pytorch out of GPU memory

Question

I am trying to implement Yolo-v2 in pytorch. However, I seem to be running out of memory just passing data through the network. The model is large and is shown below. However, I feel like I'm doing something stupid here with my network (like not freeing memory somewhere). The network works as expected on cpu.

The test code (where memory runs out) is:

x = torch.rand(32,3,416, 416).cuda()
model = Yolov2().cuda()
y = model(x.float())

Question

1. Anything that is obviously wrong with my model?
2. How do I make this more efficient with memory?
3. Other comments?

The model:

import torch
from torch import nn
import torch.nn.functional as F

class Yolov2(nn.Module):

    def __init__(self):
        super(Yolov2, self).__init__()

        self.conv1 = nn.Conv2d(in_channels=3, out_channels=32, kernel_size=3, stride=1, padding=1, bias=False)
        self.batchnorm1 = nn.BatchNorm2d(32)

        self.conv2 = nn.Conv2d(in_channels=32, out_channels=64, kernel_size=3, stride=1, padding=1, bias=False)
        self.batchnorm2 = nn.BatchNorm2d(64)

        self.conv3 = nn.Conv2d(in_channels=64, out_channels=128, kernel_size=3, stride=1, padding=1, bias=False)
        self.batchnorm3 = nn.BatchNorm2d(128)
        self.conv4 = nn.Conv2d(in_channels=128, out_channels=64, kernel_size=1, stride=1, padding=0, bias=False)
        self.batchnorm4 = nn.BatchNorm2d(64)
        self.conv5 = nn.Conv2d(in_channels=64, out_channels=128, kernel_size=3, stride=1, padding=1, bias=False)
        self.batchnorm5 = nn.BatchNorm2d(128)

        self.conv6 = nn.Conv2d(in_channels=128, out_channels=256, kernel_size=3, stride=1, padding=1, bias=False)
        self.batchnorm6 = nn.BatchNorm2d(256)
        self.conv7 = nn.Conv2d(in_channels=256, out_channels=128, kernel_size=1, stride=1, padding=0, bias=False)
        self.batchnorm7 = nn.BatchNorm2d(128)
        self.conv8 = nn.Conv2d(in_channels=128, out_channels=256, kernel_size=3, stride=1, padding=1, bias=False)
        self.batchnorm8 = nn.BatchNorm2d(256)

        self.conv9 = nn.Conv2d(in_channels=256, out_channels=512, kernel_size=3, stride=1, padding=1, bias=False)
        self.batchnorm9 = nn.BatchNorm2d(512)
        self.conv10 = nn.Conv2d(in_channels=512, out_channels=256, kernel_size=1, stride=1, padding=0, bias=False)
        self.batchnorm10 = nn.BatchNorm2d(256)
        self.conv11 = nn.Conv2d(in_channels=256, out_channels=512, kernel_size=3, stride=1, padding=1, bias=False)
        self.batchnorm11 = nn.BatchNorm2d(512)
        self.conv12 = nn.Conv2d(in_channels=512, out_channels=256, kernel_size=1, stride=1, padding=0, bias=False)
        self.batchnorm12 = nn.BatchNorm2d(256)
        self.conv13 = nn.Conv2d(in_channels=256, out_channels=512, kernel_size=3, stride=1, padding=1, bias=False)
        self.batchnorm13 = nn.BatchNorm2d(512)

        self.conv14 = nn.Conv2d(in_channels=512, out_channels=1024, kernel_size=3, stride=1, padding=1, bias=False)
        self.batchnorm14 = nn.BatchNorm2d(1024)
        self.conv15 = nn.Conv2d(in_channels=1024, out_channels=512, kernel_size=1, stride=1, padding=0, bias=False)
        self.batchnorm15 = nn.BatchNorm2d(512)
        self.conv16 = nn.Conv2d(in_channels=512, out_channels=1024, kernel_size=3, stride=1, padding=1, bias=False)
        self.batchnorm16 = nn.BatchNorm2d(1024)
        self.conv17 = nn.Conv2d(in_channels=1024, out_channels=512, kernel_size=1, stride=1, padding=0, bias=False)
        self.batchnorm17 = nn.BatchNorm2d(512)
        self.conv18 = nn.Conv2d(in_channels=512, out_channels=1024, kernel_size=3, stride=1, padding=1, bias=False)
        self.batchnorm18 = nn.BatchNorm2d(1024)

        self.conv19 = nn.Conv2d(in_channels=1024, out_channels=1024, kernel_size=3, stride=1, padding=1, bias=False)
        self.batchnorm19 = nn.BatchNorm2d(1024)
        self.conv20 = nn.Conv2d(in_channels=1024, out_channels=1024, kernel_size=3, stride=1, padding=1, bias=False)
        self.batchnorm20 = nn.BatchNorm2d(1024)

        self.conv21 = nn.Conv2d(in_channels=3072, out_channels=1024, kernel_size=3, stride=1, padding=1, bias=False)
        self.batchnorm21 = nn.BatchNorm2d(1024)

        self.conv22 = nn.Conv2d(in_channels=1024, out_channels=125, kernel_size=1, stride=1, padding=0)

    def reorg_layer(self, x):
        stride = 2
        batch_size, channels, height, width = x.size()
        new_ht = int(height/stride)
        new_wd = int(width/stride)
        new_channels = channels * stride * stride

#         from IPython.core.debugger import Tracer; Tracer()()
        passthrough = x.permute(0, 2, 3, 1)
        passthrough = passthrough.contiguous().view(-1, new_ht, stride, new_wd, stride, channels)
        passthrough = passthrough.permute(0, 1, 3, 2, 4, 5)
        passthrough = passthrough.contiguous().view(-1, new_ht, new_wd, new_channels)
        passthrough = passthrough.permute(0, 3, 1, 2)
        return passthrough

    def forward(self, x):
        out = F.max_pool2d(F.leaky_relu(self.batchnorm1(self.conv1(x)), negative_slope=0.1), 2, stride=2)
        out = F.max_pool2d(F.leaky_relu(self.batchnorm2(self.conv2(out)), negative_slope=0.1), 2, stride=2)

        out = F.leaky_relu(self.batchnorm3(self.conv3(out)), negative_slope=0.1)
        out = F.leaky_relu(self.batchnorm4(self.conv4(out)), negative_slope=0.1)
        out = F.leaky_relu(self.batchnorm5(self.conv5(out)), negative_slope=0.1)
        out = F.max_pool2d(out, 2, stride=2)

        out = F.leaky_relu(self.batchnorm6(self.conv6(out)), negative_slope=0.1)
        out = F.leaky_relu(self.batchnorm7(self.conv7(out)), negative_slope=0.1)
        out = F.leaky_relu(self.batchnorm8(self.conv8(out)), negative_slope=0.1)
        out = F.max_pool2d(out, 2, stride=2)

        out = F.leaky_relu(self.batchnorm9(self.conv9(out)), negative_slope=0.1)
        out = F.leaky_relu(self.batchnorm10(self.conv10(out)), negative_slope=0.1)
        out = F.leaky_relu(self.batchnorm11(self.conv11(out)), negative_slope=0.1)
        out = F.leaky_relu(self.batchnorm12(self.conv12(out)), negative_slope=0.1)
        out = F.leaky_relu(self.batchnorm13(self.conv13(out)), negative_slope=0.1)
#         from IPython.core.debugger import Tracer; Tracer()()
        passthrough = self.reorg_layer(out)
        out = F.max_pool2d(out, 2, stride=2)

        out = F.leaky_relu(self.batchnorm14(self.conv14(out)), negative_slope=0.1)
        out = F.leaky_relu(self.batchnorm15(self.conv15(out)), negative_slope=0.1)
        out = F.leaky_relu(self.batchnorm16(self.conv16(out)), negative_slope=0.1)
        out = F.leaky_relu(self.batchnorm17(self.conv17(out)), negative_slope=0.1)
        out = F.leaky_relu(self.batchnorm18(self.conv18(out)), negative_slope=0.1)

        out = F.leaky_relu(self.batchnorm19(self.conv19(out)), negative_slope=0.1)
        out = F.leaky_relu(self.batchnorm20(self.conv20(out)), negative_slope=0.1)

        out = torch.cat([passthrough, out], 1)
        out = F.leaky_relu(self.batchnorm21(self.conv21(out)), negative_slope=0.1)
        out = self.conv22(out)

        return out

Additional Info:

• Torch version is '0.4.1.post2'
• Running on aws p2.xlarge (limit 12gb GPU memory).
• The number of parameters for this model is 67137565. Which would occupy <500MB.
• this thread from pytorch might be related.

Answer 1

I would try to use smaller batch sizes. Start from 1 and then check what is your maximum. I can also try to reduce your input Tensor dimensions. Your network is not so small for your GPU

Answer 2

Here is the cycle you may try out:

ma = torch.cuda.memory_allocated()
print(ma)

mc = torch.cuda.memory_cached()
print(mc)

torch.cuda.empty_cache()

ma = torch.cuda.memory_allocated()
print(ma)

mc = torch.cuda.memory_cached()
print(mc)

# 653475518
# 952107008
# 383533568
# 385875968

See how memory can be freed. The other technique is called on demand loading of batches to GPU. So not the whole dataset, but a single batch or several batches at once.

Loading a single batch into GPU you basically measure in μ seconds, while tensor operations like matrix multiply are m seconds.