def trainSamples(primaryDatasets, primaryClasses, primaryInstances,
helperInstances):
for dataSet in primaryDataSets:
for cls in primaryClasses:
for instance in primaryInstances:
for helperInstance in helperInstances:
# if the number of primary instances are larger than the number of helper
# instances, no need to calculate MMD
if instance > helperInstance:
continue
# get a fixed helper class for a particular dataset and primary class
helperClass = getHelperClass(dataSet, cls)
if helperClass == -1:
continue
print(
'Primary Class: {} Helper Class: {} Primary Instances: {} Helper Instance {}'
.format(cls, helperClass, instance, helperInstance))
primaryFileName = dataSet + '_' + str(cls) + '_' + str(
instance)
helperFileName = dataSet + '_' + str(
helperClass) + '_' + str(helperInstance)
dataFolder = rootDir + str(dataSet)
x = loadDataset(dataSet, cls, instance)
y = loadDataset(dataSet, helperClass, helperInstance)
primaryTrainLoader = torch.utils.data.DataLoader(
x,
batch_size=batchSize,
shuffle=True,
num_workers=4,
drop_last=True)
helperTrainLoader = torch.utils.data.DataLoader(
y,
batch_size=batchSize,
shuffle=True,
num_workers=4,
drop_last=True)
numOutputChannels = getChannels(dataSet)
epochs = getEpochs(dataSet, instance)
print epochs
train(primaryFileName,
helperFileName,
primaryTrainLoader,
helperTrainLoader,
instance,
numOutputChannels,
epochs=epochs)
def getFakeData(dataSet, fakeClasses, instances, mmdFlag = 0, numHelperInstances=1000):
'''
Output image pixels between (-1,1)
'''
dataFolder = resultDir+'results/nonMMD/compressed/'+dataSet+'/'+dataSet
if mmdFlag==1:
dataFolder = resultDir+'results/MMDall/compressed/'+dataSet+'/'+dataSet
fakeClasses = sorted(fakeClasses)
imageArray, labelArray = getInitialArray(dataSet)
for i in fakeClasses:
if mmdFlag==1:
tempImageArray = np_load(dataFolder+'_'+str(i)+'_'+str(getHelperClass(dataSet,i))+'_'+str(instances)+'_'+str(numHelperInstances)+'.npy')
elif mmdFlag==0:
tempImageArray = np_load(dataFolder+'_'+str(i)+'_'+str(instances)+'.npy')
tempImageArray = tempImageArray.reshape(tempImageArray.shape[0],-1)
tempLabelArray = np.empty(tempImageArray.shape[0])
tempLabelArray.fill(i)
#print tempImageArray.shape, imageArray.shape
imageArray = np.concatenate([imageArray, tempImageArray])
labelArray = np.concatenate([labelArray, tempLabelArray])
imageArray = imageArray[1:]
labelArray = labelArray[1:]
# random shuffling of images and labels
p = np.random.permutation(imageArray.shape[0])
imageArray = imageArray[p]
labelArray = labelArray[p]
return imageArray, labelArray
sync(fh)
if __name__=='__main__':
from model_28 import Generator, Discriminator
numGenFilter=64
numDiscFilter=32
imageSize = 28
primaryDataSet = ['MNIST']
helperDataSet = ['SVHN-BW']
primaryClass = [1]
primaryInstances = [5000]
helperInstances = [5000]
batchSizes = [50]
for d in primaryDataSet:
for hd in helperDataSet:
for pc in primaryClass:
for pi in primaryInstances:
for hi in helperInstances:
if pi > hi:
continue
for b in batchSizes:
hc = getHelperClass(d,pc)
if hc==-1:
continue
test(d, hd, pc, hc, pi, hi)
def getFakeData(dataSet, fakeClasses, instances, mmdFlag = 0, numHelperInstances=1000):
'''
Output image pixels between (-1,1)
'''
dataFolder = resultDir+'results/nonMMD/compressed/'+dataSet+'/'+dataSet
if mmdFlag==1:
dataFolder = resultDir+'results/MMD/compressed/'+dataSet+'/'+dataSet
fakeClasses = sorted(fakeClasses)
imageArray, labelArray = getInitialArray(dataSet)
for i in fakeClasses:
if mmdFlag==1:
tempImageArray = np_load(dataFolder+'_'+str(i)+'_'+str(getHelperClass(dataSet,i))+'_'+str(instances)+'_'+str(numHelperInstances)+'.npy')
elif mmdFlag==0:
tempImageArray = np_load(dataFolder+'_'+str(i)+'_'+str(instances)+'.npy')
# no need to resize as model is producing image of same size
'''
if getChannels(dataSet)==1:
tempImageArrayResize = np.zeros((tempImageArray.shape[0],
sizeOfImage,
sizeOfImage))
for j in range(tempImageArray.shape[0]):
tempImage = PIL.Image.fromarray(np.uint8(tempImageArray[j,:,:]))
tempImage = tempImage.resize((sizeOfImage,sizeOfImage),
PIL.Image.ANTIALIAS)
tempImageArrayResize[j,:,:] = np.asarray(tempImage)
elif getChannels(dataSet)==3:
tempImageArrayResize = np.zeros((tempImageArray.shape[0],
sizeOfImage,
sizeOfImage,
3))
tempImageArray = tempImageArray.transpose([0,2,3,1])
for j in range(tempImageArray.shape[0]):
tempImage = PIL.Image.fromarray(np.uint8(tempImageArray[j,:,:,:]), 'RGB')
tempImage = tempImage.resize((sizeOfImage,sizeOfImage),
PIL.Image.ANTIALIAS)
tempImageArrayResize[j,:,:,:] = np.asarray(tempImage)
tempImageArray = tempImageArray.transpose([0,3,1,2])
'''
tempImageArray = tempImageArray.reshape(tempImageArray.shape[0],-1)
tempLabelArray = np.empty(tempImageArray.shape[0])
tempLabelArray.fill(i)
#print tempImageArray.shape, imageArray.shape
imageArray = np.concatenate([imageArray, tempImageArray])
labelArray = np.concatenate([labelArray, tempLabelArray])
imageArray = imageArray[1:]
labelArray = labelArray[1:]
# random shuffling of images and labels
p = np.random.permutation(imageArray.shape[0])
imageArray = imageArray[p]
labelArray = labelArray[p]
return imageArray, labelArray