Compute losses in try/except

app.py

@@ -56,6 +56,8 @@ def main():
     validator_df = state_vars["validator_df"]
     benchmarks = state_vars.get("benchmarks", None)
     benchmark_timestamp = state_vars.get("benchmark_timestamp", None)
+    losses_1 = state_vars["losses_1"]
+    losses_2 = state_vars["losses_2"]
 
     demo = gr.Blocks(css=".typewriter {font-family: 'JMH Typewriter', sans-serif;}")
     with demo:
@@ -82,7 +84,6 @@ def main():
             gr.HTML(EVALUATION_HEADER)
             show_stale = gr.Checkbox(label="Show Stale", interactive=True)
             competition_leaderboards = []
-            losses_1 = utils.get_losses_over_time(vali_runs, 1)
             comp_1 = competitions.COMPETITION_DETAILS[1]
             with gr.Accordion(f"{comp_1.name} Competition"):
                 gr.HTML(comp_1.html_description)
@@ -106,7 +107,6 @@ def main():
                     title="Best Average Loss Over Time",
                 )
             comp_2 = competitions.COMPETITION_DETAILS[2]
-            losses_2 = utils.get_losses_over_time(vali_runs, 2)
             # Covert the losses into % of correct answers.
             losses_2["losses"] = losses_2["losses"].apply(lambda x: 1 - x if x else None)
             with gr.Accordion(f"{comp_2.name} Competition"):

utils.py

@@ -422,6 +422,10 @@ def load_state_vars() -> dict[Any]:
             weight_keys = set()
             for uid, stats in validator_df.items():
                 weight_keys.update(stats[-1].keys())
+                
+            # Compute loss over time for all competitions.
+            losses_1 = get_losses_over_time(vali_runs, 1)
+            losses_2 = get_losses_over_time(vali_runs, 2)
 
             # Enable benchmark if the flag is set
             if BENCHMARK_FLAG:
@@ -446,6 +450,8 @@ def load_state_vars() -> dict[Any]:
         "validator_df": validator_df,
         "benchmarks": benchmarks,
         "benchmark_timestamp": benchmark_timestamp,
+        "losses_1": losses_1,
+        "losses_2": losses_2,
     }