripts are given below :
Dataloader :
import torchvision.transforms as transforms
import matplotlib.pyplot as plt
import numpy as np
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import random
import torch
import skimage
from torch.utils.data.dataset import Dataset
class CatDogDataset(Dataset):
def __init__(self, csv_file, root_dir):
self.df = pd.read_csv(csv_file)
self.root_dir = root_dir
# self.transform = transforms.Compose([transforms.ToTensor()])
def __len__(self):
return len(self.df)
# can add random rotation later on
def __getitem__(self,idx):
img_path = self.df.iloc[idx,0]
image = skimage.io.imread(img_path)
image = skimage.transform.resize(image,(3,128,128)).astype(np.float32)
image = (image-image.mean())/image.std()
image = image.astype(np.float32)
gt = np.asarray(self.df.iloc[idx,1]).astype(np.float32)
gt = torch.tensor(gt)
sample = {'image_data' : image,'gt' : gt}
return sample
Trainer :
import numpy as np
import pandas as pd
import torch
import matplotlib.pyplot as plt
import torchvision
import csv
from torch.autograd import Variable
import glob
import torch.optim as optim
import torch.nn as nn
import torch.nn.functional as F
from random import shuffle
import glob
import os
from data_loader import CatDogDataset
from torch.utils.data import DataLoader
from tqdm import tqdm
from torchsummary import summary
import sys
class Net(nn.Module):
def __init__(self):
super(Net,self).__init__()
self.conv1 = nn.Conv2d(3,base_filters,3, padding = 1)
self.pool1 = nn.MaxPool2d(2)
self.conv2 = nn.Conv2d(base_filters,base_filters*2,3, padding = 1)
self.pool2 = nn.MaxPool2d(2)
self.conv3 = nn.Conv2d(base_filters*2,base_filters*4,3, padding = 1)
self.pool3 = nn.MaxPool2d(2)
self.conv4 = nn.Conv2d(base_filters*4,base_filters*8,3, padding = 1)
self.pool4 = nn.MaxPool2d(2)
self.fc1 = nn.Linear(base_filters*8*8*8,120)
self.fc2 = nn.Linear(120,100)
self.fc3 = nn.Linear(100,1)
def forward(self,x):
x = self.pool1(F.relu(self.conv1(x)))
x = self.pool2(F.relu(self.conv2(x)))
x = self.pool3(F.relu(self.conv3(x)))
x = self.pool4(F.relu(self.conv4(x)))
x = x.view(-1,base_filters*8*8*8)
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = self.fc3(x)
#x = F.softmax(x)
return x
base_filters = 50
batch_size = 1
net = Net()
loss_fn = torch.nn.MSELoss()
dataset_train = CatDogDataset("/home/megh/work/catdog/CD_data.csv","/home/megh/work/catdog/")
train_loader = DataLoader(dataset_train, batch_size = batch_size, shuffle= True, num_workers=1)
#optimizer = optim.Adam(net.parameters(), lr = 0.1)
net.cuda()
summary(net, input_size=(3, 128, 128))
net = net.train()
tmp = []
lr = 1
for batch_idx, (animal) in enumerate(train_loader):
image = animal['image_data']
image = image.cuda()
gt = animal['gt']
gt = gt.cuda()
image, gt = Variable(image, requires_grad = True), Variable(gt, requires_grad = True)
#optimizer.zero_grad()
output = net(image)
loss = loss_fn(output,gt)
loss.backward()
#optimizer.step()
for p in net.parameters():
print(type(p.grad))
p.requires_grad_()
print(type(p.grad))
p = p - lr*(p.grad)
p.requires_grad_()
print(type(p.grad))
#print("helo")
p.grad.zero_()
print(loss)
tmp.append(net.parameters())
The error what I get is :
<class 'torch.Tensor'>
<class 'torch.Tensor'>
<class 'NoneType'>
^^ prints this before the error
AttributeError: 'NoneType' object has no attribute 'zero_'
I am not understanding why the datatype of gradient is becoming NoneType.
I have verified that the weights are not getting updated by storing the network parameters after each iteration (by commenting the p.grad.zero_()) and verifying that they are the same after every iteration.
However, when I iterate through the network parameters by writing a small script in my spyder kernel, there seems to be no issue, and the dtype doesnt become none after I do some operations on the parameters.
Any help will be greatly appreciated
Thanks
