Deeplearning4j: DL4J Keras Import: 2.2.x/2.3.x compatibility

Created on 9 Oct 2018  路  7Comments  路  Source: eclipse/deeplearning4j

Reported by @Kissy: https://github.com/deeplearning4j/deeplearning4j/issues/6527#issuecomment-427737778

I reproduce the exception compiling with Keras 2.2.3 and importing with 1.0.0-beta2.
It seems that the keras field for "config" (retrieve here modelConfig.get(config.getModelFieldConfig())) has changed from list of object in version 2.0.X to object representing map.

Keras 2.0.9 :
{ "backend": "tensorflow", "keras_version": "2.0.9", "class_name": "Sequential", "config": [ { "class_name": "Dense", "config": { "kernel_initializer": { "class_name": "VarianceScaling", "config": { "distribution": "uniform", "scale": 1.0, "seed": null, "mode": "fan_avg" }

Keras 2.2.3 :
{ "backend": "tensorflow", "keras_version": "2.2.3", "class_name": "Sequential", "config": { "name": "sequential_1", "layers": [ { "class_name": "Dense", "config": { "name": "dense_1", "trainable": true, "batch_input_shape": [ null, 100 ],

DL4J Keras

Most helpful comment

fix will be available on snapshots starting tomorrow

All 7 comments

that's disgusting... alright, let me patch this. I hope this doesn't happen again.

I just did a test working backwards from keras 2.2.4 and there is no issue for 2.2.2.

Just an INFO log about Unexpected end-of-input:

[main] INFO org.deeplearning4j.nn.modelimport.keras.Hdf5Archive - Unexpected end-of-input: was expecting closing '"' for name

I could post the two different version h5 files for cifar if you want (see issue #6577).

@timmolter that's fine, we're on top of this issue. thanks for offering help

fix will be available on snapshots starting tomorrow

I'm risking here an amateur question: are snapshots available from maven? 1.0.0-beta2 still produces the above error.

use 1.0.0-SNAPSHOT.

@bugfoot not from maven, but from sonatype.
See this page: https://deeplearning4j.org/docs/latest/deeplearning4j-config-snapshots
In the future, ask these sorts of questions on the gitter channel, not closed issues - you'll get a quicker response there.
https://gitter.im/deeplearning4j/deeplearning4j

Was this page helpful?
0 / 5 - 0 ratings