Update the doc on use of SHOULD_FLATTEN_FEATS (facebookresearch#188)

Summary: updating the doc to reflect more clearly what this parameter can be used for and vissl default. thanks to Marco for the feedback Pull Request resolved: facebookresearch#188 Reviewed By: min-xu-ai Differential Revision: D26483874 Pulled By: prigoyal fbshipit-source-id: 47afcb07e3c9dc6aeb74019f11fa365ef59a014d
yangsenwxy · Feb 18, 2021 · 2e194e5 · 2e194e5
1 parent bf6401e
commit 2e194e5
Show file tree

Hide file tree

Showing 2 changed files with 6 additions and 0 deletions.
diff --git a/configs/config/benchmark/README.md b/configs/config/benchmark/README.md
@@ -276,6 +276,8 @@ MODEL:
     ]
 ```
 
+**NOTE**: The config files have option :code:`MODEL.FEATURE_EVAL_SETTINGS.SHOULD_FLATTEN_FEATS` which can be set to True to flatten the extracted features to :code:`NxD` dimensions. By default, VISSL doesn't flatten extracted features and return the features as is.
+
 ### Full finetuning
 
 Since this only requires to initialize the model from the pre-trained model weights, there's

diff --git a/docs/source/evaluations/feature_extraction.rst b/docs/source/evaluations/feature_extraction.rst
@@ -89,6 +89,10 @@ self-supervised model training.
           ["mlp", {"dims": [2048, 2048], "use_relu": True}],
           ["mlp", {"dims": [2048, 128]}],
         ]
+        
+.. note::
+
+    The config files have option :code:`MODEL.FEATURE_EVAL_SETTINGS.SHOULD_FLATTEN_FEATS` which can be set to True to flatten the extracted features to :code:`NxD` dimensions. By default, VISSL doesn't flatten extracted features and return the features as is.
 
 How to extract features
 --------------------------
-Original file line number
+Diff line change
@@ Expand Up / @@ -276,6 +276,8 @@ MODEL: @@
         ]
     ```
+    **NOTE**: The config files have option :code:`MODEL.FEATURE_EVAL_SETTINGS.SHOULD_FLATTEN_FEATS` which can be set to True to flatten the extracted features to :code:`NxD` dimensions. By default, VISSL doesn't flatten extracted features and return the features as is.
     ### Full finetuning
     Since this only requires to initialize the model from the pre-trained model weights, there's
@@ Expand Down @@