Inference model

[mdne]

Hi, I'd trained the model on my own dataset and got training_model.h5 file. Then I've tried to load model using keras functional like that: model = tf.keras.models.load_model('training_model.h5', custom_objects={"AdaptiveL1L2": AdaptiveL1L2})

and got an error:

ValueError: Layer #0 (named "resnet" in the current model) was found to correspond to layer resnet in the save file. However the new layer resnet expects 6 weights, but the saved weights have 10 elements.

So I have several questions:

1. How to load the model for inference?
2. How to make an inference on one or more images without custom datagenerator?
3. What is the size of embeddings?

Thanks

[nixingyang]

@mdne Hi, The native load_model function would yield such error due to the reason that the l2_regularization_factor tensors are not initiated. This issue could be addressed by calling test_on_batch before load_weights as shown here. Returning to your questions: 1, Define the model using init_model. Call test_on_batch, followed by load_weights. The aforementioned steps are the same as the evaluation procedure. 2, A minimal snippet would be inference_model.predict_on_batch(...). A more detailed example can be found in the extract_features function. Please note that training_model differs from inference_model. 3, The size of embeddings is 4096. Based on my experiences from other projects, you may shrink it to 256 or 128 using PCA, while causing reasonable performance degradation. Happy experimenting :-) All the best. Xingyang

[mdne]

Thank you for the explanation. Am I right that your method assumes definition and loading the training model and makes an inference inside by calling callback? Is there a way to load the inference model directly from file(training_model.h5) and use it? Or maybe saving a model after training as an inference model?

[nixingyang]

1, Yes, that is how my implementation works. After initiating the models (training_model and inference_model) using init_model, test_on_batch and load_weights successively, you may directly compute the embeddings via inference_model.predict_on_batch(...). It is not mandatory to incorporate the evaluation pipeline as a callback. 2, It is not straightforward to load the inference model directly from the weights file, due to the same reason explained in my previous post. You may write a stand-alone script for inference if that is what you need.

[stale[bot]]

Closing as stale. Please reopen if you'd like to work on this further.