merge with readme changes

README.md

@@ -32,7 +32,7 @@ Their toolkit also reports higher BM25 and hybrid scores.
 4. Dense retriever component for inference time logic is based on FAISS index.
 
 ## New (March 2021) release
-DPR codeabse is upgraded with a number of enhancements and new models.
+DPR codebase is upgraded with a number of enhancements and new models.
 Major changes:
 1. [Hydra](https://hydra.cc/)-based configuration for all the command line tools exept the data loader (to be converted soon)
 2. Pluggable data processing layer to support custom datasets

dpr/data/download_data.py

@@ -460,6 +460,7 @@ def download(resource_key: str, out_dir: str = None):
     if resource_key not in RESOURCES_MAP:
         # match by prefix
         resources = [k for k in RESOURCES_MAP.keys() if k.startswith(resource_key)]
+        logger.info("matched by prefix resources: %s", resources)
         if resources:
             for key in resources:
                 download(key, out_dir)
@@ -517,9 +518,9 @@ def main():
     if args.resource:
         download(args.resource, args.output_dir)
     else:
-        print("Please specify resource value. Possible options are:")
+        logger.warning("Please specify resource value. Possible options are:")
         for k, v in RESOURCES_MAP.items():
-            print("Resource key=%s  :  %s", k, v["desc"])
+            logger.warning("Resource key=%s  :  %s", k, v["desc"])
 
 
 if __name__ == "__main__":

train_dense_encoder.py

@@ -740,7 +740,7 @@ def _do_biencoder_fwd_pass(
     if cfg.n_gpu > 1:
         loss = loss.mean()
     if cfg.train.gradient_accumulation_steps > 1:
-        loss = loss / cfg.gradient_accumulation_steps
+        loss = loss / cfg.train.gradient_accumulation_steps
     return loss, is_correct