Keras-retinanet: Problem to train densenet and mobilenet models

[ManuelVazquezEnriquez]

When I use any model of resnet or vgg I have no problem, but when I try to test the densenet or mobilenet models the following problems arise:

keras-retinanet/keras_retinanet/bin/train.py --backbone=densenet201 --epochs=1 --steps=1 --no-weights coco /home/manuelvazquezenriquez2/keras-retinanet-test-data/coco

loading annotations into memory...
Done (t=0.00s)
creating index...
index created!
loading annotations into memory...
Done (t=0.00s)
creating index...
index created!
Creating model, this may take a second...
Traceback (most recent call last):
  File "keras-retinanet/keras_retinanet/bin/train.py", line 493, in <module>
    main()
  File "keras-retinanet/keras_retinanet/bin/train.py", line 461, in main
    config=args.config
  File "keras-retinanet/keras_retinanet/bin/train.py", line 118, in create_models
    model          = model_with_weights(backbone_retinanet(num_classes, num_anchors=num_anchors, modifier=modifier), weights=weights, skip_mismatch=True)
  File "keras-retinanet/keras_retinanet/bin/../../keras_retinanet/models/densenet.py", line 35, in retinanet
    return densenet_retinanet(*args, backbone=self.backbone, **kwargs)
  File "keras-retinanet/keras_retinanet/bin/../../keras_retinanet/models/densenet.py", line 85, in densenet_retinanet
    backbone = densenet.DenseNet(blocks=blocks, input_tensor=inputs, include_top=False, pooling=None, weights=None)
  File "/home/manuelvazquezenriquez2/.local/lib/python2.7/site-packages/keras_applications/densenet.py", line 193, in DenseNet
    data_format=backend.image_data_format(),
AttributeError: 'NoneType' object has no attribute 'image_data_format'

keras-retinanet/keras_retinanet/bin/train.py --backbone=mobilenet128 --epochs=1 --steps=1 --no-weights coco /home/manuelvazquezenriquez2/keras-retinanet-test-data/coco

loading annotations into memory...
Done (t=0.00s)
creating index...
index created!
loading annotations into memory...
Done (t=0.00s)
creating index...
index created!
Creating model, this may take a second...
Traceback (most recent call last):
  File "keras-retinanet/keras_retinanet/bin/train.py", line 493, in <module>
    main()
  File "keras-retinanet/keras_retinanet/bin/train.py", line 461, in main
    config=args.config
  File "keras-retinanet/keras_retinanet/bin/train.py", line 118, in create_models
    model          = model_with_weights(backbone_retinanet(num_classes, num_anchors=num_anchors, modifier=modifier), weights=weights, skip_mismatch=True)
  File "keras-retinanet/keras_retinanet/bin/../../keras_retinanet/models/mobilenet.py", line 35, in retinanet
    return mobilenet_retinanet(*args, backbone=self.backbone, **kwargs)
  File "keras-retinanet/keras_retinanet/bin/../../keras_retinanet/models/mobilenet.py", line 92, in mobilenet_retinanet
    alpha = float(backbone.split('_')[1])
IndexError: list index out of range

Any idea why it can be?
Thanks

[hgaiser]

What Keras version are you using? If you're not using 2.2.2, you could try that. I've yet to update support for 2.2.4, I think something broke there.

[hgaiser]

On second look, the error of mobilenet is because you're missing the alpha in the backbone name.

It should be --backbone=mobilenet128_1.0 for an alpha of 1.0.

As for densenet, my original question still stands.

[ManuelVazquezEnriquez]

With the correct version of Keras (v2.2.2) it has worked.
Thank you

[plouvier]

So, if I understand correctly, retinanet does not work yet with keras 2.2.4 ?
Because i need to update my keras to use new parametres for another neural network but i must be sure that retinanet still work with keras 2.2.4

[rcasero]

Same problem here with DenseNet and keras 2.2.4. Trying to run

base_model = DenseNet(blocks=[6, 12, 24, 16], include_top=False, weights=None, input_shape=(401, 401, 3),
                              pooling='avg')

gives the error

Traceback (most recent call last):
  File "<input>", line 7, in <module>
  File "~/.conda/envs/cytometer_tensorflow_latest/lib/python3.6/site-packages/keras_applications/densenet.py", line 193, in DenseNet
    data_format=backend.image_data_format(),
AttributeError: 'NoneType' object has no attribute 'image_data_format'

[hgaiser]

Could you try updating keras-retinanet? I think I fixed that issue not so long ago. https://github.com/fizyr/keras-retinanet/commit/dca2d0d777a3a9daec5b2e98bcf76516eebe8887

[rcasero]

@hgaiser Sorry, I was trying to use just DenseNet, and ended up here by mistake.

from keras_applications.densenet import DenseNet

[richesh09]

Same problem here with DenseNet and keras 2.3.1. Trying to run
model = keras.applications.densenet.densenet.DenseNet(blocks=[6,10,14,12],include_top=False, weights='imagenet', input_tensor=None, input_shape=input_shape, pooling='max', classes=1000,backend=keras.backend,layers=keras.layers)