[Tracking Issue][TFLite] Expand unit test coverage for supported non-quantized operators

[tlopex]

We just rebuilt the TFLite frontend.

Today, tests/python/relax/test_frontend_tflite.py already covers a subset of operators, including some via parametrized tests (for example RELU, RELU6, TANH, SOFTMAX, LOGISTIC, ARG_MIN, and ARG_MAX). However, its test coverage is still not enough.

This issue tracks adding operator-focused unit tests for supported non-quantized operators that are still under-tested or currently have no explicit coverage in the Relax TFLite frontend.

Candidate operators to cover

The list below is based on the current Relax frontend implementation and the current contents of tests/python/relax/test_frontend_tflite.py.

Shape / indexing / layout

• SQUEEZE
• STRIDED_SLICE
• SLICE
• SHAPE
• SELECT / WHERE
• GATHER
• GATHER_ND
• REVERSE_V2
• REVERSE_SEQUENCE
• TILE
• UNPACK
• ZEROS_LIKE
• RANGE
• SPLIT_V
• SPACE_TO_BATCH_ND
• BATCH_TO_SPACE_ND
• DEPTH_TO_SPACE
• SPACE_TO_DEPTH

Elementwise / activation / comparison / normalization

• HARD_SWISH
• LEAKY_RELU
• RELU_N1_TO_1
• PRELU
• LOG
• LOG_SOFTMAX
• SQUARED_DIFFERENCE
• GREATER
• GREATER_EQUAL
• L2_NORMALIZATION
• LOCAL_RESPONSE_NORMALIZATION

Reductions

• MEAN
• REDUCE_MAX
• REDUCE_MIN
• REDUCE_PROD
• SUM

Resize

• RESIZE_BILINEAR
• RESIZE_NEAREST_NEIGHBOR

Convolution / linear / pooling

• FULLY_CONNECTED
• DEPTHWISE_CONV_2D
• TRANSPOSE_CONV
• L2_POOL_2D

Padding / sparse / other

• PAD
• PADV2
• MIRROR_PAD
• ONE_HOT
• MATRIX_DIAG
• MATRIX_SET_DIAG
• TOPK_V2
• SPARSE_TO_DENSE
• DENSIFY (sparse-to-dense tensor; may need a network-level test rather than a
  simple unit test)

Suggested contribution style

• Small PRs are welcome. A PR can add coverage for one operator or a small related group
  of operators.
• Please reuse the existing test style in tests/python/relax/test_frontend_tflite.py.
• For simple conversions, checking the generated Relax IR with tvm.ir.assert_structural_equal is preferred. We really appreciate you provide Expected IRModule to test those ops
• For larger operators, using the existing verify(...) helper pattern is also fine, following current tests in the file.
• Numerical E2E checks should continue to follow the current nightly-only behavior guarded
  by CI_ENV_NIGHTLY.

Useful reference

The old Relay test file is a good source of test shapes, edge cases, and operator-specific scenarios:

• tests/python/frontend/tflite/test_forward.py at commit 292ecfd210

Notes for contributors

This is a good-first-issue / parallelizable tracking issue:

• each operator test is mostly independent
• multiple PRs can land in parallel
• contributors can reference this issue and check off the operator(s) covered by their PR

If you pick this up, please mention which operator(s) your PR adds coverage for.
And if you find some issues with existed ops when adding tests, you're welcome to fix them as well!

cc @leandron

[Dayuxiaoshui]

Hi @tlopex — I’d like to pick up Relax TFLite frontend test coverage for these five non-quantized ops: GREATER, GREATER_EQUAL, LOG, SQUEEZE, and ZEROS_LIKE. I’ll follow the existing style in tests/python/relax/test_frontend_tflite.py (structural verify / Expected IR where reasonable). Please let me know if any of these already have an open PR. Thanks!

[fnhirwa]

Hi @tlopex, I would like to work on the activation/elementwise group. I will add test coverage for HARD_SWISH, LEAKY_RELU, RELU_N1_TO_1, PRELU, LOG, and LOG_SOFTMAX in a single PR. Following the existing assert_structural_equal style with explicit Expected IRModules.

[rknastenka]

Hi @tlopex , i'd like to work on the "Reductions" Group: SUM, MEAN, REDUCE_MAX, REDUCE_MIN, REDUCE_PROD

i'll cover them in one PR

[OmarAzizi]

Hi @tlopex, I want to claim RESIZE_BILINEAR and RESIZE_NEAREST_NEIGHBOR

[0xjah]

hey @tlopex I'd claim L2_NORMALIZATION
SLICE, REVERSE_V2

[rknastenka]

Hi @tlopex, I'd like to claim SHAPE, TILE, and RANGE.
I'll submit them together in a single PR.

[0xjah]

hey @tlopex I'd claim L2_NORMALIZATION
SLICE, REVERSE_V2 and Convolution / linear / pooling I’ll submit the pr soon