feature: estimates metrics for unavailable states

- during inference phase, the metrics are read from the offline
dataset by querying the current state. When there is no record
corresponding to the state, then the metrics are estimated by using
the twin model.

agent/backend/policy.py

@@ -19,7 +19,7 @@ $$ \max_{state} reward(model(state)) $$
         self.label = """PolicyArgMax: pursue max reward """
 
     def choose(self, model, ds, input_ranges, cur_state, cur_metrics, reward_fn):
-        print(input_ranges)
+        #print(input_ranges)
 
         input_df, output_df = predict_dict(model, ds, input_ranges)
 
@@ -28,7 +28,7 @@ $$ \max_{state} reward(model(state)) $$
         io_df['reward'] = io_df.apply(lambda row: reward_fn.calculate(row), axis=1)
         max_reward_index = io_df['reward'].argmax()
         next_state = io_df.loc[max_reward_index].to_dict()
-        print('next state', next_state)
+        #print('next state', next_state)
         return next_state
 
 

agent/backend/utils.py

@@ -46,10 +46,11 @@ def read_metrics(df, cur_state):
     cols = list(df.columns)
 
     dff = df[cols]
-    #for col, value in cur_state.items():
-    #    print(f'{col} = {value}', dff.columns)
-    #    dff = dff.query(f'{col} = {value}')
-    dff = dff.query('replica == 1')
+    for col, value in cur_state.items():
+        #print(f'{col} = {value}', dff.columns)
+        dff = dff.query(f'{col} == {value}')
+    if len(dff) == 0:
+        return None
     output_df = dff.sample(1)
     for feature in ['replica','cpu','expected_tps']:
         if feature in cols:

agent/dashboard/inference.py

@@ -44,7 +44,7 @@ def force_render():
 @solara.component
 def InferencePlots(render_count):
 
-
+    unavailable_states_in_data, set_unavailable_states_in_data = solara.use_state_or_update(0)
 
     def execute():
         #print(selected_policy_label, selected_reward_label, selected_load_label)
@@ -62,7 +62,7 @@ def InferencePlots(render_count):
 
 
         df = ds.df
-        print(df.columns)
+        #print(df.columns)
         # get all possible values for inputs
         input_ranges = {col: list(pd.unique(df[col])) for col in input_cols}
         
@@ -72,6 +72,7 @@ def InferencePlots(render_count):
         cur_hist = {}
         replica = initial_replica.value
         cpu = initial_cpu.value
+        state_not_found_in_data = 0
         for load, eod in load_profile:
             if step > nsteps.value:
                 break
@@ -83,9 +84,13 @@ def InferencePlots(render_count):
             
             if use_model_to_estimate_metrics.value:
                 cur_metrics = estimate_metrics(model, ds, cur_state)
-                print(cur_metrics)
+                #print(cur_metrics)
             else:
                 cur_metrics = read_metrics(df, cur_state)
+                if cur_metrics is None:
+                    print('there is no data for this state',cur_state)
+                    state_not_found_in_data += 1
+                    cur_metrics = estimate_metrics(model, ds, cur_state)
 
             combined_data = cur_state | cur_metrics 
             for state, value in combined_data.items():
@@ -112,6 +117,7 @@ def InferencePlots(render_count):
                 cpu = next_state['cpu']
             
             step += 1
+            set_unavailable_states_in_data(state_not_found_in_data)
 
 
     with solara.Row():
@@ -130,7 +136,8 @@ def InferencePlots(render_count):
         solara.Warning("There is no trained model ye, please train one!")
     
     solara.Button(label="Execute", on_click=execute, disabled=model is None)
-
+    if unavailable_states_in_data > 0:
+        solara.Warning(f'There are {unavailable_states_in_data} unavailable states in data. Estimatated versions are used!')
 
     #print('Interence plots')
     with solara.ColumnsResponsive():