ValueError: No such layer: P3

One issue when I convert the train model to inference model:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-7-8c208d1cf44b> in <module>()
----> 1 inference_model = retinanet_bbox(model=model)

~/RCNN/keras-retinanet/keras_retinanet/models/retinanet.py in retinanet_bbox(model, anchor_parameters, nms, class_specific_filter, name, **kwargs)
    335 
    336     # compute the anchors
--> 337     features = [model.get_layer(p_name).output for p_name in ['P3', 'P4', 'P5', 'P6', 'P7']]
    338     anchors  = __build_anchors(anchor_parameters, features)
    339 

~/RCNN/keras-retinanet/keras_retinanet/models/retinanet.py in <listcomp>(.0)
    335 
    336     # compute the anchors
--> 337     features = [model.get_layer(p_name).output for p_name in ['P3', 'P4', 'P5', 'P6', 'P7']]
    338     anchors  = __build_anchors(anchor_parameters, features)
    339 

~/.conda/envs/21/lib/python3.6/site-packages/keras/engine/network.py in get_layer(self, name, index)
    363                 return layer
    364 
--> 365         raise ValueError('No such layer: ' + name)
    366 
    367     @property

ValueError: No such layer: P3

However, in retinanet.py __create_pyramid_features(C3, C4, C5, feature_size=256), we can find the codes already included the name='P3' :

...
    # add P4 elementwise to C3
    P3 = keras.layers.Conv2D(feature_size, kernel_size=1, strides=1, padding='same', name='C3_reduced')(C3)
    P3 = keras.layers.Add(name='P3_merged')([P4_upsampled, P3])
    P3 = keras.layers.Conv2D(feature_size, kernel_size=3, strides=1, padding='same', name='P3')(P3)
...

Does anybody know what is going wrong? Thanks in advance!

[hgaiser]

I’m not sure if the multi-gpu model can easily be converted back to a non-multi-gpu model. What did you change in train.py? By default, model should be saved, which is the non-multi-gpu model even if you enable multi-gpu.

[15241192958]

I trained again and saved the inference model. The inference model is working just fine, though I don’t know if it is multi-gpu model or not. Thanks!

excuse me ? can you tell me how to save the inference model?