Hi, I reduced the workspace to 1024 and batch_size to 1 - I get the memory error same as before. I went back to using the whole batch and its running again.
Thx for the other stuff will look into that now.. On Friday, 8 April 2016 09:53:18 UTC+2, Valentin Churavy wrote: > > Also take a look at https://github.com/dmlc/MXNet.jl/pull/73 where I > implemented debug_str in Julia so that you can test your network on its > space requirements. > > On Friday, 8 April 2016 15:54:06 UTC+9, Valentin Churavy wrote: >> >> What happens if you set the batch_size to 1? Also take a look at >> https://github.com/dmlc/mxnet/tree/master/example/memcost >> >> Also workspace is per convolution and you should keep it small. >> >> On Thursday, 7 April 2016 19:13:36 UTC+9, kleinsplash wrote: >>> >>> Hi, >>> >>> I have a memory error using Quadro K5000M which has 4GB global memory. I >>> was wondering if there is some guide as to how to set my workspace and >>> Convolutional layers. >>> >>> My current settings: >>> >>> training_data = 128x128x1x800 >>> batch_size = 128x128x1x8 >>> workspace = 2048 (I think this can go up to 4096 because of the >>> .deviceQuery) >>> >>> This is my net (still to be designed so just basic ): >>> >>> # first conv >>> conv1 = @mx.chain mx.Convolution(data=data, kernel=(5,5), num_filter=20, >>> workspace=workspace) => >>> mx.Activation(act_type=:relu) => >>> mx.Pooling(pool_type=:max, kernel=(2,2), stride=(2,2)) >>> # second conv >>> conv2 = @mx.chain mx.Convolution(data=conv1, kernel=(5,5), num_filter=50 >>> , workspace=workspace) => >>> mx.Activation(act_type=:relu) => >>> mx.Pooling(pool_type=:max, kernel=(2,2), stride=(2,2)) >>> # first fully-connected >>> fc1 = @mx.chain mx.Flatten(data=conv2) => >>> mx.FullyConnected(num_hidden=500) => >>> mx.Activation(act_type=:relu) >>> # second fully-connected >>> fc2 = mx.FullyConnected(data=fc1, num_hidden=10) >>> # third fully-connected >>> fc3 = mx.FullyConnected(data=fc2, num_hidden=2) >>> # softmax loss >>> net = mx.SoftmaxOutput(data=fc3, name=:softmax) >>> >>> So far if I reduce my image to 28x28 it all works - but I need to up the >>> resolution to pick out features. Anyone have any ideas on thumb sucking >>> initial values for at least getting past memory issues to the design of the >>> net? >>> >>> >>> Thx >>> >>>
