Passing SparseVector to model predictor not supported

[kapilkd13]

ISSUE

I trained Factorization machine model on sagemaker with MXNET. Now for training I followed this notebook . My train data is sparse(300 M rows, 4M columns) I used smac.write_spmatrix_to_sparse_tensor to write my sparse training data to s3 for training. I was able to train and deploy an endpoint for inference. But problem starts when I try to inference, my feature dimension is 4 Million. So when I try to pass a dense vector of this size to predict for inference, I get Request Entity Too Large. On checking I found that sagemaker has 5mb size limit. only option I can think of is to somehow pass the sparse vector but predict is not accepting that. Can you guys help?
I tried sending sparse tensor, with content-type as protobuf, didn't work.
I am using RealTimePredictor class as predictor_cls, is there anything I can do with this class so that it accept sparse vector and maybe convert it on server side if needed. Any Suggestions?

[yastasho]

The algorithm supports JSON format for sparse data as described in https://docs.aws.amazon.com/sagemaker/latest/dg/cdf-inference.html#common-in-formats. Sparse data can also be provided in protobuf format. write_spmatrix_to_sparse_tensor can be used to convert data https://github.com/aws/sagemaker-python-sdk/blob/master/src/sagemaker/amazon/common.py#L172