Qinlong Wang

>  > > This one looks a connected graph of processes in a distributed job. Vote this logo. The point in the logo represents a node in a job....

Yes, we can import from elasticdl_preprocessing in `census_wide_deep_model`.

Can we write the those analysis results into a file? Then we read those results to build the model in the container using the same image.

Make sure that the preprocessing module `elasticdl_preprocessing` can be pip installed.

Is `elasticdl/elasticdl:v1.0.0` the image of elasticdl-test-mnist-allreduce-master? If yes, you can run `pip show elasticdl` in a container with the image to check the elasticdl package.

你可以先用 `tf.saved_model.save` 将 Keras Model 保存成`tf.savedmodel`格式. 然后使用 TensorFlow 的 [Java API](https://www.tensorflow.org/install/lang_java?hl=zh-cn) 中的 [SavedModelBundle](https://www.tensorflow.org/api_docs/java/org/tensorflow/SavedModelBundle?hl=zh-cn) 加载模型，并使用 Java API 的 [Session.Runner](https://www.tensorflow.org/api_docs/java/org/tensorflow/Session.Runner?hl=zh-cn) 来预测样本做预估。

这种情况下，应该使用 TensorFlow 的 dataset pipeline，常用的接口就是 `tf.data.Dataset.from_generator` 来构造dataset。PyTroch 也有相应的 dataloader 接口。TensorFlow 的例子可以参考我们针对 k8s 分布式训练的例子的 `input_fn` https://github.com/sql-machine-learning/elasticdl/blob/develop/model_zoo/iris/dnn_estimator.py#L101-L110

I solved it by using `DeepSpeedEngine.backward(loss)` and `DeepSpeedEngine.step()` not torch native`loss.backward()` and `optimizer.step()`.

Now, we are developing features to support running `torch_npu` on Ascend 910 processor. If you have any ideas and solutions to support Mindspore, feel free to submit PRs.

Yeah, we are writing the guide and designing the APIs to use DLRover with new AI chips. We can discuss the detail of API design together if you are interested...