#issue help wanted 8 reference for mask matrix of 1s

[bibuzz]

@fsalv @EscVM
With reference to #8 (comment)
Matrix of how many ones should I set cropped_y_mask as ?

cropped_y_mask = tf.cast([1,1,1,1,1], tf.float32)

    cropped_predictions_masked = tf.cast(
        cropped_predictions, tf.float32)*cropped_y_mask
    cropped_labels_masked = cropped_labels*cropped_y_mask

    total_pixels_masked = tf.reduce_sum(cropped_y_mask, axis=[1, 2])

I am getting error
InvalidArgumentError: Invalid reduction dimension (1 for input with 1 dimension(s) [Op:Sum]
D:\SuperResolution\RAMS\utils\loss.py in l1_loss(y_true, y_pred, HR_SIZE)
50 cropped_labels_masked = cropped_labels*cropped_y_mask
51
---> 52 total_pixels_masked = tf.reduce_sum(cropped_y_mask, axis=[1, 2])
53
54 # bias brightness

I tried all combos
cropped_y_mask =[1,1,1,1,1]
cropped_y_mask =[1]
cropped_y_mask =[1,1]

FullError.txt
Any help appreciated

[fsalv]

cropped_y_mask should have the same dimension of the output, so B x HR_SIZE x HR_SIZE x 1. All the combos you tried are one dimensional, so they won't work. Try using np.ones.

[bibuzz]

Hi @fsalv @EscVM

I printed out the crop_prediction and the y_mask to see if dimensions same, and it does
show the same now (48, 90, 90, 1),
Like you said batch_sizehr_sizehr_size*1
(here my hr_size = 96, but some border pixel shift is done to make it 90 i guess
I did try the other option as well (48, 96, 96, 1) and it resulted in same error)

cropped_y_mask = np.ones(shape=(48, 90, 90, 1), dtype=np.float32)

I have attached the loss.py and the complete error

Output
------
Epoch 1/100
crop predictions tf.Tensor(
[[[[4612.5356 ]
[3837.1716 ]
[4633.922 ]
...
[2775.706 ]
[2106.822 ]
[2689.7336 ]]
...
[2897.199 ]
[2584.3728 ]
[2875.7666 ]]]], shape=(48, 90, 90, 1), dtype=float32)

crop y mask tf.Tensor(
[[[[1.]
[1.]
[1.]
...
[1.]
[1.]
[1.]]
...
[1.]]]], shape=(48, 90, 90, 1), dtype=float32)

Error
------
D:\SuperResolution\RAMS\utils\loss.py in l1_loss(y_true, y_pred, HR_SIZE)
48 print("crop predictions",tf.cast(cropped_predictions, tf.float32));
49 print("crop y mask", cropped_y_mask);
---> 50 cropped_predictions_masked = tf.cast(
51 cropped_predictions, tf.float32)cropped_y_mask
52 cropped_labels_masked = cropped_labelscropped_y_mask

InvalidArgumentError: required broadcastable shapes at loc(unknown) [Op:Mul]
loss1.txt
outputwitherror.txt