Minor changes

- Adds cpu prediction support
- Fixes exception w/ not existing Raganato dataset directory
- Fixes protobuf lib latest version incompatibility protocolbuffers/protobuf#10051

.gitignore

@@ -1,3 +1,8 @@
+data/*
+predictions/*
+
+!.placeholder
+
 # Byte-compiled / optimized / DLL files
 __pycache__/
 *.py[cod]

esc/predict.py

@@ -305,6 +305,7 @@ def parse_args() -> argparse.Namespace:
     parser.add_argument("--prediction-types", type=str, required=True, nargs="+")
     parser.add_argument("--evaluate", action="store_true", default=False)
     # default + not required
+    parser.add_argument("--cpu", action="store_true", default=False)
     parser.add_argument("--device", type=int, default=0)
     parser.add_argument("--tokens-per-batch", type=int, default=4000)
     parser.add_argument("--output-errors", action="store_true", default=False)
@@ -328,7 +329,9 @@ def main() -> None:
     wsd_model = ESCModule.load_from_checkpoint(args.ckpt)
     wsd_model.freeze()
 
-    if args.device >= 0:
+    if args.cpu:
+        wsd_model.to("cpu")
+    elif args.device >= 0:
         wsd_model.to(torch.device(args.device))
 
     tokenizer = get_tokenizer(

esc/utils/commons.py

@@ -27,6 +27,9 @@ def chunks(lst, n):
 
 def list_elems_in_dir(dir_path: str, only_files: bool = False, only_dirs: bool = False) -> List[str]:
 
+    if not isdir(dir_path):
+        return list()
+
     elems_in_dir = [e for e in listdir(dir_path)]
 
     if only_files:

requirements.txt

@@ -3,4 +3,5 @@ pytorch-lightning==0.9.0
 wandb
 nltk==3.4.5
 nlp
-black==21.5b2
+black==21.5b2
+protobuf<=3.20.1