Is this the problem about graphic card

I am using "Object Detection Using Faster R-CNN Deep Learning" example code. In this code, the convolution layer and the pooling layer are slightly modified by me. My graphics card is the GTX660.

I tested with two sets of data. The total size of A data is 290MB and the total size of B data is 23MB.

I monitored the load of the graphics card to see if there was a problem with the graphics card. A data in this code yields the following results. MATLAB appeared busy for more than 30 minutes in the status window. At that time, only 1 to 3% of the graphics card load was available. When the code halted with an error indication of out-of-memory, the load instantly reached 100%.

And, B data in this code yields the following results. The numbers processed within 30 minutes appear on the screen. At this time, the load of the graphics card was between 70% and 90%.

What I'm curious about is: When I using A data, the graphical load initially reaches 1% to 3% but suddenly reaches 100%.

Please, Somebody help me.

Answer 1

[Vikaasa Kumar]

Hello Jungmoon,

Based on this information, it appears that this issue is likely caused due to the following reasons:

1) Large input dimensions of the images.

2) Limited memory availability on the GPU card.

3) Size and complexity of the network architecture.

Furthermore, the Faster R-CNN method processes the entire input image without resizing. This is in contrast to the 'RCNNObjectDetector' method which performs inherent cropping and resizing of regions to be comparable to the input dimensions of the network.

A couple of workarounds that you may consider to continue using the 'trainFasterRCNNObjectDetector' method are mentioned below:

1) Setting the 'SmallestImageDimension' parameter of the 'trainFasterRCNNObjectDetector' to a value of 400. However, this parameter may need to be adjusted to account for your current GPU specifications. Setting this parameter will resize the images during training and help avoid 'Out of memory' errors.

2) Using a smaller network in addition to setting the 'SmallestImageDimension' parameter as shown in this example: https://www.mathworks.com/help/vision/examples/object-detection-using-faster-r-cnn-deep-learning.html

3) Updating the GPU card. To train a network like AlexNet using the Faster R-CNN method, a GPU with more memory might be required.

I believe that the above suggestions should help resolve this issue.

Best,

Vikaasa

[Jungmoon Kim]

Thank you very much, Vikaasa! Your answer is very useful for me. I will try the method which you mentioned.

I have an additional question. In the second method, there is good example. Actually I am using the example now.

I am curious about the optimal network for me. I want to detect the object which is small and which has not particular features like a ball or a cylinder. Furthermore, the background of the object has some noise.

I am trying to find a optimal network for me by using the example. I used to change the number of filter, the size of filter, the size of inputlayer. And I used to add the another maxpooling layer in the middle layer. I conducted the code and I checked the result. Some networks are good, but other networks are bad. I think that I have been to just try without better plan.

Would you recommend better way for finding solution?