This PR covers 1 major feature upgrade and several minor changes:
Feature Upgrade:
Added functions to pre-allocate array space on Tiled Client, and saving single frame on the fly during inference.
Minor Changes:
Now 2 uids will be taken from yaml file: uid_save for model and metric saving, and uid_retrieve for model retrieval.
Added conversion codes to convert str of lists to np array, list, or float tensor when needed.
To Test:
Modify uri and api_key in example yaml files
run make file commands
In order to test full inference saving, select correct uid_retrieve for previous trained models, and comment out mask_tiled_uri, mask_tiled_api_key (or set them to null in yaml).
Note:
It would be worthwhile to try at least 1 full inference to see if all slices have been saved properly, eg. a very small MSDNet. In my test case, I encountered unstable connection to public tiled server due to spiky internet, and apparently there is no mechanism in code for recovery in such situation. Don't know if this would be a concern for the beam time.
As a reference, the error code I got was: httpx.ReadError: [Errno 54] Connection reset by peer
The SMSNet Ensemble methods should still be tested out as this hasn't been well addressed in my last PR due to time concern.
This PR covers 1 major feature upgrade and several minor changes:
Feature Upgrade:
Added functions to pre-allocate array space on Tiled Client, and saving single frame on the fly during inference.
Minor Changes:
uid_save
for model and metric saving, anduid_retrieve
for model retrieval.To Test:
uid_retrieve
for previous trained models, and comment outmask_tiled_uri
,mask_tiled_api_key
(or set them tonull
in yaml).Note:
httpx.ReadError: [Errno 54] Connection reset by peer