tflite export and preprocess_input/decode_predictions

[avber]

Hi @leondgarse ,

If the intention is to export model to tflite and use it in tfjs, I guess preprocess_input/decode_predictions will not be available.

Is it possible to export them as signature functions? Ideally decode_predictions should have at least max_output_size, iou_or_sigma and score_threshold parameters.

Thanks

[leondgarse]

For PreprocessInput, currently will not make it a layer.

• Firstly, the main process for PreprocessInput is resize and rescale. For resize, it's not helpful for TFLite model, I'm not sure if the dynamic input shape is support by TFLite:

  # Define a simple model with dynamic input shape
  inputs = keras.layers.Input([None, None, 3])
  nn = tf.image.resize(inputs, [512, 512])
  bb = keras.models.Model(inputs, nn)
  
  # Convert TFLite
  converter = tf.lite.TFLiteConverter.from_keras_model(bb)
  open(bb.name + ".tflite", "wb").write(converter.convert())
  
  # Check input shape, it is not dynamic anymore...
  tf.lite.Interpreter(model_path=bb.name + ".tflite").get_input_details()[0]["shape"]
  # array([1, 1, 1, 3], dtype=int32)

• For rescale, it can be easily added to model. A case can be found in colab kecam_test.ipynb.

  import tensorflow as tf
  from tensorflow import keras
  from keras_cv_attention_models import efficientdet
  
  model = efficientdet.EfficientDetD0(pretrained="coco")
  inputs = model.inputs[0]
  # rescale_mode torch
  mean = tf.constant([0.485, 0.456, 0.406]) * 255.0
  std = tf.constant([0.229, 0.224, 0.225]) * 255.0
  nn = (inputs - mean) / std
  nn = model(nn)
  model.decode_predictions.use_static_output = True
  nn = model.decode_predictions(nn)
  bb = keras.models.Model(inputs, nn)
  bb.summary()
  _________________________________________________________________
   Layer (type)                Output Shape              Param #   
  =================================================================
   input_3 (InputLayer)        [(None, 512, 512, 3)]     0         
                                                                   
   tf.math.subtract_1 (TFOpLam  (None, 512, 512, 3)      0         
   bda)                                                            
                                                                   
   tf.math.truediv_1 (TFOpLamb  (None, 512, 512, 3)      0         
   da)                                                             
                                                                   
   efficientdet_d0 (Functional  (None, 49104, 94)        3927203   
   )                                                               
                                                                   
   decode_predictions_2 (Decod  (None, 100, 6)           0         
   ePredictions)                                                   
                                                                   
  =================================================================
  Total params: 3,927,203
  Trainable params: 3,880,067
  Non-trainable params: 47,136
  _________________________________________________________________

• Anyway, both resize and rescale are rather easy processes, that can be implemented by other process before calling model.

[avber]

Thank you, will try it in my app.

Definitely agree on resize in the preprocessing stage.