在Watson机器学习上部署使用自定义损失功能训练的模型

Question

I've trained a model on Watson Machine learning with a custom loss function but when I try to deploy it I get an error saying:

Invalid Input data: Unknow loss function:loss

I am aware that normally in Keras you can load a model that have custom loss function using:

load_model('model.h5', custom_object={'loss': loss_function}) 

I am now looking for a way to also add the definition of my loss function in my deployment.

Answer 1

You should be able to follow the method described in this doc for Custom Compnents in TF Models for WML

https://dataplatform.cloud.ibm.com/docs/content/analyze-data/ml-custom_libs_tensorflow.html?audience=wdp&context=analytics

the specific requirements are detailed here: https://dataplatform.cloud.ibm.com/docs/content/analyze-data/ml-custom_libs_overview.html#requirements

The main idea is that you need to specify a runtime for your deployment other thank the default runtime and package up your custom component in the manner specified. Hope this helps.

Answer 2

Reply from official support:

Custom functions and layers for Keras aren't currently supported in WML. You can refer to what is supported in this dock: https://dataplatform.cloud.ibm.com/docs/content/analyze-data/pm_service_supported_frameworks.html?audience=wdp&context=analytics It seems the part about loss functions is not mentioned here. We will update the documentation.

Date: 2019-01-28