Cuda out of memory - problem in code or gpu?

Hello all!. I am currently working on a computer vision project. I keep getting a runtime error that says “CUDA out of memory”. I have tried all possible ways like reducing batch size and image resolution, clearing the cache, deleting variables after training starts, reducing image data and so on… Unfortunately, this error doesn’t stop. I have a Nvidia Geforce 940MX graphics card on my HP Pavilion laptop. I have installed cuda 10.2 and cudNN from the pytorch installation page. My aim was to create a flask website out of this model but I am stuck with this issue. Any suggestions to this problem will be helpful.

This is my code

import pandas as pd
import numpy as np
import torch
import torch.nn as nn
import os
import cv2
import matplotlib.pyplot as plt
from sklearn.model_selection import train_test_split
import albumentations as A
from import TensorDataset, DataLoader,Dataset
from torchvision import models
from collections import defaultdict
from import RandomSampler
import torch.optim as optim
from torch.optim import lr_scheduler
from sklearn import model_selection
from tqdm import tqdm
import gc

# generate data from csv file
class Build_dataset(Dataset):

    def __init__(self, csv, split, mode, transform=None):
        self.csv = csv.reset_index(drop=True)
        self.split = split
        self.mode = mode
        self.transform = transform

    def __len__(self):
        return self.csv.shape[0]

    def __getitem__(self, index):
        row = self.csv.iloc[index]

        image = cv2.imread(row.filepath)
        image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)

        if self.transform is not None:
            res = self.transform(image=image)
            image = res['image'].astype(np.float32)
            image = image.astype(np.float32)

        image = image.transpose(2, 0, 1)
        data = torch.tensor(image).float()

        if self.mode == 'test':
            return data
            return data, torch.tensor(self.csv.iloc[index].target).long()

# training epoch          
def train_epoch(model, loader, optimizer,loss_fn,device, scheduler,n_examples):

model = model.train()

losses = []
correct_predictions = 0

for inputs, labels in tqdm(loader):
    inputs =
    labels =

    outputs = model(inputs)

    _, preds = torch.max(outputs, dim=1)
    loss = loss_fn(outputs, labels)
    correct_predictions += torch.sum(preds == labels)
# here you delete inputs and labels and then use gc.collect
    del inputs, labels

return correct_predictions.double() / n_examples, np.mean(losses)

# validation epoch 
def val_epoch(model, loader,loss_fn, device,n_examples):

model = model.eval()

losses = []
correct_predictions = 0

with torch.no_grad():
    for inputs, labels in tqdm(loader):
        inputs =
        labels =
        outputs = model(inputs)
        _, preds = torch.max(outputs, dim=1)
        loss = loss_fn(outputs, labels)
        correct_predictions += torch.sum(preds == labels)
        # here you delete inputs and labels and then use gc.collect
        del inputs, labels

return correct_predictions.double() / n_examples, np.mean(losses)

def train(model,device, num_epochs):
# generate data
dataset_train = Build_dataset(df_train,  'train', 'train', transform=transforms_train)
dataset_valid = Build_dataset(df_valid, 'train', 'val', transform=transforms_val)

#load data 
train_loader = DataLoader(dataset_train, batch_size = 16,sampler=RandomSampler(dataset_train), num_workers=4)
valid_loader = DataLoader(dataset_valid, batch_size = 16,shuffle = True, num_workers= 4 )

dataset_train_size = len(dataset_train)

dataset_valid_size = len(dataset_valid)

optimizer = optim.Adam(model.parameters(), lr = 3e-5)

model =

scheduler = lr_scheduler.ReduceLROnPlateau(optimizer, patience = 3,threshold = 0.001, mode = 'max')

loss_fn = nn.CrossEntropyLoss().to(device)

history = defaultdict(list)

best_accuracy = 0.0

for epoch in range(num_epochs):
    print(f'Epoch {epoch+1} / {num_epochs}')
    print ('-'*30)
    train_acc, train_loss = train_epoch(model, train_loader, optimizer, loss_fn, device, scheduler, dataset_train_size)
    print(f'Train loss {train_loss} accuracy {train_acc}')
    valid_acc, valid_loss = val_epoch(model, valid_loader, loss_fn, device,dataset_valid_size)
    print(f'Val   loss {valid_loss} accuracy {valid_acc}')
    if valid_acc > best_accuracy:, 'best_model_state.bin')
        best_accuracy = valid_acc
print('Best Accuracy: {best_accuracy}')


return model, history

if __name__ == '__main__':
#competition data -2020
data_dir = "C:\\Users\\Aniruddh\\Documents\\kaggle\\jpeg_melanoma_2020"
#competition data - 2019
data_dir2 = "C:\\Users\\Aniruddh\\Documents\\kaggle\\jpeg_melanoma_2019"
# device
device = torch.device("cuda")

# augmenting images

image_size = 384
transforms_train = A.Compose([
    A.RandomBrightness(limit=0.2, p=0.75),
    A.RandomContrast(limit=0.2, p=0.75),
        A.GaussNoise(var_limit=(5.0, 30.0)),
    ], p=0.7),

        A.GridDistortion(num_steps=5, distort_limit=1.),
    ], p=0.7),

    A.CLAHE(clip_limit=4.0, p=0.7),
    A.HueSaturationValue(hue_shift_limit=10, sat_shift_limit=20, val_shift_limit=10, p=0.5),
    A.ShiftScaleRotate(shift_limit=0.1, scale_limit=0.1, rotate_limit=15, border_mode=0, p=0.85),
    A.Resize(image_size, image_size),
    A.Cutout(max_h_size=int(image_size * 0.375), max_w_size=int(image_size * 0.375), num_holes=1, p=0.7),    

transforms_val = A.Compose([
    A.Resize(image_size, image_size),
# create data
df_train = pd.read_csv("C:\\Users\\Aniruddh\\Documents\\kaggle\\jpeg_melanoma_2020\\train.csv")  #/kaggle/input/siim-isic-melanoma-classification/train.csv

df_train['is_ext'] = 0
df_train['filepath'] = df_train['image_name'].apply(lambda x: os.path.join(data_dir, 'train', f'{x}.jpg'))

# dataset from 2020 data
df_train['diagnosis'] = df_train['diagnosis'].apply(lambda x: x.replace('seborrheic keratosis', 'BKL'))
df_train['diagnosis'] = df_train['diagnosis'].apply(lambda x: x.replace('lichenoid keratosis', 'BKL'))
df_train['diagnosis'] = df_train['diagnosis'].apply(lambda x: x.replace('solar lentigo', 'BKL'))
df_train['diagnosis'] = df_train['diagnosis'].apply(lambda x: x.replace('lentigo NOS', 'BKL'))
df_train['diagnosis'] = df_train['diagnosis'].apply(lambda x: x.replace('cafe-au-lait macule', 'unknown'))
df_train['diagnosis'] = df_train['diagnosis'].apply(lambda x: x.replace('atypical melanocytic proliferation', 'unknown'))

# dataset from 2019 data
df_train2 = pd.read_csv('/content/drive/My Drive/siim_melanoma images/train_2019.csv')
df_train2 = df_train2[df_train2['tfrecord'] >= 0].reset_index(drop=True)
#df_train2['fold'] = df_train2['tfrecord'] % 5
df_train2['is_ext'] = 1
df_train2['filepath'] = df_train2['image_name'].apply(lambda x: os.path.join(data_dir2, 'train', f'{x}.jpg'))

df_train2['diagnosis'] = df_train2['diagnosis'].apply(lambda x: x.replace('NV', 'nevus'))
df_train2['diagnosis'] = df_train2['diagnosis'].apply(lambda x: x.replace('MEL', 'melanoma'))

#concat both 2019 and 2020 data
df_train = pd.concat([df_train, df_train2]).reset_index(drop=True)

# shuffle data
df = df_train.sample(frac=1).reset_index(drop=True)

# creating 8 different target values
new_target = {d: idx for idx, d in enumerate(sorted(df.diagnosis.unique()))}
df['target'] = df['diagnosis'].map(new_target)
mel_idx = new_target['melanoma']

df = df[['filepath','diagnosis', 'target', 'is_ext']]

class_names = list(df['diagnosis'].unique())

# splitting train and validation data by 20%
df_valid = df[:11471]
df_train = df[11472:].reset_index()
df_train = df_train.drop(columns = ['index'])

# create model

def create_model(n_classes):
    model = models.resnet50(pretrained=True)

    n_features = model.fc.in_features
    model.fc = nn.Linear(n_features, n_classes)

# model    
base_model = create_model(len(class_names)) 

# train model      
base_model, history = train(base_model, device, num_epochs = 15) 

Code Objective

The purpose of the project is to classify skin cancer images by creating 8 different target variables from the given datasets (i.e the competition was about classifying benign and malignant images but I used the diagnosis column on the dataset as my target variable as the data was really skewed). The model used is Resnet-50 from torchvision models.

These were the data used
skin images (this year competition):
skin images (last year competition):

I decided to create a Flask application out of this but, the CUDA memory was always causing a runtime error

RuntimeError: CUDA out of memory. Tried to allocate 144.00 MiB (GPU 0; 2.00 GiB total capacity; 1.21 GiB already allocated; 43.55 MiB free; 1.23 GiB reserved in total by PyTorch)

These are the details about my Nvidia GPU

Sun Sep 13 19:09:34 2020
| NVIDIA-SMI 451.67       Driver Version: 451.67       CUDA Version: 11.0     |
| GPU  Name            TCC/WDDM | Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|   0  GeForce 940MX      WDDM  | 00000000:01:00.0 Off |                  N/A |
| N/A   63C    P8    N/A /  N/A |     37MiB /  2048MiB |      0%      Default |

| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|  No running processes found                                                 |

# more information about my GPU
==============NVSMI LOG==============

Timestamp                                 : Sun Sep 13 19:11:22 2020
Driver Version                            : 451.67
CUDA Version                              : 11.0

Attached GPUs                             : 1
GPU 00000000:01:00.0
Product Name                          : GeForce 940MX
Product Brand                         : GeForce
Display Mode                          : Disabled
Display Active                        : Disabled
Persistence Mode                      : N/A
Accounting Mode                       : Disabled
Accounting Mode Buffer Size           : 4000
Driver Model
    Current                           : WDDM
    Pending                           : WDDM
Serial Number                         : N/A
GPU UUID                              : GPU-9a8c69df-26f2-2a98-3712-ea22f6add038
Minor Number                          : N/A
VBIOS Version                         : 82.08.6D.00.8C
MultiGPU Board                        : No
Board ID                              : 0x100
GPU Part Number                       : N/A
Inforom Version
    Image Version                     : N/A
    OEM Object                        : N/A
    ECC Object                        : N/A
    Power Management Object           : N/A
GPU Operation Mode
    Current                           : N/A
    Pending                           : N/A
GPU Virtualization Mode
    Virtualization Mode               : None
    Host VGPU Mode                    : N/A
    Relaxed Ordering Mode             : N/A
    Bus                               : 0x01
    Device                            : 0x00
    Domain                            : 0x0000
    Device Id                         : 0x134D10DE
    Bus Id                            : 00000000:01:00.0
    Sub System Id                     : 0x83F9103C
    GPU Link Info
        PCIe Generation
            Max                       : 3
            Current                   : 1
        Link Width
            Max                       : 4x
            Current                   : 4x
    Bridge Chip
        Type                          : N/A
        Firmware                      : N/A
    Replays Since Reset               : 0
    Replay Number Rollovers           : 0
    Tx Throughput                     : 0 KB/s
    Rx Throughput                     : 0 KB/s
Fan Speed                             : N/A
Performance State                     : P8
Clocks Throttle Reasons
    Idle                              : Not Active
    Applications Clocks Setting       : Not Active
    SW Power Cap                      : Not Active
    HW Slowdown                       : Not Active
        HW Thermal Slowdown           : N/A
        HW Power Brake Slowdown       : N/A
    Sync Boost                        : Not Active
    SW Thermal Slowdown               : Not Active
    Display Clock Setting             : Not Active
FB Memory Usage
    Total                             : 2048 MiB
    Used                              : 37 MiB
    Free                              : 2011 MiB
BAR1 Memory Usage
    Total                             : 256 MiB
    Used                              : 225 MiB
    Free                              : 31 MiB
Compute Mode                          : Default
    Gpu                               : 0 %
    Memory                            : 0 %
    Encoder                           : N/A
    Decoder                           : N/A
Encoder Stats
    Active Sessions                   : 0
    Average FPS                       : 0
    Average Latency                   : 0
FBC Stats
    Active Sessions                   : 0
    Average FPS                       : 0
    Average Latency                   : 0
Ecc Mode
    Current                           : N/A
    Pending                           : N/A
ECC Errors
        Single Bit
            Device Memory             : N/A
            Register File             : N/A
            L1 Cache                  : N/A
            L2 Cache                  : N/A
            Texture Memory            : N/A
            Texture Shared            : N/A
            CBU                       : N/A
            Total                     : N/A
        Double Bit
            Device Memory             : N/A
            Register File             : N/A
            L1 Cache                  : N/A
            L2 Cache                  : N/A
            Texture Memory            : N/A
            Texture Shared            : N/A
            CBU                       : N/A
            Total                     : N/A
        Single Bit
            Device Memory             : N/A
            Register File             : N/A
            L1 Cache                  : N/A
            L2 Cache                  : N/A
            Texture Memory            : N/A
            Texture Shared            : N/A
            CBU                       : N/A
            Total                     : N/A
        Double Bit
            Device Memory             : N/A
            Register File             : N/A
            L1 Cache                  : N/A
            L2 Cache                  : N/A
            Texture Memory            : N/A
            Texture Shared            : N/A
            CBU                       : N/A
            Total                     : N/A
Retired Pages
    Single Bit ECC                    : N/A
    Double Bit ECC                    : N/A
    Pending Page Blacklist            : N/A
Remapped Rows                         : N/A
    GPU Current Temp                  : 60 C
    GPU Shutdown Temp                 : 99 C
    GPU Slowdown Temp                 : 94 C
    GPU Max Operating Temp            : 90 C
    Memory Current Temp               : N/A
    Memory Max Operating Temp         : N/A
Power Readings
    Power Management                  : N/A
    Power Draw                        : N/A
    Power Limit                       : N/A
    Default Power Limit               : N/A
    Enforced Power Limit              : N/A
    Min Power Limit                   : N/A
    Max Power Limit                   : N/A
    Graphics                          : 405 MHz
    SM                                : 405 MHz
    Memory                            : 405 MHz
    Video                             : 396 MHz
Applications Clocks
    Graphics                          : 1006 MHz
    Memory                            : 1001 MHz
Default Applications Clocks
    Graphics                          : 1004 MHz
    Memory                            : 1001 MHz
Max Clocks
    Graphics                          : 1241 MHz
    SM                                : 1241 MHz
    Memory                            : 1001 MHz
    Video                             : 1216 MHz
Max Customer Boost Clocks
    Graphics                          : N/A
Clock Policy
    Auto Boost                        : N/A
    Auto Boost Default                : N/A
Processes                             : None

Could it be possible that u loaded other things in the CUDA device too other than the training data features, labels and the model
Deleting variables after training start won’t help coz most variables are stored and handled on the RAM and cpu except the ones specified on the CUDA enabled gpu which should be just training data and model

Pls can u post ur code if possible

1 Like

@Aniruddh honestly I see nothing wrong in ur code have u tried running this code on the cpu instead or try running on a another cuda gpu pc
This might be from the configuration of ur gpu on the nnvidia panel (this is just a speculation)

yes! if I try running this on the CPU, the whole system freezes to the point where I have to manually restart the computer. Also if I try running the code with lower image resolution, lower batch sizes etc, each epoch takes around 12 hours to complete on a CPU which is definitely impractical.
Now…about using other GPU environments, I used the google colab to train the model, but it shows an error after a few hours of training saying ‘page not responsive’. I thought about colab pro (which is a paid version of colab) but it is available only in the United States…and I am from a different country. I am also thinking about aws EC2 p3 instances for renting a cloud GPU…but again, I don’t really have sufficient documentation about it’s usage…or how it is going to help me to create a deep learning website.

I do have a question though. Suppose I somehow successfully train the model on colab and save the model as the bin file (just like how I did in the code above), is it possible to download that trained model file to my local directory for prediction without the gpu? My question maybe complicated…apologies. I am just trying different ways.

Yes you can download the code file to your local directory.
Have u checked if their’s any configuration that looks off on ur nnvidia panel and also have u tried updating ur gpu drivers

Also this is just a speculation, but I don’t think u should use the ‘+=’ operator in the training loop to update coz this can accumulate gradients
Can u find a better way for this.

Another thing I want to ask is this: does u neural network have only linear layers and no convolutional layers? If ur neural network has convolutional layers can u pls post your neural network code here pls let me see how it downsamples images
The thing is if u r using just linear layers it can cause memory issues for ur gpu and if u are convolutional layers that do not downsample these images it will cause similar issues to when u use only linear layers.

I realized that u are using restnet to train so no need to post ur code
In that case ur restnet model is too big for ur gpu to handle as it is 50 layers deep (as the name implies ‘restnet50’)

Measuring the memory use of ResNet -50 training with a mini-batch of 32 on a typical high performance GPU shows that it needs over 7.5 GB of local DRAM. … This significantly slows down the training time and considerably increases power consumption

Oh! I didn’t see your recent post. Yes I kinda had a little feeling about the model size too.

In terms of my NVIDIA panel, I have submitted on the post about the systems management interface(smi) and all the information. To my understanding, the graphics card that I have is a mobile GPU with a low memory buffer that cannot accept over-sized memory space from these training models. Understandable. Updating drivers didn’t really work as well.

However here is a small update on the model but still a query.
I tried by (a) reducing the resnet 50 to resnet 18 and…
(b) reducing the training and validation data
and the model actually starting training successfully in my local system on cuda.
Here are the changes that I did to the data.

after specifying the training and validation in this line on the code.

df_valid = df[:11471]
df_train = df[11472:].reset_index()
df_train = df_train.drop(columns = ['index'])

I further reduced it to 10,000 train image data and 7000 valid image data

df_train = df_train.head(10000)
df_valid = df_valid.head(-7000)

Now that the model is training in this scenario, here is the question. Is this image data considered ‘small’ for training? or should I add more?

In that case the restnet50 was too big. In the case of neural networks, the more training data the better, although it’s not really necessary if ur model is really doing well and able to generalize rather than overfit on the training data, but then once in a while it’s good to train it on new datasets.
Though if u want to actually use the restnet50 then u can use a gpu cloud services for that given that it’s too much for ur gpu to handle

1 Like

Thanks a lot mate! :+1:t2: :+1:t2: