v1.3.0 updates (#149)

* generalize method of accessing likelihood in apps

* generalize subdirectory names used by video app

* move version def from setup.py to __init__

* refactor train_hydra script to import a train function for greater flexibility

Author: themattinthehatt

lightning_pose/__init__.py

@@ -0,0 +1 @@
+__version__ = "1.3.0"

lightning_pose/apps/utils.py

@@ -16,7 +16,7 @@
 
 
 @st.cache_resource
-def update_labeled_file_list(model_preds_folders: list, use_ood: bool = False):
+def update_labeled_file_list(model_preds_folders: List[str], use_ood: bool = False) -> List[list]:
     per_model_preds = []
     for model_pred_folder in model_preds_folders:
         # pull labeled results from each model folder
@@ -40,15 +40,20 @@ def update_labeled_file_list(model_preds_folders: list, use_ood: bool = False):
 
 
 @st.cache_resource
-def update_vid_metric_files_list(video: str, model_preds_folders: list):
+def update_vid_metric_files_list(
+    video: str,
+    model_preds_folders: List[str],
+    video_subdir: str = "video_preds",
+) -> List[list]:
     per_vid_preds = []
     for model_preds_folder in model_preds_folders:
         # pull each prediction file associated with a particular video
         # wrap in Path so that it looks like an UploadedFile object
+        video_dir = os.path.join(model_preds_folder, video_subdir)
+        if not os.path.isdir(video_dir):
+            continue
         model_preds = [
-            f
-            for f in os.listdir(os.path.join(model_preds_folder, "video_preds"))
-            if os.path.isfile(os.path.join(model_preds_folder, "video_preds", f))
+            f for f in os.listdir(video_dir) if os.path.isfile(os.path.join(video_dir, f))
         ]
         ret_files = []
         for file in model_preds:
@@ -59,16 +64,17 @@ def update_vid_metric_files_list(video: str, model_preds_folders: list):
 
 
 @st.cache_resource
-def get_all_videos(model_preds_folders: list):
+def get_all_videos(model_preds_folders: List[str], video_subdir: str = "video_preds") -> list:
     # find each video that is predicted on by the models
     # wrap in Path so that it looks like an UploadedFile object
     # returned by streamlit's file_uploader
     ret_videos = set()
     for model_preds_folder in model_preds_folders:
+        video_dir = os.path.join(model_preds_folder, video_subdir)
+        if not os.path.isdir(video_dir):
+            continue
         model_preds = [
-            f
-            for f in os.listdir(os.path.join(model_preds_folder, "video_preds"))
-            if os.path.isfile(os.path.join(model_preds_folder, "video_preds", f))
+            f for f in os.listdir(video_dir) if os.path.isfile(os.path.join(video_dir, f))
         ]
         for file in model_preds:
             if "temporal" in file:
@@ -97,7 +103,7 @@ def concat_dfs(dframes: Dict[str, pd.DataFrame]) -> Tuple[pd.DataFrame, List[str
 
 
 @st.cache_data
-def get_df_box(df_orig, keypoint_names, model_names):
+def get_df_box(df_orig: pd.DataFrame, keypoint_names: list, model_names: list) -> pd.DataFrame:
     df_boxes = []
     for keypoint in keypoint_names:
         for model_curr in model_names:
@@ -112,7 +118,13 @@ def get_df_box(df_orig, keypoint_names, model_names):
 
 
 @st.cache_data
-def get_df_scatter(df_0, df_1, data_type, model_names, keypoint_names):
+def get_df_scatter(
+    df_0: pd.DataFrame,
+    df_1: pd.DataFrame,
+    data_type: str,
+    model_names: list,
+    keypoint_names: list
+) -> pd.DataFrame:
     df_scatters = []
     for keypoint in keypoint_names:
         df_scatters.append(
@@ -147,7 +159,7 @@ def get_full_name(keypoint: str, coordinate: str, model: str) -> str:
 # ----------------------------------------------
 @st.cache_data
 def build_precomputed_metrics_df(
-    dframes: Dict[str, pd.DataFrame], keypoint_names: List[str], **kwargs
+    dframes: Dict[str, pd.DataFrame], keypoint_names: List[str], **kwargs,
 ) -> dict:
     concat_dfs = defaultdict(list)
     for model_name, df_dict in dframes.items():
@@ -179,7 +191,7 @@ def build_precomputed_metrics_df(
 
 @st.cache_data
 def get_precomputed_error(
-    df: pd.DataFrame, keypoint_names: List[str], model_name: str
+    df: pd.DataFrame, keypoint_names: List[str], model_name: str,
 ) -> pd.DataFrame:
     # collect results
     df_ = df
@@ -192,17 +204,17 @@ def get_precomputed_error(
 
 @st.cache_data
 def compute_confidence(
-    df: pd.DataFrame, keypoint_names: List[str], model_name: str, **kwargs
+    df: pd.DataFrame, keypoint_names: List[str], model_name: str, **kwargs,
 ) -> pd.DataFrame:
+
     if df.shape[1] % 3 == 1:
-        # get rid of "set" column if present
-        tmp = df.iloc[:, :-1].to_numpy().reshape(df.shape[0], -1, 3)
+        # collect "set" column if present
         set = df.iloc[:, -1].to_numpy()
     else:
-        tmp = df.to_numpy().reshape(df.shape[0], -1, 3)
         set = None
 
-    results = tmp[:, :, 2]
+    mask = df.columns.get_level_values("coords").isin(["likelihood"])
+    results = df.loc[:, mask].to_numpy()
 
     # collect results
     df_ = pd.DataFrame(columns=keypoint_names)
@@ -219,7 +231,7 @@ def compute_confidence(
 
 # ------------ utils related to model finding in dir ---------
 # write a function that finds all model folders in the model_dir
-def get_model_folders(model_dir):
+def get_model_folders(model_dir: str) -> List[str]:
     # strip trailing slash if present
     if model_dir[-1] == os.sep:
         model_dir = model_dir[:-1]
@@ -232,7 +244,7 @@ def get_model_folders(model_dir):
 
 
 # just to get the last two levels of the path
-def get_model_folders_vis(model_folders):
+def get_model_folders_vis(model_folders: List[str]) -> List[str]:
     fs = []
     for f in model_folders:
         fs.append(f.split("/")[-2:])

lightning_pose/apps/video_diagnostics.py

@@ -32,6 +32,7 @@
 
 
 def run():
+
     args = parser.parse_args()
 
     st.title("Video Diagnostics")
@@ -53,23 +54,19 @@ def run():
     # get the last two levels of each path to be presented to user
     model_folders_vis = get_model_folders_vis(model_folders)
 
-    selected_models_vis = st.sidebar.multiselect(
-        "Select models", model_folders_vis, default=None
-    )
+    selected_models_vis = st.sidebar.multiselect("Select models", model_folders_vis, default=None)
 
     # append this to full path
-    selected_models = [
-        "/" + os.path.join(args.model_dir, f) for f in selected_models_vis
-    ]
+    selected_models = ["/" + os.path.join(args.model_dir, f) for f in selected_models_vis]
 
     # ----- selecting videos to analyze -----
-    all_videos_: list = get_all_videos(selected_models)
+    all_videos_: list = get_all_videos(selected_models, video_subdir=args.video_subdir)
 
     # choose from the different videos that were predicted
     video_to_plot = st.sidebar.selectbox("Select a video:", [*all_videos_], key="video")
 
     prediction_files = update_vid_metric_files_list(
-        video=video_to_plot, model_preds_folders=selected_models
+        video=video_to_plot, model_preds_folders=selected_models, video_subdir=args.video_subdir,
     )
 
     model_names = copy.copy(selected_models_vis)
@@ -100,9 +97,7 @@ def run():
                     dframe = pd.read_csv(model_pred_file_path, index_col=None)
                     dframes_metrics[model_name][str(model_pred_file)] = dframe
                 else:
-                    dframe = pd.read_csv(
-                        model_pred_file_path, header=[1, 2], index_col=0
-                    )
+                    dframe = pd.read_csv(model_pred_file_path, header=[1, 2], index_col=0)
                     dframes_traces[model_name] = dframe
                     dframes_metrics[model_name]["confidence"] = dframe
                 # data_types = dframe.iloc[:, -1].unique()
@@ -221,6 +216,7 @@ def run():
     parser = argparse.ArgumentParser()
 
     parser.add_argument("--model_dir", type=str, default=[])
+    parser.add_argument("--video_subdir", type=str, default="video_preds")
     parser.add_argument("--make_dir", action="store_true", default=False)
 
     run()