apply rounding before calculating pareto front

Author: mertalev

docs/docs/features/searching.md

@@ -118,7 +118,7 @@ Memory and execution time estimates were obtained without acceleration on a 7800
 | ViT-H-14__laion2b-s32b-b79k                          | 4676         | 39.06               | 82.36      | ❌             |
 | ViT-B-32-SigLIP2-256__webli                          | 3061         | 3.31                | 82.28      | ✅             |
 | ViT-B-16-SigLIP__webli                               | 1081         | 5.77                | 81.9       | ✅             |
-| ViT-B-16-SigLIP-256__webli                           | 1102         | 7.11                | 81.9       | ✅             |
+| ViT-B-16-SigLIP-256__webli                           | 1102         | 7.11                | 81.9       | ❌             |
 | ViT-L-14__laion2b-s32b-b82k                          | 2233         | 20.56               | 80.82      | ❌             |
 | nllb-clip-base-siglip__mrl                           | 4696         | 16.95               | 80.65      | ❌             |
 | nllb-clip-base-siglip__v1                            | 4675         | 15.17               | 80.16      | ❌             |
@@ -423,7 +423,7 @@ Memory and execution time estimates were obtained without acceleration on a 7800
 | Model                                                | Memory (MiB) | Execution Time (ms) | Recall (%) | Pareto Optimal |
 |------------------------------------------------------|--------------|---------------------|------------|----------------|
 | ViT-SO400M-16-SigLIP2-384__webli                     | 3854         | 56.57               | 86.5       | ✅             |
-| ViT-SO400M-16-SigLIP2-512__webli                     | 4050         | 107.67              | 86.5       | ✅             |
+| ViT-SO400M-16-SigLIP2-512__webli                     | 4050         | 107.67              | 86.5       | ❌             |
 | ViT-SO400M-14-SigLIP2-378__webli                     | 3940         | 72.25               | 86.39      | ❌             |
 | ViT-gopt-16-SigLIP2-384__webli                       | 6585         | 146.84              | 86.15      | ❌             |
 | ViT-H-14-378-quickgelu__dfn5b                        | 5049         | 108.4               | 86.1       | ❌             |
@@ -698,7 +698,7 @@ Memory and execution time estimates were obtained without acceleration on a 7800
 | ViT-B-16-SigLIP-512__webli                           | 1828         | 26.17               | 76.51      | ✅             |
 | ViT-B-16-SigLIP-384__webli                           | 1128         | 13.53               | 76.08      | ✅             |
 | ViT-B-16-SigLIP__webli                               | 1081         | 5.77                | 75.29      | ✅             |
-| ViT-B-16-SigLIP-256__webli                           | 1102         | 7.11                | 75.29      | ✅             |
+| ViT-B-16-SigLIP-256__webli                           | 1102         | 7.11                | 75.29      | ❌             |
 | ViT-SO400M-14-SigLIP-384__webli                      | 4417         | 72.19               | 74.84      | ❌             |
 | ViT-H-14__laion2b-s32b-b79k                          | 4676         | 39.06               | 56.32      | ❌             |
 | ViT-L-14__laion2b-s32b-b82k                          | 2233         | 20.56               | 47.25      | ❌             |

machine-learning/export/immich_model_exporter/parse_eval_data.py

@@ -77,10 +77,11 @@ def collapsed_table(language: str, df: pl.DataFrame) -> str:
 )
 eval_df = eval_df.with_columns(
     recall=(
-        pl.col("image_retrieval_recall@1") + pl.col("image_retrieval_recall@5") + pl.col("image_retrieval_recall@10")
-    )
-    / 3
-)
+        (pl.col("image_retrieval_recall@1") + pl.col("image_retrieval_recall@5") + pl.col("image_retrieval_recall@10"))
+        * (100 / 3)
+    ).round(2)
+).collect()
+eval_df.write_parquet("model_info.parquet")
 
 pareto_front = eval_df.join_where(
     eval_df.select("language", "peak_rss", "exec_time_ms", "recall").rename(
@@ -103,14 +104,11 @@ def collapsed_table(language: str, df: pl.DataFrame) -> str:
 )
 eval_df = eval_df.join(pareto_front, on=["pretrained_model", "language"], how="left")
 eval_df = eval_df.with_columns(is_pareto=pl.col("recall_other").is_null())
-eval_df = (
-    eval_df.drop("peak_rss_other", "exec_time_ms_other", "recall_other", "language_other")
-    .unique(subset=["pretrained_model", "language"])
-    .collect()
+eval_df = eval_df.drop("peak_rss_other", "exec_time_ms_other", "recall_other", "language_other").unique(
+    subset=["pretrained_model", "language"]
 )
-eval_df.write_parquet("model_info.parquet")
 
-eval_df = eval_df.filter(pl.col("recall") >= 0.2)
+eval_df = eval_df.filter(pl.col("recall") >= 20)
 eval_df = eval_df.select(
     pl.col("pretrained_model").alias("Model"),
     (pl.col("peak_rss") / 1024).round().cast(pl.UInt32).alias("Memory (MiB)"),
@@ -119,7 +117,7 @@ def collapsed_table(language: str, df: pl.DataFrame) -> str:
     # pl.col("image_retrieval_recall@1").mul(100).round(2).alias("Recall@1 (%)"),
     # pl.col("image_retrieval_recall@5").mul(100).round(2).alias("Recall@5 (%)"),
     # pl.col("image_retrieval_recall@10").mul(100).round(2).alias("Recall@10 (%)"),
-    pl.col("recall").mul(100).round(2).alias("Recall (%)"),
+    pl.col("recall").alias("Recall (%)"),
     pl.when("is_pareto").then(pl.lit("✅")).otherwise(pl.lit("❌")).alias("Pareto Optimal"),
 )
 eval_df = eval_df.sort("Recall (%)", "Memory (MiB)", descending=[True, False])