matlab/RBM/jbatch.m

function jbatch(numSamplesPerBatch, numBatches)
numLabels = 10;

maxNumSamples = -1;
for d=1:numLabels,
  Dall{d} = load(['test' num2str(d-1) '.ascii'],'-ascii');
  fprintf('%5d Digits of class %d\n',size(Dall{d},1),d-1);
  maxNumSamples = max(maxNumSamples, size(Dall{d},1));
%  jconv(Dall{d}/256, d-1);
%  jimage(Dall{d}/256, d-1, 28, 28);
end;

already_used = zeros(numLabels, maxNumSamples);

for k=1:numBatches,
	s = ['mnist_batch_' num2str(k) '.trainingStates.dat']
	fid = fopen(s, 'w');
	fprintf(fid, '%d\n', numSamplesPerBatch);

	for m=1:numSamplesPerBatch,
		label = floor(numLabels*rand()) + 1;
		D = Dall{label};
		[numTrain, numPixel] = size(D);
		while(1)
			sample = floor(numTrain*rand()) + 1;
			if already_used(label, sample) == 0
				break;
			end
		end
		already_used(label, sample) = 1;
		jwrite(fid, D(sample, :)/256);
	end
	fclose(fid);
end

for k=1:numBatches,
	s = ['mnist_retina_batch_' num2str(k) '.trainingStates.dat']
	fid = fopen(s, 'w');
	fprintf(fid, '%d\n', numSamplesPerBatch);

	for m=1:numSamplesPerBatch,
		label = floor(numLabels*rand()) + 1;
		D = Dall{label};
		[numTrain, numPixel] = size(D);
		while(1)
			sample = floor(numTrain*rand()) + 1;
			if already_used(label, sample) == 0
				break;
			end
		end
		already_used(label, sample) = 1;
        jwrite_retina(fid, D(sample, :)/256, 28, 28);
	end
	fclose(fid);
end

s = ['mnist_batch_0_9.trainingStates.dat']
fid = fopen(s, 'w');
fprintf(fid, '%d\n', numLabels*numSamplesPerBatch);

already_used = zeros(10, maxNumSamples);
for label=1:numLabels,
	for m=1:numSamplesPerBatch,
		D = Dall{label};
		[numTrain, numPixel] = size(D);
		while(1)
			sample = floor(numTrain*rand()) + 1;
			if already_used(label, sample) == 0
				break;
			end
		end
		already_used(label, sample) = 1;
		jwrite(fid, D(sample, :)/256);
	end
end
fclose(fid);