calculate_inception_scores.py 文件源码

def get_batch(self, batch_size):
        print("Loading %d real images of max possible %d..." % (batch_size, self.train_index.size(0)))
        batch = torch.Tensor(batch_size, 3, self.opt.image_size, self.opt.image_size)
        for j in range(batch_size):
            img = self.get_data(self.train_index[self.counter])
            #print(img.size())
            batch[j].copy_(img)
            self.counter = (self.counter + 1) % self.train_index.size(0)
            #print("counter", self.counter)
        #print(np.average(batch.cpu().numpy()), np.min(batch.cpu().numpy()), np.max(batch.cpu().numpy()))
        batch_np = (batch.cpu().numpy() * 255).astype(np.uint8).transpose((0, 2, 3, 1))
        #return Variable(batch.cuda(), volatile=volatile)
        #from scipy import misc
        #misc.imshow(batch_np[0])
        return batch_np

#
# Background processes below
# Tensorflow and pytorch stuff is separated in different processes, because
# otherwies the two will bitch at each other
# The whole measuring of inception scores for checkpoints is in background
# processes, because otherwise python will not free the memory properly and
# run into out of memory errors sooner or later
#