vanvalenlab / kiosk-redis-consumer / 698

96%

push

travis-ci-com

Use GetModelMetadata and batching to pad/tile images for specific model sizes. (#90)

* update protos for 1.15, add model metadata request.

* add function for `get_model_metadata` to `grpc_client` and `TensorFlowServingConsumer`.

* convert `detect_scale` and `detect_label` to use model metadata, and migrate to `ImageFileConsumer`.

* update every model request to query the model_metadata server for model size and dtype.

* deprecate `process_big_image` for `predict`, which handles too big and too small

* remove `cuts`, `field`, and `streaming` from `ImageConsumer`.

* migrate retry login into PredictClient and use context to auto-close the channel.

* cast batch size as an int and only expand image if no batches exist.

* add GatewayTimeout to list of retry-able storage exceptions.

160 of 160 new or added lines in 4 files covered. (100.0%)

1125 of 1239 relevant lines covered (90.8%)

3.63 hits per line