0.14.0release here: https://github.com/combust/mleap/releases
Does anyone have any code that involves a workaround to the problem that mleap doesn't support SQLTransformer that they can share ? I've externalized the sql transformation as suggested here :
so that I can serialize my model to mleap but I can't seem to get the same results
when I compare the model I trained with the SparkML SQLTransformer vs doing the sql transformation on the training data before training the model and also on the input data for evaluation using the serialized mleap version.
wrkopen source software available to assist with the http load testing). Initially, I started with loading the same MLeap Pipeline into MLeap 20 times with a different model name. Then using an already constructed Leap Frame, I did a
wrkcommand configured to slam MLeap with 3 concurrent "clients" across 3 threads generating a lot of transform requests. Using a c5.4xlarge instance (16 vCPUs), the latency was roughly: 10.71ms avg, 43ms 99th perc, 17ms stddev, 300ms max. This didn't cut it.... so, bumped up the instance type to c5.9xlarge (36 vCPUs), and I was able to bring the latency across the entire test down to something more expected: 7.8ms avg, 21ms 99th perc, 3.5ms stddev, 49ms max. BETTER! :) In addition, I didn't notice large Java heap memory usage either.... roughly the same throughout all the tests (sawtoothed as expected between 500MB and 2GB). One other interesting comparison is that I did the same testing with MLeap v0.10.1 about a year ago... with one model on v0.10.1, the latency looked like this: 4.36ms avg, 18.15ms 99th perc, 3.36ms stddev, 40.59ms max. On v0.14.0, with one model loaded doing the same load test I got: 2.49ms avg, 10.86ms 99th perc, 2.09ms stdddev, 53.71ms max). A slight performance bump with the newest version! w00t!
Hi, new user of mleap here!
Is there a way to access individual stages of an ml pipeline?
My use case is this: I train a spark pipeline with string indexer & model stages, which is then converted to mleap pipeline.
At runtime I make predictions with this mleap pipeline, which returns for me a prediction for a given index. However I need to transform the index back into a human-readable string.
In spark I could do something like
val indexToString = pipelineModel.stages(0).asInstanceOf[StringIndexerModel].labels.zipWithIndex.map(el => (el._2, el._1)).toMap
I couldn't find a solution in the docs, can someone point me in the right direction?
docker run -e MLEAP_GRPC_PORT=9091 -e server.port=9090 --name=mleap_serving combustml/mleap-serving:0.14.1-SNAPSHOTfor example. we have a MLEAP_HTTP_PORT environment variable, but i realised now that’s ignored by the spring boot service, will raise a PR to fix that sometime soon. but for now, MLEAP_GRPC_PORT and server.port should do the job.