Unknown layer: AnchorBoxes about ssd_keras HOT 10 CLOSED

@gloddream There definitely is, i just haven't figured out how yet. i've recently implemented the get_config() method in both of the custom layers (L2Normalization and AnchorBoxes), so you should theoretically be able to load the model using

model = load_model('./model/ssd7_0.h5', custom_objects={'AnchorBoxes': AnchorBoxes, 'L2Normalization': L2Normalization, 'SSDLoss': SSDLoss})

but for some reason either model.save or load_model isn't able to use the dictionaries that get_config() provides correctly. Since Keras documentation regarding saving and loading models with custom objects is basically non-existent and as far as I can tell I implemented the relevant methods in my layers exactly like the Keras core layers implement them, I didn't bother trying around any longer since I had other stuff to do. If you figure out how to make it work, let me know 🙂

from ssd_keras.

Instead of loading the model, build the model (by calling build_model()) and then do model.load_weights() instead.

from ssd_keras.

is there some way to use load_model('./model/ssd7_0.h5') ?

from ssd_keras.

In case it's still relevant for you guys, using load_model() works now.

from ssd_keras.

Hi @pierluigiferrari ,
According to Official Keras Documentation here, now there is support for saving model's architecture + weights and optimizer state in a single HDF5 file.
So, is it possible to load full model into keras, without passing the custom_objects parameters.
This way any custom trained model can be loaded to run predictions, without knowing custom objects to run predictions.
I did try to save one of you models into a single HDF5 file as mentioned in the documentation, but while loading I get an error ValueError: Unknown layer: L2Normalization.

I understand this can be done by passing custom_objects, but is there way to load a model without doing that.

from ssd_keras.

@vinay0410 can you point me to that documentation?

from ssd_keras.

Hi @pierluigiferrari ,
Sorry I forgot to attach the link, I have updated my comment with the link.
https://keras.io/getting-started/faq/#how-can-i-save-a-keras-model

from ssd_keras.

@vinay0410 I'm afraid you're mixing things up. Saving a model saves its architecture, training configuration, etc., but you still need to provide any custom objects that the model requires when you load it. What you describe above is not possible and neither does the documentation that you link to suggest it is.

from ssd_keras.

Hi @pierluigiferrari ! I am so thankful for finding your repository! Great work!!

I know this issue is resolved. But I was trying to apply quantization(tensorflow model optimisation technique) to your repository and again encountered the AnchorBoxes issue. As an author of this repository, I thought you would be the best person to check the issue with. Any inputs from your end will be of great help.

I have detailed this on tensorflow model optimization issues page. Link here

from ssd_keras.

Hi @pierluigiferrari ! I am so thankful for finding your repository! Great work!!

I know this issue is resolved. But I was trying to apply quantization(tensorflow model optimisation technique) to your repository and again encountered the AnchorBoxes issue. As an author of this repository, I thought you would be the best person to check the issue with. Any inputs from your end will be of great help.

I have detailed this on tensorflow model optimization issues page. Link here

The above mentioned problem is now resolved under tensorflow/model-optimization#620.

All thanks to @Hackerman28 !! 😄

from ssd_keras.