🎨 Refactor ModelABC to Help Use Default Torch Models
#867
Conversation
Signed-off-by: Shan E Ahmed Raza <[email protected]>
Signed-off-by: Shan E Ahmed Raza <[email protected]>
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## dev-define-engines-abc #867 +/- ##
==========================================================
- Coverage 91.52% 91.43% -0.09%
==========================================================
Files 71 71
Lines 8920 8925 +5
Branches 1184 1184
==========================================================
- Hits 8164 8161 -3
- Misses 742 750 +8
Partials 14 14 ☔ View full report in Codecov by Sentry. 🚨 Try these New Features:
|
| with torch.inference_mode(): | ||
| output = model(img_patches_device) | ||
| # Output should be a single tensor or scalar | ||
| return {"probabilities": output.cpu().numpy()} |
There was a problem hiding this comment.
In the current develop branch, neither CNNModel, nor CNNBackbone returned dictionaries as output of their infer_batch() methods. Also, CNNModel currently returns an array, while CNNBackbone returns a list with the array. It might be fine, just wanted to highlight this.
CNNModel
CNNBackbone
There was a problem hiding this comment.
Thanks. We are aware of this. Our preference is to use torch nn models but to generalise for multi modal output we may need dictionaries. This PR is to check if we can move to generic torch models or we will need a sub class.
…rch-nn-model # Conflicts: # tests/models/test_arch_vanilla.py # tiatoolbox/models/architecture/vanilla.py
…rch-nn-model # Conflicts: # tiatoolbox/models/engine/engine_abc.py
ModelABCto Help Use Default Torch Modelsinfer_batchfromModelABC