Closed
Milestone
DAN-P5: Improve training
See https://redmine.teklia.com/issues/4401
Will start when https://gitlab.com/teklia/atr/dan/-/milestones/3#tab-issues is done
Unstarted Issues (open and unassigned)
0
Ongoing Issues (open and assigned)
0
Completed Issues (closed)
20
- Make normalization optional
- Replace DPI Adjusting by albumentation's RandomScale
- Update labels.json parsing during training
- Fix img_format argument in data extraction
- Is curriculum still used in training
- Remove OCRManager
- Move parse_tokens
- Reorganize parameters in training configuration
- Reorganize repository
- Remove all mentions of SIMARA
- Create train, evaluate and predict commands
- Rename predict and evaluation functions
- Always use DropoutScheduler
- Simplify MetricManager
- Remove curriculum learning
- Allow specifying device to use when training
- Remove never changed training parameters
- Generate the correct `parameters.yml` file directly during training
- Document training on multiple GPUs at once
- Correctly support "DistributedDataParallel"
Loading
Loading
Loading