Freeze layers in Tensorflow Object Detection API

Research & Models
, ,
#1

Is it possible to freeze specific layers when using Tensorflow Object Detection API?

For example, I am using EfficientDet downloaded from the Tensorflow 2 Model Zoo. When I train the model, I am going to make it predict whether an object is car,plane or motorcycle. The model is already trained on these types of objects (COCO 2017 dataset), but I want to train it more on my specific use case . But, since it is already trained on these types of object I do not want it to “forget” what it has already learned. Hence, I need to freeze some of the layers. So, is this possible? And, if it is possible, how do I know which layers I actually need to freeze? I have found that I might can use the freeze_variables parameter in the pipeline.config file?

Thanks for any help!

#4

This might be what you’re after
Transfer Learning example
Specifically these lines:

base_model.trainable = True
# Let's take a look to see how many layers are in the base model
print("Number of layers in the base model: ", len(base_model.layers))

# Fine-tune from this layer onwards
fine_tune_at = 100

# Freeze all the layers before the `fine_tune_at` layer
for layer in base_model.layers[:fine_tune_at]:
    layer.trainable =  False
#5

The problem with this is that it is not a Keras model. So as far as I can see it cannot be done like this when models are from the Model Zoo.

#7

I dont think that model will forget these classes if you have them in your training data.

#9

After I trained it on some data the accuracy of the model dropped about 20% for the same classes that I want to make prediction on. I would expect it to be around 94% which was the case before training. So something is going on.

#11

If your model are from the Model Zoo, you also can use Transfer Learning, i.e. to continue train process, but with very small learning_rate, such as 10-7 or 10-8 …based on your train set. In this case, you will improve an accuracy of model, but by a little steps.

#12

So if I reduce the learning rate it should work to train it on my own data without losing all its “progress”?

#13

perhaps yes, it should be conceived as the classical approach