Error while train in GPU with tensorboard

#1

I have inserted network as well as input to the device, yet an error has emerged
This error vanishes when i comment writer.add_graph(model,example_data) in main.py
Exception has occurred: RuntimeError
Expected object of device type cuda but got device type cpu for argument #1 ‘self’ in call to _thnn_conv2d_forward
File “C:\Users\Vatsal\Documents\PROJECTS\Personal Projects\COVID-CT\model.py”, line 18, in forward
x = self.conv1(x)
File “C:\Users\Vatsal\Documents\PROJECTS\Personal Projects\COVID-CT\main.py”, line 123, in
writer.add_graph(model_basic_CNN,example_data)

image
image1275×412 23.6 KB

Full Code Available through the LINK
Following implementation for the model.py was used

import torch
from torch import nn
from torch.nn import functional as F
class basic_CNN(nn.Module):
    """
    Basic CNN architecture deployed
    """
    def __init__(self):
        super(basic_CNN, self).__init__()
        self.conv1 = nn.Conv2d(3,32,4,stride=2)
        self.conv2 = nn.Conv2d(32,64,3,stride=2)
        self.conv3 = nn.Conv2d(64,128,3,stride=2)
        self.conv4 = nn.Conv2d(128,256,3,stride=2)
        self.fc1 = nn.Linear(43264,64)
        self.fc2 = nn.Linear(64,1)

    def forward(self, x):
        x = self.conv1(x)
        x = self.conv2(x)
        x = F.relu(x)
        x = self.conv3(x)
        x = self.conv4(x)
        x = F.relu(x)
        x = x.view(-1,self.num_flat_features(x))
        x = F.relu(self.fc1(x))
        x = torch.sigmoid(self.fc2(x))
        return x

    def num_flat_features(self, x):
        size = x.size()[1:]
        num_features = 1
        for s in size:
            num_features *= s
        return num_features

Following implementation for the main.py was used

import torch
import torchvision
from DataLoader import *
from model import *
#from utils import *
from torchvision import transforms
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
import os
import sys
import warnings
import numpy as np
from sklearn.metrics import roc_auc_score, confusion_matrix

log_path = r'C:\Users\Vatsal\Documents\PROJECTS\logs'
writer = SummaryWriter(log_dir=log_path)
#HYPER-PARAMETERS
batchsize=16
bs = 10
lr = 1e-4

device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
print(device)


normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
train_transformer = transforms.Compose([
    transforms.Resize(256),
    transforms.RandomResizedCrop((224),scale=(0.5,1.0)),
    transforms.RandomHorizontalFlip(),
    transforms.ToTensor(),
    normalize
])

val_transformer = transforms.Compose([
    transforms.Resize(224),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    normalize
])

train_set = CovidCTDataset(root_dir=os.getcwd(), type='train', transform=train_transformer)
val_set = CovidCTDataset(root_dir=os.getcwd(), type='val', transform=val_transformer)
test_set = CovidCTDataset(root_dir=os.getcwd(), type='test', transform=val_transformer)
print("Size of Train Set {}".format(train_set.__len__()))
print("Size of Val Set {}".format(val_set.__len__()))
print("Size of Test Set {}".format(test_set.__len__()))

train_loader = DataLoader(train_set, batch_size=batchsize, drop_last=False, shuffle=True)
val_loader = DataLoader(val_set, batch_size=batchsize, drop_last=False, shuffle=True)
test_loader = DataLoader(test_set, batch_size=batchsize, drop_last=False, shuffle=False)


def train(epoch, optimizer, criterion):
    model.train()
    train_loss = 0
    train_correct = 0
    for batch_index, batch_samples in enumerate(train_loader):
        data, target = batch_samples['img'].to(device), batch_samples['label'].to(device)
        target = target.view(-1,1)
        optimizer.zero_grad()
        #print((data).is_cuda)
        output = model(data)
        loss = criterion(output.float(), target.float())
        train_loss = train_loss + loss.item()
        loss.backward()
        optimizer.step()

    return model, train_loss    


def val(epoch, criterion):
    model.eval()
    # Don't update model
    val_loss=0
    with torch.no_grad():
        tpr_list = []
        fpr_list = []
        
        predlist=[]
        targetlist=[]
        # Predict
        for batch_index, batch_samples in enumerate(val_loader):
            data, target = batch_samples['img'].to(device), batch_samples['label'].to(device)
            target = target.view(-1,1)
            output = model(data)
            loss = criterion(output.float(), target.float())
            val_loss += loss.item()
            output = [round(int(x)) for x in output]

            targetcpu=target.long().cpu().numpy()
            predlist=np.append(predlist, np.array(output))
            targetlist=np.append(targetlist,targetcpu)
            
    return targetlist, predlist, val_loss



# train
bs = 10
votenum = 10

#warnings.filterwarnings('ignore')

r_list = []
p_list = []
acc_list = []

model = basic_CNN().cuda()
model.to(device)
print(next(model.parameters()).is_cuda)
modelname = 'basic_CNN'
criterion = torch.nn.BCELoss()
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
examples = next(iter(train_loader))
example_data = examples['img']
example_label = examples['label']
img_grid = torchvision.utils.make_grid(example_data)
writer.add_image('Coivid Img',img_grid)
example_data.to(device)
writer.add_graph(model,example_data)

total_epoch = 3000
for epoch in range(1, total_epoch+1):
    train_loss = train(epoch=epoch, optimizer=optimizer, criterion=criterion)
    targetlist, predlist, val_loss = val(epoch=epoch, criterion=criterion)

    TN, FP, FN, TP = confusion_matrix(targetlist, predlist).ravel()
    p = TP / (TP + FP)
    r = TP / (TP + FN)
    F1 = 2 * r * p / (r + p)
    acc = (TP + TN) / (TP + TN + FP + FN)
    writer.add_scalar('train loss', np.asarray(train_loss[-1]), epoch)
    writer.add_scalar('val loss', np.array(val_loss) , epoch)
    writer.add_scalar('True Positive', TP, epoch)
    writer.add_scalar('False Positive', FP, epoch)
    writer.add_scalar('True Negative', TN, epoch)
    writer.add_scalar('False Negative', FN, epoch)
    writer.add_scalar('Accuracy', acc, epoch)
    writer.add_scalar('F1 score', F1, epoch)
    writer.add_scalar('Precision', p, epoch)
    writer.add_scalar('Recall', r, epoch)

    if epoch % votenum == 0:    
        print('TP=',TP,'TN=',TN,'FN=',FN,'FP=',FP)
        print('Total Correct',TP+FP)
        print('precision',p)
        print('recall',r)
        print('F1',F1)
        print('acc',acc)
        print('\n The epoch is {}, average recall: {:.4f}, average precision: {:.4f},average F1: {:.4f}, average accuracy: {:.4f}'.format(
        epoch, r, p, F1, acc))




writer.close()

Additionally,
Can anyone please verify whether for binary classification what should be dimensions of last linear layer should it be torch.nn.Linear(64,1) or I need to change to torch.nn.Linear(64,2)

Any help is deeply appreciated !!!

#2

The to() operation is not an in-place op for tensors, so you would need to reassign it:

example_data = example_data.to(device)

For nn.BCELoss or nn.BCEWithLogitsLoss the former shape of [batch_size, 1] is expected.

You could use the output shape of [batch_size, 2] in case you want to apply a multi-class classification to this binary use case, which could be applied via nn.CrossEntropyLoss.

#3

Could you please clarify again on the first part where shall I reassign !?

#4

I think he is referring to the bold lines below from your code. You need to assign them to some variables.

model = basic_CNN().cuda()
model.to(device)
print(next(model.parameters()).is_cuda)
modelname = ‘basic_CNN’
criterion = torch.nn.BCELoss()
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
examples = next(iter(train_loader))
example_data = examples[‘img’]
example_label = examples[‘label’]
img_grid = torchvision.utils.make_grid(example_data)
writer.add_image(‘Coivid Img’,img_grid)
example_data.to(device)
writer.add_graph(model,example_data)

Change them to

model = model.to(device)
example_data = example_data.to(device)
#5

Thanks @ptrblck and @Shreyansh_IITB for you help