No easy way to change the input shapes for a built model for transfer learning

TensorFlow.js version

0.11.7

Browser version

N/A

Describe the problem or feature request

If I load a pretrained TFJS model, there is no easy way to change the input shapes it takes in. For instance, if I load Mobilenet, and I want to perform transfer learning on it, I have no easy way to get the model to take different shapes (even if the model is fully convolutional).

Code to reproduce the bug / link to feature request

  const model = await tf.loadModel('https://storage.googleapis.com/tfjs-models/tfjs/mobilenet_v1_0.25_224/model.json')
  
  const test = tf.zeros([1, 512, 512, 3])
  const preds = model.predict(test) 

Expected output:

preds is a tensor of the output of the model

Actual output:

Error: Error when checking : expected input_1 to have shape [null,224,224,3] but got array with shape [1,512,512,3].

This is especially needed when you need to transfer some layers of a loaded model into another:

const someLayers = tf.sequential({layers: model.layers.slice(0, 5)})
someLayers.predict(test)

It would be good if there was some way to unbuild a layer, or if there was some easy way to clone a layer (keeping everything but the shapes & topology), potentially like:

const convLayers = model.layers.slice(0, 5).map(l => l.unbuilt()) // or l.clone(), l.clone({withoutShape: true}), etc

const someLayers = tf.sequential({layers: convLayers})

WDYT? I would be happy to author a PR for it unless someone else is working on something similar.

[aman-tiwari]

This works for now:

type ShapeMap = { [name: string]: [number, number, number, number] };

function mapInputShapes<T extends tf.Model>(model: T, newShapeMap: ShapeMap) {
  const cfg = {...model.getConfig()};
  cfg.layers = (cfg.layers as any[]).map(l => {
    if(l.name in newShapeMap) {
      return {...l, config: {
        ...l.config,
        batchInputShape: newShapeMap[l.name]
      }};
    } else {
      return l;
    }
  });

  const map = tf.serialization.SerializationMap.getMap().classNameMap;
  const [cls, fromConfig] = map[model.getClassName()];

  return fromConfig(cls, cfg) as T;
}

But not sure if tf.serialization.SerializationMap.getMap().classNameMap is public or whether this will break for certain kinds of models.

[schematical]

I used @aman-tiwari’s solution above but now getting this:

Input 0 is incompatible with layer dense_Dense1: expected axis -1 of input shape to have value 25088 but got shape 51,38400.

Any ideas?