6

I'm trying to feed images of dimension (350x350x3) as the Input shape and I want to train the network to output a (1400x1400x3) image (4x upscale).

My training dataset consists of 8 images 1400x1400x3 which I flip around and such to get a total of 32 images for validation.

Then, I scale those 32 images down to 350x350x3 to obtain the input images which will be cross-validated with their 32 other counterparts.

print(type(validateData))
print(validateData.shape)
print(type(validateData[0].shape))
print(validateData[0].shape)

returns

<class 'numpy.ndarray'>
(32,)
<class 'tuple'>
(1400, 1400, 3)

And, similarly:

print(type(trainingData))  # <class 'numpy.ndarray'>
print(trainingData.shape)  # (32,)
print(type(trainingData[0].shape))  # <class 'tuple'>
print(trainingData[0].shape)  # (350, 350, 3)

So when I do

model.fit(trainingData,
          validateData,
          epochs=5,
          verbose=2,
          batch_size=4)  # 32 images-> 8 batches of 4

What exactly am I supposed to feed as the two first parameters of the .fit function?

As it is, I am getting this error:

ValueError: Error when checking input: expected input_1 to have 4 dimensions, but got array with shape (32, 1)

Here is my full code if you want to look into it.

Keras API isn't very explicit about the formatting of the data that should be fed:

fit
fit(x=None, y=None, batch_size=None, epochs=1, verbose=1, callbacks=None, validation_split=0.0, validation_data=None, shuffle=True, class_weight=None, sample_weight=None, initial_epoch=0, steps_per_epoch=None, validation_steps=None)
Trains the model for a given number of epochs (iterations on a dataset).

Arguments

x: Numpy array of training data (if the model has a single input), or list of Numpy arrays (if the model has multiple inputs). If input layers in the model are named, you can also pass a dictionary mapping input names to Numpy arrays.  x can be None (default) if feeding from framework-native tensors (e.g. TensorFlow data tensors).
y: Numpy array of target (label) data (if the model has a single output), or list of Numpy arrays (if the model has multiple outputs). If output layers in the model are named, you can also pass a dictionary mapping output names to Numpy arrays.  y can be None (default) if feeding from framework-native tensors (e.g. TensorFlow data tensors).

Here is the full code of another implementation that I've tried using. This time, I changed the parameters to be python lists of np_arrays (each image being the 3D np_array). I now get this error:

ValueError: Error when checking model input: the list of Numpy arrays that you are passing to your model is not the size the model expected. Expected to see 1 array(s), but instead got the following list of 32 arrays: [array([[[0.6774938 , 0.64219969, 0.60690557],
        [0.67257049, 0.63743775, 0.60206295],
        [0.67203473, 0.6418085 , 0.60398018],
        ...,
        [0.55292714, 0.5253832 , 0.46217287],
  ...

It's hard to know if I'm closer or further.

payne
  • 4,691
  • 8
  • 37
  • 85

1 Answers1

3
fit(x=None, y=None, batch_size=None, epochs=1, verbose=1, callbacks=None, validation_split=0.0, validation_data=None, shuffle=True, class_weight=None, sample_weight=None, initial_epoch=0, steps_per_epoch=None, validation_steps=None)

In the fit function x is your training data. Perhaps it is more appropriately called X_train or something similar rather than trainData. In this case if you have 32 training images of shape (350, 350, 3) they should be stacked in a four dimensional (32, 350, 350, 3) array. If you have an array of your images as objects like you seem to have you can reshape using the following code

X_train = np.array([x for x in trainingData])

Then y in the fit function is the values that you want your network to output for the corresponding input, usually a label and in this case the larger images. You call it validateData but it's more commonly called something closer to y_train. Validation data is data that is not used during training but only during evaluation of the model after every epoch. So shape y_train in the same way that X_train was shaped. Expected shape of y_train would be (32, 1400, 1400, 3).

dennissv
  • 756
  • 5
  • 12
  • `should be stacked in a four dimensional (32, 350, 350, 3) array. If you have an array of your images` : this is precisely where I'm getting lost. I do not understand what python data type I should be using as input (np_array, or list of each image as np_array, or list of each image as arrays, or array of images as np_array, or... etc.). – payne Sep 21 '18 at 23:13
  • 1
    It should be a single four dimensional numpy array of shape (nr_of_images, image_height, image_width, color_channels). You seem to have a numpy array of 32 objects. I've edited another piece of code in if the reshape command doesn't work (did you try it?). – dennissv Sep 21 '18 at 23:19
  • Also realized I had some typos in the code examples, sorry about that. Should be fixed now. – dennissv Sep 21 '18 at 23:32
  • not working. Would you mind [joining the chat here](https://chat.stackoverflow.com/rooms/169987/dawgs-waffle-haus) so that we do not spam the comment section? :) – payne Sep 21 '18 at 23:39
  • For clarification purposes, I would like to let the community know that sds solved my issue. My numpy array was created with 2 different sized images, and -then- I would separate it to separate these 2 classes of images. However, because of the way the array was created, the separation would leave an artifact of `numpy array containing numpy arrays` instead of being a single `numpy array`. The `X_train = np.array([x for x in trainingData])` line of code, applied after the split, worked out! He also solved my issue regarding the input format: one numpy array of size `(num_imgs x W x H x D)`. – payne Sep 23 '18 at 06:08