train bdd detection in coco format (#8)

* train bdd detection in coco format * Update coco.py * Gradient clipping
haofengac · Mar 17, 2019 · 1c8c931 · 1c8c931 · vikki1234-cloud · Jan 10, 2021
1 parent 534ec22
commit 1c8c931
Show file tree

Hide file tree

Showing 2 changed files with 14 additions and 1 deletion.
diff --git a/maskrcnn_benchmark/config/paths_catalog.py b/maskrcnn_benchmark/config/paths_catalog.py
@@ -33,6 +33,12 @@ class DatasetCatalog(object):
         "bdd100k_val": (
             "bdd100k/images/100k/val", "bdd100k/labels/bdd100k_labels_images_val.json"
         ),
+        "bdd100k_det_train": (
+            "bdd100k/images/100k/train", "bdd100k/labels/bdd100k_labels_images_det_coco_train.json"
+        ),
+        "bdd100k_det_val": (
+            "bdd100k/images/100k/val", "bdd100k/labels/bdd100k_labels_images_det_coco_val.json"
+        ),
         "kitti_tracking_train": (
             "kitti_tracking/train", ""
         ),
@@ -46,7 +52,7 @@ def get(name):
         factory_names = {
             "coco": "COCODataset",
             "kitti": "KittiDataset",
-            "bdd100k": "Bdd100kDataset"
+            "bdd100k": "COCODataset"
         }
         for k in factory_names.keys():
             if (k in name):

diff --git a/maskrcnn_benchmark/engine/trainer.py b/maskrcnn_benchmark/engine/trainer.py
@@ -74,6 +74,13 @@ def do_train(
 
         optimizer.zero_grad()
         losses.backward()
+
+        accum_grad = 0
+        for p in list(filter(lambda p: p.grad is not None, model.parameters())):
+            accum_grad += p.grad.data.norm(2).item()
+
+        if iteration > 500 and accum_grad > 200:
+            torch.nn.utils.clip_grad_norm_(model.parameters(), 200)
         optimizer.step()
 
         batch_time = time.time() - end