fix: add grad spike detection

changelog.md

@@ -10,6 +10,7 @@
 - `docs/tutorials/tuning.md`: New tutorial for hyperparameter tuning.
 - Provided a [detailed tutorial](./docs/tutorials/tuning.md) on hyperparameter tuning, covering usage scenarios and configuration options.
 - `ScheduledOptimizer` (e.g., `@core: "optimizer"`) now supports importing optimizers using their qualified name (e.g., `optim: "torch.optim.Adam"`).
+- Added grad spike detection to the `edsnlp.train` script, and per weight layer gradient logging.
 
 ### Changed
 
@@ -27,6 +28,7 @@
   - Ensure we don't overwrite the RNG of the data reader when calling `stream.shuffle()` with no seed
   - Raise an error if the batch size in `stream.shuffle(batch_size=...)` is not compatible with the stream
 - `eds.split` now keeps doc and span attributes in the sub-documents.
+- Fixed mini-batch accumulation for multi-task training
 
 # v0.15.0 (2024-12-13)
 

docs/tutorials/training.md

@@ -179,7 +179,7 @@ EDS-NLP supports training models either [from the command line](#from-the-comman
       val_data: ${ val_data }
       max_steps: 2000
       validation_interval: ${ train.max_steps//10 }
-      max_grad_norm: 1.0
+      grad_max_norm: 1.0
       scorer: ${ scorer }
       optimizer: ${ optimizer }
       # Do preprocessing in parallel on 1 worker
@@ -284,7 +284,7 @@ EDS-NLP supports training models either [from the command line](#from-the-comman
         val_data=val_data,
         scorer={"ner": ner_metric},
         optimizer=optimizer,
-        max_grad_norm=1.0,
+        grad_max_norm=1.0,
         output_dir="artifacts",
         # Do preprocessing in parallel on 1 worker
         num_workers=1,

docs/tutorials/tuning.md

@@ -233,7 +233,7 @@ train:
   val_data: ${ val_data }
   max_steps: 400
   validation_interval: ${ train.max_steps//2 }
-  max_grad_norm: 1.0
+  grad_max_norm: 1.0
   scorer: ${ scorer }
   optimizer: ${ optimizer }
   num_workers: 2