[Bug] Missing validation for sequence_stride in stft leads to negative stride acceptance and backend crashes

[amadhan882]

Description

The keras.ops.stft operation lacks proper input validation for the sequence_stride parameter. This leads to two specific issues across all backends:

1. Negative Stride Failure: A negative sequence_stride causes the operation to return an unexpected object type (tuple) instead of a tensor, leading to 'tuple' object has no attribute 'shape' errors during downstream processing. This indicates a silent internal failure.
2. Zero Stride Crash: A zero sequence_stride leaks to the backend and causes a raw InvalidArgumentError (Integer division by zero) instead of a unified Keras ValueError.

Minimal Reproducible Example

import os
import numpy as np
import keras
import keras.ops as ops

# Set backend before running
# os.environ["KERAS_BACKEND"] = "tensorflow"

def reproduce_stft_bug():
    print(f"Testing Backend: {keras.backend.backend()}")
    x = ops.convert_to_tensor(np.random.rand(1, 1024).astype("float32"))

    print("\n--- Test 1: Negative Stride ---")
    try:
        # Expected: ValueError | Observed: Returns a tuple, causing AttributeError on .shape
        res_neg = ops.stft(x, sequence_length=256, sequence_stride=-5, fft_length=256)
        print(f"Result shape: {res_neg.shape}")
    except Exception as e:
        print(f"Caught negative stride issue: {e}")

    print("\n--- Test 2: Zero Stride ---")
    try:
        # Expected: ValueError | Observed: Raw Backend Crash
        res_zero = ops.stft(x, sequence_length=256, sequence_stride=0, fft_length=256)
    except Exception as e:
        print(f"Observed: Caught by backend with {type(e).__name__}: {e}")

if __name__ == "__main__":
    reproduce_stft_bug()

Observed Output

Testing Backend: tensorflow

--- Test 1: Negative Stride ---
SUCCESS: Caught negative stride: 'tuple' object has no attribute 'shape'

--- Test 2: Zero Stride ---
OBSERVED: Caught by backend with InvalidArgumentError: {{function_node __wrapped__FloorDiv_device_/job:localhost/replica:0/task:0/device:GPU:0}} Integer division by zero [Op:FloorDiv] name: 
NOTE: This should be caught by Keras validation as a ValueError, not a raw backend error.

Observed Behavior

• Negative Stride: All backends proceed without validating the stride. In TensorFlow, the operation fails internally and returns a tuple, which breaks any subsequent tensor operations.
• Zero Stride: Triggers InvalidArgumentError: Integer division by zero [Op:FloorDiv].

Expected Behavior

keras.ops.stft should validate that sequence_stride > 0 and raise a ValueError consistently across all backends.

System Information (Google Colab)

==============================
      System Info
==============================
OS              : Linux 6.6.105+
Python version  : 3.12.12
Keras version   : 3.10.0
Keras Backend   : tensorflow

==============================
      GPU Info
==============================
GPU Model       : Tesla T4
CUDA Available  : Yes
CUDA Version    : 12.8

==============================
    Library Versions
==============================
jax                                      0.7.2
jax-cuda12-pjrt                          0.7.2
jax-cuda12-plugin                        0.7.2
jaxlib                                   0.7.2
numpy                                    2.0.2
tensorflow                               2.19.0
tensorflow-datasets                      4.9.9
tensorflow_decision_forests              1.12.0
tensorflow-hub                           0.16.1
tensorflow-metadata                      1.17.3
tensorflow-probability                   0.25.0
tensorflow-text                          2.19.0
torch                                    2.9.0+cu128
torchao                                  0.10.0
torchaudio                               2.9.0+cu128
torchcodec                               0.8.0+cu128
torchdata                                0.11.0
torchsummary                             1.5.1
torchtune                                0.6.1
torchvision                              0.24.0+cu128

Conclusion

The current implementation of keras.ops.stft relies on backend-specific behavior for input validation instead of a unified Keras validation layer. This results in:

• Silent Logic Errors: Negative strides are accepted, causing internal failures and returning inconsistent object types (tuples).
• Process Crashes: Zero strides leak to the backend kernels, triggering raw InvalidArgumentError or C++ level division-by-zero instead of a high-level ValueError.

A centralized validation check ensuring sequence_stride > 0 is necessary to maintain cross-backend consistency and framework stability.

[dhantule]

Hi @amadhan882, thanks for reporting this.

I tested your code with Keras 3.13.2 and I was able to reproduce the errors in this gist.

[amadhan882]

Hi @dhantule,

Thanks for confirming the issue and sharing the reproduction gist. It’s great to see that it’s verified on Keras 3.13.2.

Also, thank you @maitry63 for highlighting the issue early on, and @Abhinavexists for the comprehensive PR (#22204) that covers both stft and istft. Looking forward to the fix being merged!