Key transfer learning insight: when adapating a pre-trained DL model, one must freeze the inner network and change the input & output layers, then train on new data. The fine-tuning is all w/in the _new_ (unfrozen!) I/O layers.

key_transfer_learning_insight.py

# https://github.com/intel-analytics/analytics-zoo#transfer-learning
from zoo.pipeline.api.net import *
from zoo.pipeline.api.keras.layers import *
from zoo.pipeline.api.keras.models import *


def transfer_learn_model(def_path:str, model_path:str) -> Model:
  # load the pre-trained model
  full_model = Net.load_caffe(def_path, model_path)

  # create a new model by removing layers after pool5/drop_7x7_s1
  model = full_model.new_graph(["pool5/drop_7x7_s1"])
  # freeze layers from input to pool4/3x3_s2 inclusive
  model.freeze_up_to(["pool4/3x3_s2"])

  # wrap the frozen model w/ new I/O layers
  inputs = Input(name="input", shape=(3, 224, 224))
  inception = model.to_keras()(inputs)
  flatten = Flatten()(inception)
  logits = Dense(2)(flatten)
  return Model(inputs, logits)

https://github.com/intel-analytics/BigDL

and

https://github.com/intel-analytics/analytics-zo