Custom Layer incompatible dimensions

[weingaunity]

Hi, i try to realize a custom layer (a dense layer with bias and relu activation function only). I implemented the demonstration setup in a single page web-app and end up with following error when i start training:

Uncaught (in promise) Error: Input 0 is incompatible with layer dense_Dense1: expected min_ndim=2, found ndim=1.

Can you give me some advice or resources where different kind of custom layers are implemented in javascript (not typescript).

Here the complete web-app: index.html

Here is the custom layer i created:

class FixedRampLayer extends tf.layers.Layer {
  constructor(config) {
    super(config);
    this.units = config.units;
  }
  computeOutputShape(inputShape) {
    return [inputShape[0], this.units];
  }

  build(inputShape) {
    this.bias = this.addWeight(
      'bias',
      [this.units],
      'float32',
      tf.initializers.randomUniform({minval: -1, maxval: 1})
    );
  }

  call(inputs) {
    return tf.relu(tf.add(inputs, this.bias.read()));
  }

  static get className() {
    return 'FixedRampLayer';
  }
}

Here the model:

const model = tf.sequential();
model.add(new FixedRampLayer({units: 10, inputShape: [1]}));             
model.add(tf.layers.dense({units: 1, useBias:true}));

[avmohansai]

Hi @weingaunity,
Thank you for raising this issue.
As per the error, min_ndim=2, found ndim=1, the Dense layer expects a 2D tensor (Rank 2) because it processes data in batches: [batch_size, input_features]. Even if you are training on a single number, it must be wrapped as a 2D array.
Could you please confirm if this works for you.
Thank you.

[weingaunity]

Hi @avmohansai ,
thanks to your answer. I have no idea how to wrap this exactly without using try/error approach.
I do not see where this wrapping happens in other examples. What i found and had a closer look were:

• https://www.tensorflow.org/js/guide/models_and_layers

• tfjs/tfjs-layers/src/layers/core.ts

  Line 265 in 0fc04d9

  override call(inputs: Tensor|Tensor[], kwargs: Kwargs): Tensor|Tensor[] {

• https://github.com/tensorflow/tfjs-examples/blob/master/custom-layer/custom_layer.js

Can you provide me a code snippet which should work?

Thx.