conv3d, maxPool3d and avgPool3d with pad 'same' use the wrong window when the filter is smaller than the stride

Open Beginner friendly
  • Is something wrong?
#8,753 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Assessment

Difficulty
2/5
Estimated time
1-3 hours
Newbie friendliness
86/100
Issue type
Bug
Clarity
Clearly specified
Activity status
Active
Tech stack
javascript, typescript
Domain
backend, machine-learning

Research direction

Start with tfjs-core/src/ops/conv_util.ts, comparing get3DPadAndOutInfo with the 2D getPadAndOutInfo implementation. Reproduce the issue with the provided stride-2, 1×1×1 example and check the 3D convolution and pooling tests. Done means all three ops sample positions 0 and 2 with pad: 'same', while their output shape remains correct.

Written by the indexing model from the issue text.

Description

Please make sure that this is a bug. As per our
GitHub Policy,
we only address code/doc bugs, performance issues, feature requests and
build/installation issues on GitHub. tag:bug_template

System information

  • Have I written custom code (as opposed to using a stock example script provided in TensorFlow.js): yes
  • OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Linux x86_64 (Debian 12)
  • Mobile device (e.g. iPhone 8, Pixel 2, Samsung Galaxy) if the issue happens on mobile device: n/a
  • TensorFlow.js installed from (npm or script link): npm
  • TensorFlow.js version (use command below): @tensorflow/tfjs-core, @tensorflow/tfjs-backend-cpu, @tensorflow/tfjs-backend-wasm 4.22.0 (master 7f5309fe is unchanged), Node.js 26.13.0
  • Browser version: n/a
  • Tensorflow.js Converter Version: n/a

Describe the current behavior

With pad: 'same', the 3D ops (conv3d, maxPool3d, avgPool3d) read the wrong input positions when TensorFlow's SAME padding formula would be negative before clamping in some dimension, which happens whenever the filter is smaller than the stride (for example a 1x1x1 filter with stride 2, as in a strided 1x1x1 projection). The output has the right shape, [1, 1, 1, 2, 1], but is shifted by one element ([2, 4]). The CPU and WASM backends produce the same result.

Describe the expected behavior

Positions 0 and 2 are sampled, giving [1, 3], matching TensorFlow Python and TensorFlow.js 2D ops.

Standalone code to reproduce the issue

const tf = require("@tensorflow/tfjs-core");
require("@tensorflow/tfjs-backend-cpu");
(async () => {
  await tf.setBackend("cpu");
  const x = tf.tensor5d([1, 2, 3, 4], [1, 1, 1, 4, 1]);
  tf.maxPool3d(x, [1, 1, 1], [1, 1, 2], "same").print();
  tf.avgPool3d(x, [1, 1, 1], [1, 1, 2], "same").print();
  tf.conv3d(x, tf.ones([1, 1, 1, 1, 1]), [1, 1, 2], "same").print();
})();

All three print:

Tensor
    [[[[[2],
        [4]]]]]

TensorFlow (tf.nn.max_pool3d, tf.nn.avg_pool3d, tf.nn.conv3d with "SAME") and the 2D equivalent tf.maxPool(tf.tensor4d([1, 2, 3, 4], [1, 1, 4, 1]), 1, 2, "same") print [1, 3].

Other info / logs

get3DPadAndOutInfo in tfjs-core/src/ops/conv_util.ts computes padAlongDepth, padAlongHeight, and padAlongWidth as (out - 1) * stride + filter - in without Math.max(0, ...), whereas the 2D version getPadAndOutInfo clamps each to at least 0. Here the width padding is -1, so left becomes -1 and every window starts one element late.

Credit: https://ucsbplse.github.io/

Dominant language
TypeScript
Stars
19.2k
Forks
2k
No merged PRs in 30d

Getting set up

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from tensorflow/tfjs

All issues in tensorflow/tfjs

Similar issues

More TypeScript issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.