Generative-Moment-Matching-Networks
Generative-Moment-Matching-Networks copied to clipboard
Support for CIFAR10?
Hello, I am trying to train your model on CIFAR10 by loading cifar picture and modify some hyperparameter in trainDataSpaceNetwork
and generateFigure
def loadCIFAR():
with open("cifar-10-batches-py/data_batch_1", 'rb') as f:
dict = pickle.load(f, encoding='bytes')
data1 = dict[b'data']
with open("cifar-10-batches-py/data_batch_2", 'rb') as f:
dict = pickle.load(f, encoding='bytes')
data2 = dict[b'data']
with open("cifar-10-batches-py/data_batch_3", 'rb') as f:
dict = pickle.load(f, encoding='bytes')
data3 = dict[b'data']
with open("cifar-10-batches-py/data_batch_4", "rb") as f:
dict = pickle.load(f, encoding='bytes')
data4 = dict[b'data']
with open("cifar-10-batches-py/data_batch_5", "rb") as f:
dict = pickle.load(f, encoding='bytes')
data5 = dict[b'data']
return np.concatenate((data1, data2, data3, data4, data5), axis=0)
def trainDataSpaceNetwork(dataset):
# something...
elif dataset == 'cifar':
input_dim = 3072
image_side = 32
num_examples = 50000
train_x = loadCIFAR()
but the generated pictures are of very low quality, just like noise. So have you used your model for CIFAR10? Should I change other hyperparameter like num_iterations
or batch_size
?
Thank You.