[relax][frontend][tflite] Add tests for l2_normalization/slice/reverse_v2#19371
[relax][frontend][tflite] Add tests for l2_normalization/slice/reverse_v2#193710xjah wants to merge 3 commits intoapache:mainfrom
Conversation
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request adds unit tests for the TFLite frontend in Relax, specifically for L2 normalization, slice, and reverse_v2 operations. The review feedback suggests improving test coverage by parameterizing these tests to handle various input shapes, axes, and edge cases, such as negative size values in the slice operation.
…everse_v2
- Fix test_l2_normalization: Remove expected module using non-existent R.nn.l2_normalize
- l2_normalize is not yet implemented in relax.op.nn
- Keep test for TensorFlow conversion without structural equality check
- Fix test_slice: Add missing 'axes' parameter to strided_slice
- strided_slice() requires axes parameter alongside begin and end
- Changed from R.strided_slice(x, begin=..., end=...)
to R.strided_slice(x, axes=[0, 1], begin=..., end=...)
- Fix test_reverse_v2: Replace non-existent R.reverse with R.flip
- R.reverse doesn't exist in relax.op
- Changed from R.reverse(x, axis=1) to R.flip(x, axis=1)
|
Hi @0xjah Can you have a look at the CI error and fix it? Thanks! |
…se_v2 - Fix convert_l2_normalization: Implement manually since relax.op.nn.l2_normalize doesn't exist - L2 norm formula: output = input / sqrt(sum(input^2) + eps) - Uses existing ops: square, sum, sqrt, divide - Applied along last axis with epsilon=1e-12 - Fix convert_slice: Add missing axes parameter to strided_slice call - strided_slice requires axes alongside begin and end parameters - Changed from strided_slice(data, begin, end) to strided_slice(data, axes=list(range(rank)), begin, end) - Fix convert_reverse_v2: Replace non-existent relax.op.reverse with relax.op.flip - Both operations reverse along specified axis but flip is the available function in relax
|
It looks like the Gemini bot suggestions on The parameterization suggestions are worth doing since single-shape tests don't catch edge cases like |
| # L2 normalization is applied along the last axis | ||
| squared = relax.op.square(in_expr) | ||
| sum_squared = relax.op.sum(squared, axis=input_tensor_rank - 1, keepdims=True) | ||
| denom = relax.op.sqrt(sum_squared + 1e-12) |
There was a problem hiding this comment.
sum_squared is a Relax Expr, so you can't use the + operator with a raw Python float directly. This is what's causing the TypeError: Please convert 1e-12 failure in the CI. You need to wrap the epsilon:
| denom = relax.op.sqrt(sum_squared + 1e-12) | |
| eps = relax.const(1e-12, dtype="float32") | |
| denom = relax.op.sqrt(relax.op.add(sum_squared, eps)) |
| axes = list(range(input_tensor_rank)) | ||
| out = relax.op.strided_slice(in_expr, axes=axes, begin=begin, end=end) |
There was a problem hiding this comment.
The begin and end values read from the TFLite flatbuffer come back as numpy.int32, and strided_slice doesn't accept those. That's why you get the TypeError: Cannot convert 1 with type <class 'numpy.int32'> CI failure. Cast them to plain Python ints before passing:
| axes = list(range(input_tensor_rank)) | |
| out = relax.op.strided_slice(in_expr, axes=axes, begin=begin, end=end) | |
| axes = list(range(len(begin))) | |
| begin = [int(b) for b in begin] | |
| end = [int(e) for e in end] | |
| axes = list(range(input_tensor_rank)) | |
| out = relax.op.strided_slice(in_expr, axes=axes, begin=begin, end=end) |
3 participants
This PR adds Relax TFLite frontend test coverage for:
Part of fixing #18971.