convolutional-pose-machines-release
convolutional-pose-machines-release copied to clipboard
shihenw
error == cudaSuccess (2 vs 0) out memory
Hi Shihen,
I tried to retrain your model, my GPU version is K80, 24GB, but I got the following error: error == cudaSuccess (2 vs 0) out of memory
Are you using batchsize = 1?
in file cpm_data_layer.cpp:106, there are lines like:
if (batch_size == 1 && crop_size == 0) { Datum& datum = *(reader_.full().peek()); if (datum.encoded()) { if (force_color) { DecodeDatum(&datum, true); } else { DecodeDatumNative(&datum); } } batch->data_.Reshape(1, datum.channels(), datum.height(), datum.width()); this->transformed_data_.Reshape(1, datum.channels(), datum.height(), datum.width()); }
The code will resize the data_ to the origin image size.
No I using batch size at least 4 on each GPU. You are right, that code has problems with batch size 1. Will update.
I got the same error when i forgot to set the USE_CUDNN := 1 in Makefile.config
aronfothi's setting in the building of Caffe is crucial. I tried several times w/o USE_CUDNN on various computers (even one w/ a GPU card of 12 GB)--you must set USE_CUDNN to 1!
Btw, use CUDNN v. 3 and DO NOT use OpenCV 3 per Shih-En's recommendation.
Hello everyone, In gpu mode, one nvidia gtx 1080 which has 8 g graphics memory can use to train the network and real-time forecasting?
If anyone knows, please tell me, much thanks!
Yes, you should be able to train it but testing will not be realtime. Issue #14 talks about realtime testing with 4 GPUs (!).
Hi, @schelian Thanks for you reply, it`s a great help to me. Sir, may i ask a another question? I am confused with ‘multiplier ’in applyModel.m line 26
multiplier = 2.^(log2(starting_scale):(1/octave):log2(ending_scale));
% data container for each scale and stage score = cell(nstage, length(multiplier)); pad = cell(1, length(multiplier)); ori_size = cell(1, length(multiplier));
net = caffe.Net(model.deployFile, model.caffemodel, 'test'); % change outputs to enable visualizing stagewise results % note this is why we keep out own copy of m-files of caffe wrapper
colors = hsv(length(multiplier)); for m = 1:length(multiplier) scale = multiplier(m);
and 'scale_search = 0.7:0.1:1.3; % fit training' in run_benchmark.m line 4
one is by hand drag a bounding box ,one is by pre-setting , i notice that this parameter affects the size of the detection scale, how should I adjust this parameter to fit the size of my photo(e.g 640*480 )?
I don't think you have to adjust the parameter. multiplier is derived from starting_scale (line 26); starting_scale is derived from middle_range (line 21); middle_range is derived from the size of the image (line 20). If you happen to know your testing images have people within a certain scale then you can change the parameter. This will increase speed but could introduce false negatives / misses.