#acumos-meeting: Architecture Committee
Meeting started by farheen_att at 15:12:26 UTC
(full logs).
Meeting summary
- Federated Training (farheen_att, 15:12:55)
- Guy there may be overlap in the docker image
creation. You don't need to rebuild a docker image to use the
training part. (farheen_att,
15:14:56)
- Wenting we don't need to build but we need
storage for version control. (farheen_att,
15:15:11)
- Sayee version control can be done in many ways.
are we going to have 90 versions deployed? (farheen_att,
15:15:34)
- Sayee running model will have a wrapper to
train the model delete the cache. (farheen_att,
15:16:04)
- W: you need to specify where the model is
onboarded. Besides predict you have to have update. (farheen_att,
15:16:27)
- Sayee - if it is a rest API to call a model you
will not have another version correct? (farheen_att,
15:16:50)
- S: do you do a post call to update the
model? (farheen_att,
15:17:06)
- W: it doesn't have to be a REST api it can be
handled by CI/CD. If the API is required then we need to be the API
interface but we don't have that yet. (farheen_att,
15:17:41)
- S: we build the model outside and we pull the
model based on the flag or setting. It will then restart all the
pods to clear the cache. (farheen_att,
15:18:13)
- S: If you are going to train on predict as well
as update. (farheen_att,
15:18:35)
- W: you want to generate a new ms? (farheen_att,
15:18:52)
- S: No you can re-use. (farheen_att,
15:19:03)
- Guy: PMML has a fairly clean separation between
code and data. For a lot of our models the code and data is
intertwined. Data is the working memory of the model. (farheen_att,
15:19:44)
- G: suppose you're training neural net model.
After training you have weights, threshholds, and graphs. That
object is the trained model in order to persist that you have to
serialize that object into working memory. (farheen_att,
15:20:35)
- G: All the packages provide a way to do. We
have to find a way to put them into Acumos. (farheen_att,
15:20:51)
- W: We are not going to re-do the MS we are
going to do a hyper visor change. (farheen_att,
15:21:15)
- G: Changing the model architecture will be the
same. Building a Random forest tree powers a predictor. You just
have to load this data structure. (farheen_att,
15:21:59)
- W: Agree, we will not change the architecture.
We have to have a mechanism to update this set of parameters.
(farheen_att,
15:22:36)
- G: Parameters is tens of thousands of MB should
not be a parameter. (farheen_att,
15:23:13)
- W: We will call it parameter until we decide
what we want to call it. (farheen_att,
15:24:55)
- W: I agree we have to make use of the flow but
we need the basic functionality. (farheen_att,
15:30:10)
- S: We need a flow or thought process.
(farheen_att,
15:30:21)
- W: yes, I agree whether the bi-directional
communication is based on supplier or subscriber. (farheen_att,
15:31:01)
- S: Bi-directional means when I have a model
running then I want a container. (farheen_att,
15:31:31)
- W: info from model will be passed from
subscriber to the supplier. The supplier is the creator of the
model. (farheen_att,
15:32:40)
- Guy: What are the next steps? (farheen_att,
15:33:20)
- W: we can have broader discussion. (farheen_att,
15:33:53)
- Demos (farheen_att, 15:34:13)
- there were 3 demos proposed for today.
(farheen_att,
15:34:27)
- first is from portal and then licensing.
(farheen_att,
15:34:36)
- Rajesh reschedule portal. (farheen_att,
15:34:50)
- We will have the OAM demo instead. (farheen_att,
15:35:15)
- OA&M Demeter Sprint 1 demo (farheen_att, 15:35:36)
- demo being recorded and made available at
wiki.acumos.org. (farheen_att,
15:35:58)
- sprint management. (farheen_att, 15:47:31)
- topic Config issue for Acumos / Azure K8
deployment. (farheen_att,
15:53:15)
- Sayee asking for Ken's support (farheen_att,
15:54:03)
- Ken - We have been working on something similar
on our setup. I will talk to Will later and include you.
(farheen_att,
15:54:27)
- Sayee: That would be helpful. (farheen_att,
15:54:34)
- Model Chaining RTU licenses in DS (farheen_att, 15:55:53)
- Sayee: There are two checks in DS. we need an
API to list the model instead of calling by version, Alex will give
us that RTU API. (farheen_att,
15:56:34)
- Interworking with ONAP Jessica (farheen_att, 15:57:23)
- Jessica wants to showcase the interworking with
ONAP and E2E. How stable is the interworking between Acumos.
(farheen_att,
16:00:30)
Meeting ended at 16:00:48 UTC
(full logs).
Action items
- (none)
People present (lines said)
- farheen_att (46)
- collabot` (3)
Generated by MeetBot 0.1.4.