Iteration on images with Pytorch: error due to CUDA memory issue with batch size 1

Question

During training, the architecture generates three models and now encoder is used to encode images with iterations=16. After performing 6 iteration, i got an error. "CUDA out of memory". I have 4 gpus and i also implemented dataparallel class but it does not work

Transformation while training

train_transform = transforms.Compose([
#transforms.RandomCrop((32, 32)),
transforms.ToPILImage(),
transforms.Resize((512, 512)),
#transforms.ColorJitter(brightness=0.5),
transforms.ToTensor(),
])

Encoder.py file

import argparse

import numpy as np
from scipy.misc import imread, imresize, imsave

import torch
from torch.autograd import Variable

parser = argparse.ArgumentParser()
parser.add_argument(
    '--model', '-m', required=True, type=str, help='path to model')
parser.add_argument(
    '--input', '-i', required=True, type=str, help='input image')
parser.add_argument(
    '--output', '-o', required=True, type=str, help='output codes')
parser.add_argument('--cuda', '-g', action='store_true', help='enables cuda')
parser.add_argument(
    '--iterations', type=int, default=16, help='unroll iterations')
args = parser.parse_args()

image = imread(args.input, mode='RGB')
image = torch.from_numpy(
    np.expand_dims(
        np.transpose(image.astype(np.float32) / 255.0, (2, 0, 1)), 0))
batch_size, input_channels, height, width = image.size()
assert height % 32 == 0 and width % 32 == 0

image = Variable(image, volatile=True)

torch.cuda.empty_cache()


import network

encoder = network.EncoderCell()
binarizer = network.Binarizer()
decoder = network.DecoderCell()

encoder.eval()
binarizer.eval()
decoder.eval()

print("Khawar", torch.cuda.current_device())
#torch.cuda.set_device(2)

#print("Khawar", torch.cuda.current_device())

encoder.load_state_dict(torch.load(args.model))
binarizer.load_state_dict(
    torch.load(args.model.replace('encoder', 'binarizer')))
decoder.load_state_dict(torch.load(args.model.replace('encoder', 'decoder')))

encoder_h_1 = (Variable(
    torch.zeros(batch_size, 256, height // 4, width // 4), volatile=True),
               Variable(
                   torch.zeros(batch_size, 256, height // 4, width // 4),
                   volatile=True))
encoder_h_2 = (Variable(
    torch.zeros(batch_size, 512, height // 8, width // 8), volatile=True),
               Variable(
                   torch.zeros(batch_size, 512, height // 8, width // 8),
                   volatile=True))
encoder_h_3 = (Variable(
    torch.zeros(batch_size, 512, height // 16, width // 16), volatile=True),
               Variable(
                   torch.zeros(batch_size, 512, height // 16, width // 16),
                   volatile=True))

decoder_h_1 = (Variable(
    torch.zeros(batch_size, 512, height // 16, width // 16), volatile=True),
               Variable(
                   torch.zeros(batch_size, 512, height // 16, width // 16),
                   volatile=True))
decoder_h_2 = (Variable(
    torch.zeros(batch_size, 512, height // 8, width // 8), volatile=True),
               Variable(
                   torch.zeros(batch_size, 512, height // 8, width // 8),
                   volatile=True))
decoder_h_3 = (Variable(
    torch.zeros(batch_size, 256, height // 4, width // 4), volatile=True),
               Variable(
                   torch.zeros(batch_size, 256, height // 4, width // 4),
                   volatile=True))
decoder_h_4 = (Variable(
    torch.zeros(batch_size, 128, height // 2, width // 2), volatile=True),
               Variable(
                   torch.zeros(batch_size, 128, height // 2, width // 2),
                   volatile=True))

if args.cuda:
    encoder = encoder.cuda()
    binarizer = binarizer.cuda()
    decoder = decoder.cuda()

    image = image.cuda()

    encoder_h_1 = (encoder_h_1[0].cuda(), encoder_h_1[1].cuda())
    encoder_h_2 = (encoder_h_2[0].cuda(), encoder_h_2[1].cuda())
    encoder_h_3 = (encoder_h_3[0].cuda(), encoder_h_3[1].cuda())

    decoder_h_1 = (decoder_h_1[0].cuda(), decoder_h_1[1].cuda())
    decoder_h_2 = (decoder_h_2[0].cuda(), decoder_h_2[1].cuda())
    decoder_h_3 = (decoder_h_3[0].cuda(), decoder_h_3[1].cuda())
    decoder_h_4 = (decoder_h_4[0].cuda(), decoder_h_4[1].cuda())

codes = []
res = image - 0.5
for iters in range(args.iterations):
    encoded, encoder_h_1, encoder_h_2, encoder_h_3 = encoder(
        res, encoder_h_1, encoder_h_2, encoder_h_3)

    code = binarizer(encoded)

    output, decoder_h_1, decoder_h_2, decoder_h_3, decoder_h_4 = decoder(
        code, decoder_h_1, decoder_h_2, decoder_h_3, decoder_h_4)

    res = res - output
    codes.append(code.data.cpu().numpy())

    print('Iter: {:02d}; Loss: {:.06f}'.format(iters, res.data.abs().mean()))

codes = (np.stack(codes).astype(np.int8) + 1) // 2

export = np.packbits(codes.reshape(-1))

np.savez_compressed(args.output, shape=codes.shape, codes=export)[!

Error image

It may be due to your retaining all the graphs of all iterations when you do res=res - output. Can you try res = res - output.detach() ? Also, can you monitor the memory usage of your gpu (with nvidia-smi -l 1 for example) and check whether the memory usage increases linearly with iterations ? — trialNerror
I implemented your approach and watch memory usage. GPU 0 usage is 100% — Khawar Islam

David David · Accepted Answer · 2020-08-09T21:45:40

Try using torchsummary (pip install torchsummary).

from torchsummary import summary

summary(model)

edit: summary(model, (3,299,299)) but replace (3,299,299) with shape of your samples.

It will give you the size of each of the models. You want to see if the forward and back pass size in memory are too large for your gpu. Especially since you are using three models (encoder, binarizer, and decoder).

Iteration on images with Pytorch: error due to CUDA memory issue with batch size 1

1 Answers