Stuck on an issue?

Lightrun Answers was designed to reduce the constant googling that comes with debugging 3rd party libraries. It collects links to all the places you might be looking at while hunting down a tough bug.

And, if you’re still stuck at the end, we’re happy to hop on a call to see how we can help out.

Unknown padding parameter: explicit

See original GitHub issue

TensorFlow.js version

Tested with both: 1.7.4 & 2.0.0-rc.4

Browser version

Any.

Describe the problem or feature request

The model.json converted using tensorflowjs_converter 1.7.4r1 contained explicit_paddings in the following lines:

...
      {
        "name": "StatefulPartitionedCall/sequential/conv2d/Relu",
        "op": "_FusedConv2D",
        "input": [
          "conv2d_input",
          "Func/StatefulPartitionedCall/input/_2",
          "Func/StatefulPartitionedCall/input/_3"
        ],
        "device": "/device:CPU:0",
        "attr": {
          "dilations": { "list": { "i": ["1", "1", "1", "1"] } },
          "T": { "type": "DT_FLOAT" },
          "strides": { "list": { "i": ["1", "1", "1", "1"] } },
          "data_format": { "s": "TkhXQw==" },
          "use_cudnn_on_gpu": { "b": true },
          "explicit_paddings": {
            "list": { "i": ["0", "0", "0", "0", "0", "0", "0", "0"] }
          },
          "num_args": { "i": "1" },
          "epsilon": { "f": 0.0 },
          "padding": { "s": "RVhQTElDSVQ=" },
          "fused_ops": { "list": { "s": ["Qmlhc0FkZA==", "UmVsdQ=="] } }
        }
      },
...

I get this when trying to use the model. (2.0.0-rc.4)

    Unknown padding parameter: explicit

      at getPadAndOutInfo (node_modules/@tensorflow/tfjs-core/src/ops/conv_util.ts:434:11)
      at conv_util.computeConv2DInfo (node_modules/@tensorflow/tfjs-core/src/ops/conv_util.ts:149:9)
      at fusedConv2d_ (node_modules/@tensorflow/tfjs-core/src/ops/fused_ops.ts:387:20)
      at fusedConv2d (node_modules/@tensorflow/tfjs-core/src/ops/operation.ts:45:24)
      at convolution.executeOp (node_modules/@tensorflow/tfjs-converter/src/operations/executors/convolution_executor.ts:102:15)
      at node_modules/@tensorflow/tfjs-converter/src/operations/operation_executor.ts:64:23
      at node_modules/@tensorflow/tfjs/node_modules/@tensorflow/tfjs-core/src/engine.ts:425:20
      at Engine.scopedRun (node_modules/@tensorflow/tfjs/node_modules/@tensorflow/tfjs-core/src/engine.ts:436:19)
      at Engine.tidy (node_modules/@tensorflow/tfjs/node_modules/@tensorflow/tfjs-core/src/engine.ts:423:17)
      at Object.tidy (node_modules/@tensorflow/tfjs-core/src/globals.ts:182:17)

Pseudo-code of what I did. The other models without explicit_paddings on the converted json did work without issues.

  const model = await loadGraphModel(
    'http://localhost:5000/model2a/model.json'
  );
  const dTensor = tensor([...d, ...d]);
  const reshape = dTensor.reshape([2, 18, 4, 1]);
  console.log(reshape.shape);
  const prediction = model.predict(reshape);

Code to reproduce the bug / link to feature request

Unfortunately, I can’t share the code at the moment. But Also I have other models that work without problems using the same pattern of approach. What common for those that worked is that they do not have explicit_paddings in their generated model.json.