Tensorflow GAN outputting negative numbers in an image, causing PIL errors

Question

I have a GAN set up with Tensorflow and Keras, and training has been smooth. The only problem that I am having is that when I try to convert output tensors into images, I get an error.

Upon further inspection of the output tensor, I find that the image in RGB format contains negative numbers for channel values.

arr = generated_image.numpy() # (128, 128, 3)

# Zeroing in on a specific pixel to examine RGB values

print(arr[0][0][0]) # array([0.80051363, 0.55783302, 0.34086022])

print(arr[0][0][1]) # array([0.1622794 , 0.40731752, 0.41627714])

print(arr[0][0][2]) # array([-22.26079941, -17.90978622, -17.2147789 ])

After looking around on the site a bit, I found this question addressing the error I am having, written below. It said that my types should be adjusted to be in 0-255 RGB instead of a float between 0 and 1. That is why I believe these negative numbers in the array are the cause of the below error.

KeyError                                  Traceback (most recent call last)
/opt/conda/lib/python3.7/site-packages/PIL/Image.py in fromarray(obj, mode)
   2927         try:
-> 2928             mode, rawmode = _fromarray_typemap[typekey]
   2929         except KeyError as e:

KeyError: ((1, 1, 3), '<f8')

The above exception was the direct cause of the following exception:

TypeError                                 Traceback (most recent call last)
/tmp/ipykernel_33/14769953.py in <module>
----> 1 img = Image.fromarray(arr[0])

/opt/conda/lib/python3.7/site-packages/PIL/Image.py in fromarray(obj, mode)
   2928             mode, rawmode = _fromarray_typemap[typekey]
   2929         except KeyError as e:
-> 2930             raise TypeError("Cannot handle this data type: %s, %s" % typekey) from e
   2931     else:
   2932         rawmode = mode

TypeError: Cannot handle this data type: (1, 1, 3), <f8

My question is

1. How did my output tensor include negative numbers and still be a valid image, able to be shown with matplotlib.pyplot.imshow()?

2. How can I adjust the GAN to output the image entirely with floats between 0 and 1?

3. Does this even matter? Is there another image encoding that is being used here? If so, how do I extract images from it?

The generator:

generator = keras.Sequential([
    keras.layers.Dense(8*8*64, use_bias=False, input_shape=(100,)),
    keras.layers.BatchNormalization(),
    keras.layers.LeakyReLU(),
    
    keras.layers.Reshape((8, 8, 64)),
    
    keras.layers.Conv2DTranspose(32, (3, 3), strides=(4, 4), use_bias=False, padding="same"),
    keras.layers.BatchNormalization(),
    keras.layers.LeakyReLU(),
    
    keras.layers.Conv2DTranspose(16, (3, 3), strides=(2, 2), use_bias=False, padding="same"),
    keras.layers.BatchNormalization(),
    keras.layers.LeakyReLU(),
    
    keras.layers.Conv2DTranspose(3, (3, 3), strides=(2, 2), use_bias=False, padding="same"),
    
])

Answer 1

Thanks to @Dr.Snoopy, I have found an answer.

The last layer of my model was missing an activation function that would restrict the values between 0 and 1, which is why I was getting random values.

Simply putting on the sigmoid activation function solved the problem.

keras.layers.Conv2DTranspose(3, (3, 3), strides=(2, 2), use_bias=False, padding="same", activation="sigmoid)```