Question

我正在按照Jeff Heaton's tutorial的方法使用keras创建GAN。即使使用我自己的数据集，一切都可以正常工作。但是，我不知道如何创建一个新图像。（希顿先生以拼贴的形式创建了28张图像！）

我尝试失败的尝试：

new_fixed_seed = np.random.normal(0, 1, (1, 100))
generated_images = generator.predict(new_fixed_seed)
im = Image.fromarray(generated_images)

结果：TypeError: Cannot handle this data type

我在做什么错了？

Answer 1

通常，对我来说，当我计算generated images时，我正在使用以下代码将它们存储在本地：

# combine a squared number of images.
def combine_images(generated_images):

  generated_images =  np.transpose(generated_images , (0, 3, 1, 2))

  num = generated_images.shape[0]
  width = int(math.sqrt(num))
  height = int(math.ceil(float(num)/width))
  shape = generated_images.shape[2:]
  image = np.zeros((3,(height+3)*shape[0], (width+3)*shape[1]),
                 dtype=generated_images.dtype)

  for index, img in enumerate(generated_images):

    new_shape = (img.shape[0], img.shape[1]+ 4, img.shape[2] + 4)
    img_ = np.zeros(new_shape)
    img_[:, 2:  2+img.shape[1], 2:  2+img.shape[2]] = img

    i = int(index/width)
    j = index % width
    image[:, i*new_shape[1]: (i+1)*new_shape[1], j*new_shape[2]: (j+1)*new_shape[2]] = img_[:, :, :]
  return image

# store combined images
def store_image_maps(images_db, filename):
  image = combine_images(images_db)
  image = image * 127.5 + 127.5
  image = np.swapaxes(image, 0, 1)
  image = np.swapaxes(image, 1, 2)
  cv2.imwrite(filename,image)

Answer 2

我可以使用它，但是我并不完全满意，因为我认为它可能更干净，而且我认为其中涉及不必要的步骤：

# SEED_SIZE is 100
fixed_seed = np.random.normal(0, 1, (1, SEED_SIZE))

# used 64x64 because those were my image sizes
image_array = np.full((   
      64,64, 3),           
      255, dtype=np.uint8)

generated_images = generator.predict(fixed_seed)

#if you don't use 255 here the images are black
image_array[:] = generated_images * 255  

im = Image.fromarray(image_array)

im.save('/content/drive/My Drive/Dataset/test.png')

如何使用具有权重的成品GAN模型创建新图像？

2 个答案: