[luci/service] Support BMM dynamic shape inferece

[zetwhite]

This PR supports dynamic shpae inference ofr BatchMatMul Op.

ONE-DCO-1.0-Signed-off-by: SeungHui Youn [email protected]

for : #13697

[shs-park]

I made some test cases with tensorflow python code.
I could get exceptions in Test 4 and Test 6 cases.

Please note when you make test cases.

Python Test Code

import tensorflow as tf

# Example 1: Simple 2D Matrices Batch
A = tf.constant([[[1, 2, 3], [4, 5, 6]],
                 [[7, 8, 9], [10, 11, 12]]], dtype=tf.float32)  # Shape: (2, 2, 3)

B = tf.constant([[[1, 4], [2, 5], [3, 6]],
                 [[7, 10], [8, 11], [9, 12]]], dtype=tf.float32)  # Shape: (2, 3, 2)

# Perform batch matrix multiplication using tf.matmul
result = tf.matmul(A, B)
print("Example 1 Result (Shape: {}):\n{}".format(result.shape, result.numpy()))

# Example 2: Broadcasting in Batch Dimensions
A = tf.constant([[[1, 2, 3], [4, 5, 6]],
                 [[7, 8, 9], [10, 11, 12]],
                 [[13, 14, 15], [16, 17, 18]]], dtype=tf.float32)  # Shape: (3, 2, 3)

B = tf.constant([[[1, 4], [2, 5], [3, 6]]], dtype=tf.float32)  # Shape: (1, 3, 2)

# Perform batch matrix multiplication with broadcasting using tf.matmul
result = tf.matmul(A, B)
print("Example 2 Result (Shape: {}):\n{}".format(result.shape, result.numpy()))

# Example 3: Edge case with zero-dimension batch
A = tf.random.normal([0, 2, 3])  # Shape: (0, 2, 3)
B = tf.random.normal([0, 3, 2])  # Shape: (0, 3, 2)

# Perform batch matrix multiplication using tf.matmul
result = tf.matmul(A, B)
print("Example 3 Result (Shape: {}):\n{}".format(result.shape, result.numpy()))

# Example 3-2: Edge case with zero-dimension batch
A = tf.random.normal([3, 2, 3])  # Shape: (3, 2, 3)
B = tf.random.normal([3, 2])  # Shape: (3, 2)

# Perform batch matrix multiplication using tf.matmul
result = tf.matmul(A, B)
print("Example 3-2 Result (Shape: {}):\n{}".format(result.shape, result.numpy()))

# Example 4: Edge case with zero-dimension batch
A = tf.random.normal([2, 2, 3])  # Shape: (2, 2, 3)
B = tf.random.normal([0, 3, 2])  # Shape: (0, 3, 2)

# Perform batch matrix multiplication using tf.matmul
#result = tf.matmul(A, B)
print("Example 4 Result EXCEPTION")

# Example 5: Edge case with rank-2 batch
A = tf.random.normal([2, 3])  # Shape: (2, 3)
B = tf.random.normal([3, 2])  # Shape: (3, 2)

# Perform batch matrix multiplication using tf.matmul
result = tf.matmul(A, B)
print("Example 5 Result (Shape: {}):\n{}".format(result.shape, result.numpy()))

# Example 6: Edge case with rank-2 and rank-1 atch
A = tf.random.normal([2, 3])  # Shape: (2, 3)
B = tf.random.normal([3])  # Shape: (3)

# Perform batch matrix multiplication using tf.matmul
#result = tf.matmul(A, B)
print("Example 6 Result EXCEPTION")

[shs-park]

I left some comments, PTAL
=)

[jinevening]

LGTM with minor comment https://github.com/Samsung/ONE/pull/13759/files#r1733773189

[seanshpark]

LGTM

[zetwhite]

(note) to resolve https://github.com/Samsung/ONE/pull/13759/files#r1732922382

[zetwhite]

(newly added) to handle empty input

[zetwhite]

@ys44kim I added the TC for the case you asked about. :)

[seanshpark]

LGTM

[shs-park]

Thank you, LGTM!
=)

[jinevening]

LGTM

[jinevening]

Ping @ys44kim

[ys44kim]

LGTM

[seanshpark]

@zetwhite , please check #13780 (review)

can you apply here too? or as getting approvals takes time and you've go all ACKs,
apply in another PR?

[seanshpark]

if go with another PR, you can apply changes to all Ops (I think there aren't too many)

[zetwhite]

I'd like to merge this PR as is. (not for waiting approvals again)
If fine, I'll make another PR to fix it.

[zetwhite]

you can apply changes to all Ops (I think there aren't too many)

Oh, I see. I apply it at once.

[seanshpark]

ah... commit title message had typo ... -_-;

This PR supports dynamic shpae inference ofr BatchMatMul Op.