Merge pull request #67 from yurujaja/regression_fix

Minor regression code and command fixes

README.md

@@ -211,7 +211,7 @@ torchrun --nnodes=1 --nproc_per_node=1 pangaea/run.py \
    encoder=prithvi \
    decoder=reg_upernet \
    preprocessing=reg_default \
-   criterion=cross_entropy \
+   criterion=mse \
    task=regression
 ```
 To use SatlasNet encoder, the `configs/encoder/satlasnet_si.yaml` is required.
@@ -230,7 +230,7 @@ torchrun --nnodes=1 --nproc_per_node=1 pangaea/run.py \
    encoder=prithvi \
    decoder=reg_upernet_mt_ltae \
    preprocessing=reg_default \
-   criterion=cross_entropy \
+   criterion=mse \
    task=regression
 ```
 

configs/dataset/biomassters.yaml

@@ -1,10 +1,10 @@
 _target_: pangaea.datasets.biomassters.BioMassters
 dataset_name: BioMassters
 root_path: ./data/Biomassters
-# download_url: #https://huggingface.co/datasets/ibm-nasa-geospatial/hls_burn_scars/resolve/main/hls_burn_scars.tar.gz?download=true
+download_url: 
 auto_download: False
 img_size: 256
-temporal: 6 #6 (summer month use if multi_temp is 1)
+temp: 6 #6 (select month to use if single temporal (multi_temp : 1))
 multi_temporal: 12
 multi_modal: True
 

pangaea/datasets/biomassters.py

@@ -13,8 +13,8 @@ def read_imgs(multi_temporal, temp , fname, data_dir, img_size):
     imgs_s1, imgs_s2, mask = [], [], []
     if multi_temporal==1:
         month_list = [temp]        
-    else:
-        month_list = list(range(12))
+    else:        
+        month_list = list(range(int(multi_temporal)))
 
     for month in month_list:
 
@@ -47,10 +47,8 @@ def read_imgs(multi_temporal, temp , fname, data_dir, img_size):
 
     imgs_s1 = np.stack(imgs_s1, axis=1)
     imgs_s2 = np.stack(imgs_s2, axis=1)
-
     return imgs_s1, imgs_s2, mask
 
-# @DATASET_REGISTRY.register()
 class BioMassters(GeoFMDataset):
     def __init__(
         self,
@@ -122,7 +120,6 @@ def __init__(
             data_max=data_max,
             download_url=download_url,
             auto_download=auto_download,
-            # temp = temp,
         )
 
         self.root_path = root_path

pangaea/decoders/upernet.py

@@ -426,6 +426,7 @@ def __init__(
             finetune=finetune,
         )
 
+        self.model_name = "Reg_UPerNet"
         if not self.finetune:
             for param in self.encoder.parameters():
                 param.requires_grad = False

pangaea/engine/trainer.py

@@ -82,7 +82,8 @@ def __init__(
         ], f"Invalid precision {precision}, use 'fp32', 'fp16' or 'bfp16'."
         self.enable_mixed_precision = precision != "fp32"
         self.precision = torch.float16 if (precision == "fp16") else torch.bfloat16
-        self.scaler = torch.GradScaler("cuda", enabled=self.enable_mixed_precision)
+        # self.scaler = torch.GradScaler("cuda", enabled=self.enable_mixed_precision)
+        self.scaler = torch.cuda.amp.GradScaler("cuda", enabled=self.enable_mixed_precision)
 
         self.start_epoch = 0