Skip to content

Latest commit

 

History

History
56 lines (36 loc) · 3.19 KB

README.md

File metadata and controls

56 lines (36 loc) · 3.19 KB

Keras Model Downloading

Instructions

Installation of the required libraries is easiest with Anaconda

  • conda create -n tf_downloading python=3.7
  • conda activate tf_downloading
  • conda install tensorflow=1.14.0
  • conda install keras
  • conda install efficientnet

There are two main files, keras_download.py and fix_efficientnet.py

keras_download.py

This file simply loops through all models defined in models.py, downloads the weights, and saves the .h5 file.

Run this script first: (tf_downloading) wekaDeeplearning4j $ python keras_download.py

fix_efficientnet.py

The EfficientNet model as supplied by the Qubvel Implementation has two problems that stop it working in DL4J:

  1. Custom FixedDropout layer type. This is fixed by replacing all FixedDropout layers with a simple Dropout layer (which is supported in DL4J). It's unlikely this will cause any noticable harm to the model.
  2. Mismatching activation shapes (i.e., multiplying [112, 112, 32] with [1, 1, 32]). This isn't a problem in Keras as it implicitly broadcasts the non-matching dimensions (e.g., [1, 1, 32] is broadcasted to [112, 112, 32]), however, due to the way DL4J parses the Multiply layer, it requires the two sets of activations to be of equal shape. This is fixed by inserting a broadcasting Lambda layer to the Keras model.

Fixes within DL4J

Deeplearning4j cannot parse Lambda layers by itself; it needs a matching Java implementation to be specified which it can then use. To load the broadcasting Lambda layers inserted into the EfficientNet model in DL4J, the CustomBroadcast layer needs to be used. This CustomBroadcast layer simply needs to be intialized with the resulting width the activations should be broadcasted to.

Note that this is done by the KerasModelConverter automatically. Once a DL4J .zip file has been created, no Lambda layers need to be specified.

Other Notes

KerasModelConverter uses a feature which isn't currently supported in DL4J - InputType.setDefaultCNN2DFormat(). When parsing the EfficientNet models, some layers are created with the 'default' channel order. Previously this has always been channels first (CNN2DFormat.NCHW), however, the EfficientNet models require the default to be opposite (CNN2DFormat.NHWC). This functionality isn't currently supported in DL4J (as of 1.0.0-beta7), so a custom .jar must be created.

Installing Custom Jar

To fix this, a fork of Deeplearning4J has been created: link

You need to clone this repository, and build a new deeplearning4j-nn jar file.

Then, copy this file into the place of the currently downloaded deeplearning4j-nn jar file (e.g. deeplearning4j-nn-1.0.0-beta7.jar) and rename this new jar to replace the old one.

Your IDE should now load the newly created sources (from the custom jar), and KerasModelConverter should be able to run.

Once Deeplearning4j merges the PR and puts out a new release (after 1.0.0-beta7), this custom jar process won't be necessary; the setDefaultCNN2DFormat() method will already be on the InputType class.