CODEWITHSUNDEEP
pythonnumpytensorflowmath
Tejas Pandey
Tejas Pandey

Reputation: 37

Tensor Scalar Multiplication Tensor Flow

I'm currently trying to implement my own loss function.

I have three tensors.

A [batch, row, col, keypoints] # Actual Values
B [batch, row, col, keypoints] # Predicted Values
C [batch, keypoints_mask]      # Mask

keypoints_mask is either 1 or 0. I want to treat the tensors as arrays and do scalar multiplication of the last dimension.

E.g something like this:

A [5, 100, 100, 10]
B [5, 100, 100, 10]
C [5, 10]

A[-1][0] = A[-1][0] * C[-1][0]
A[-1][1] = A[-1][1] * C[-1][1]
...

B[-1][0] = B[-1][0] * C[-1][0]
B[-1][1] = B[-1][1] * C[-1][1]
...

Loss = Mean_Squared_Error(A, B)

What would be the best approach to do implement this?

Edit:

The data is an image, where for every pixel I have 10 values.

Psuedo Code

for b in batch:
    for r in row:
        for c in col:
            for i in enumerate(keypoints):
                A[b, r, c, i] = A[b, r, c, i] * C[b, i]
                B[b, r, c, i] = B[b, r, c, i] * C[b, i]

Upvotes: 0

Views: 260

Answers (1)

Tejas Pandey
Tejas Pandey

Reputation: 37

This is what I ended up doing and it seems to work for now.

A [5, 100, 100, 10] # Actual
B [5, 100, 100, 10] # Predicted
C [5, 10]           # Mask

Loss = A - B
Loss = Loss * Loss
Loss = tf.reduce_mean(Loss, [1,2]) # [5, 100, 100, 10] -> [5, 10]
Loss = Loss * C

Upvotes: 1

Related Questions