Fix aggregation for historical data

CHANGELOGS.rst

@@ -4,7 +4,8 @@ Change Logs
 0.7.1
 +++++
 
-* :pr:`151`: adds command line ``agg``, class CubeLogsPerformance to produce timeseries
+* :pr:`155`: better aggregation of historical data
+* :pr:`151`, :pr:`153`: adds command line ``agg``, class CubeLogsPerformance to produce timeseries
 * :pr:`152`: add a function to compute fully dynamic shapes given any inputs
 
 0.7.0

_unittests/ut_helpers/test_log_helper.py

@@ -3,6 +3,7 @@
 import textwrap
 import unittest
 import zipfile
+import numpy as np
 import pandas
 from onnx_diagnostic.ext_test_case import ExtTestCase, hide_stdout
 from onnx_diagnostic.helpers.log_helper import (
@@ -46,7 +47,6 @@ def test_cube_logs_load_df(self):
         cube = CubeLogs(df)
         text = str(cube)
         self.assertIsInstance(text, str)
-        self.assertRaise(lambda: cube.load(verbose=1), AssertionError)
         cube = CubeLogs(
             self.df1(),
             recent=True,
@@ -130,12 +130,12 @@ def test_cube_logs_view(self):
         self.assertEqual((2, 6), view.shape)
         self.assertEqual(
             [
-                ("time_baseline", "export", "phi3"),
-                ("time_baseline", "export", "phi4"),
-                ("time_baseline", "onnx-dynamo", "phi4"),
-                ("time_latency", "export", "phi3"),
-                ("time_latency", "export", "phi4"),
-                ("time_latency", "onnx-dynamo", "phi4"),
+                ("time_baseline", "phi3", "export"),
+                ("time_baseline", "phi4", "export"),
+                ("time_baseline", "phi4", "onnx-dynamo"),
+                ("time_latency", "phi3", "export"),
+                ("time_latency", "phi4", "export"),
+                ("time_latency", "phi4", "onnx-dynamo"),
             ],
             list(view.columns),
         )
@@ -229,6 +229,55 @@ def test_cube_logs_performance(self):
         )
         self.assertExists(output)
 
+    def test_duplicate(self):
+        df = pandas.DataFrame(
+            [
+                dict(date="2025/01/01", time_engine=0.5, model_name="A", version_engine="0.5"),
+                dict(date="2025/01/01", time_engine=0.5, model_name="A", version_engine="0.5"),
+            ]
+        )
+        cube = CubeLogs(df)
+        self.assertRaise(lambda: cube.load(), AssertionError)
+        CubeLogs(df, recent=True).load()
+
+    def test_historical(self):
+        # case 1
+        df = pandas.DataFrame(
+            [
+                dict(date="2025/01/01", time_p=0.51, exporter="E1", m_name="A", m_cls="CA"),
+                dict(date="2025/01/02", time_p=0.62, exporter="E1", m_name="A", m_cls="CA"),
+                dict(date="2025/01/01", time_p=0.53, exporter="E2", m_name="A", m_cls="CA"),
+                dict(date="2025/01/02", time_p=0.64, exporter="E2", m_name="A", m_cls="CA"),
+                dict(date="2025/01/01", time_p=0.55, exporter="E2", m_name="B", m_cls="CA"),
+                dict(date="2025/01/02", time_p=0.66, exporter="E2", m_name="B", m_cls="CA"),
+            ]
+        )
+        cube = CubeLogs(df, keys=["^m_*", "exporter"]).load()
+        view, view_def = cube.view(CubeViewDef(["^m_.*"], ["^time_.*"]), return_view_def=True)
+        self.assertEqual(
+            "CubeViewDef(key_index=['^m_.*'], values=['^time_.*'])", repr(view_def)
+        )
+        self.assertEqual(["METRICS", "exporter", "date"], view.columns.names)
+        got = view.values.ravel()
+        self.assertEqual(
+            sorted([0.51, 0.62, 0.53, 0.64, -1, -1, 0.55, 0.66]),
+            sorted(np.where(np.isnan(got), -1, got).tolist()),
+        )
+
+        # case 2
+        df = pandas.DataFrame(
+            [
+                dict(date="2025/01/02", time_p=0.62, exporter="E1", m_name="A", m_cls="CA"),
+                dict(date="2025/01/02", time_p=0.64, exporter="E2", m_name="A", m_cls="CA"),
+                dict(date="2025/01/01", time_p=0.51, exporter="E1", m_name="B", m_cls="CA"),
+                dict(date="2025/01/02", time_p=0.66, exporter="E2", m_name="B", m_cls="CA"),
+            ]
+        )
+        cube = CubeLogs(df, keys=["^m_*", "exporter"]).load()
+        view, view_def = cube.view(CubeViewDef(["^m_.*"], ["^time_.*"]), return_view_def=True)
+        self.assertEqual((2, 3), view.shape)
+        self.assertEqual(["METRICS", "exporter", "date"], view.columns.names)
+
 
 if __name__ == "__main__":
     unittest.main(verbosity=2)

onnx_diagnostic/_command_lines_parser.py

@@ -656,29 +656,31 @@ def get_parser_agg() -> ArgumentParser:
     parser.add_argument(
         "-w",
         "--values",
-        default="^time_.*,^disc.*,^ERR_.*,CMD,^ITER.*,^onnx_.*,^op_onnx_.*",
+        default="^time_.*,^disc.*,^ERR_.*,CMD,^ITER.*,^onnx_.*,^op_onnx_.*,^peak_gpu_.*",
         help="List of columns to consider as values, "
         "multiple values are separated by `,`\n"
         "regular expressions are allowed",
     )
     parser.add_argument(
-        "-i", "--ignored", default="version_python", help="List of columns to ignore"
+        "-i", "--ignored", default="^version_.*", help="List of columns to ignore"
     )
     parser.add_argument(
         "-f",
         "--formula",
-        default="speedup,bucket[speedup],ERR1,n_models,n_eager,"
-        "n_running,n_acc01,n_acc001,n_dynamic,n_pass,n_faster,"
-        "n_faster2x,n_faster3x,n_faster4x,n_attention,"
-        "peak_gpu_torch,peak_gpu_nvidia,n_control_flow,"
-        "n_constant,n_shape,n_expand,"
-        "n_function,n_initializer,n_scatter,time_export_unbiased",
+        default="speedup,bucket[speedup],ERR1,n_models,n_model_eager,"
+        "n_model_running,n_model_acc01,n_model_acc001,n_model_dynamic,"
+        "n_model_pass,n_model_faster,"
+        "n_model_faster2x,n_model_faster3x,n_model_faster4x,n_node_attention,"
+        "peak_gpu_torch,peak_gpu_nvidia,n_node_control_flow,"
+        "n_node_constant,n_node_shape,n_node_expand,"
+        "n_node_function,n_node_initializer,n_node_scatter,"
+        "time_export_unbiased",
         help="Columns to compute after the aggregation was done.",
     )
     parser.add_argument(
         "--views",
         default="agg-suite,disc,speedup,time,time_export,err,cmd,"
-        "bucket-speedup,raw-short,counts",
+        "bucket-speedup,raw-short,counts,peak-gpu",
         help="Views to add to the output files.",
     )
     parser.add_argument(
@@ -702,7 +704,7 @@ def _cmd_agg(argv: List[Any]):
             args.inputs, verbose=args.verbose, filtering=lambda name: bool(reg.search(name))
         )
     )
-    assert csv, f"No csv files in {args.inputs}"
+    assert csv, f"No csv files in {args.inputs}, csv={csv}"
     if args.verbose:
         from tqdm import tqdm
 
@@ -712,7 +714,9 @@ def _cmd_agg(argv: List[Any]):
     dfs = []
     for c in loop:
         df = open_dataframe(c)
-        assert args.time in df.columns, f"Missing time column {args.time!r} in {c.head()!r}"
+        assert (
+            args.time in df.columns
+        ), f"Missing time column {args.time!r} in {c!r}\n{df.head()}\n{sorted(df.columns)}"
         dfs.append(df)
 
     cube = CubeLogsPerformance(