item-matching - npm Package Compare versions

Comparing version 0.0.106 to 0.0.107

notebooks/benchmark_model/create_img_embed.py

 import duckdb
 import polars as pl
-from transformers import SiglipVisionModel, Dinov2WithRegistersModel
+from transformers import (
+    SiglipVisionModel,
+    Dinov2WithRegistersModel,
+    Siglip2VisionModel,
+    AutoProcessor,
+    Siglip2VisionConfig,
+)
 from PIL import Image
@@ -11,5 +17,5 @@ from accelerate import Accelerator
 from torchvision import transforms
-from numpy.lib.format import open_memmap
 from core_pro.ultilities import make_sync_folder
 from tqdm.auto import tqdm
+import numpy as np
 
@@ -25,5 +31,5 @@ device = Accelerator().device
             [
-                transforms.Resize(
-                    img_size, interpolation=transforms.InterpolationMode.BICUBIC
-                ),
+                # transforms.Resize(
+                #     img_size, interpolation=transforms.InterpolationMode.BICUBIC
+                # ),
                 transforms.CenterCrop(img_size),
@@ -43,3 +49,4 @@ transforms.ConvertImageDtype(torch.float32),  # to [0,1]
         tensor = transforms.ToTensor()(img)  # HWC→CHW float32
-        return self.transform(tensor)
+        tensor = self.transform(tensor)
+        return tensor
 
@@ -61,5 +68,29 @@
     )
-    return torch.compile(img_model)
+    # return torch.compile(img_model)
+    return img_model
 
 
+def setup_siglip2():
+    pretrain_name = "google/siglip2-base-patch16-224"
+
+    config = Siglip2VisionConfig(
+        image_size=224,  # 224/16 = 14 patches per side → 196 total
+        patch_size=16,
+        num_channels=3,
+        embed_dim=768,
+        patch_embed_type="conv",  # if SigLIP-2 supports choosing Conv vs Linear
+    )
+    img_model = Siglip2VisionModel(config)
+
+    img_model = (
+        img_model.from_pretrained(
+            pretrain_name, torch_dtype=torch.bfloat16, ignore_mismatched_sizes=True
+        )
+        .to(device)
+        .eval()
+    )
+    processor = AutoProcessor.from_pretrained(pretrain_name)
+    return torch.compile(img_model), processor
+
+
 def setup_dinov2():
@@ -75,3 +106,4 @@ pretrain_name = "facebook/dinov2-with-registers-base"
     )
-    return torch.compile(img_model)
+    # return torch.compile(img_model)
+    return img_model
 
@@ -86,3 +118,12 @@
 ):
-    # 1) Prepare DataLoader ---
+    # 1) Load & compile model in mixed precision
+    device = torch.device("cuda")
+    if model == "siglip":
+        img_model = setup_siglip()
+    elif model == "siglip2":
+        img_model, processor = setup_siglip2()
+    else:
+        img_model = setup_dinov2()
+
+    # 2) Prepare DataLoader ---
     ds = ImagePathsDataset(file_paths, img_size=224)
@@ -98,22 +139,5 @@ loader = DataLoader(
 
-    # 2) Load & compile model in mixed precision
-    device = torch.device("cuda")
-    if model == "siglip":
-        img_model = setup_siglip()
-    else:
-        img_model = setup_dinov2()
-
-    # 3) Pre‑allocate a .npy memmap for all embeddings
-    total = len(ds)
-    dim = img_model.config.hidden_size  # e.g. 1024
-    mmap = open_memmap(
-        filename=str(path / f"{model}_embeds.npy"),
-        mode="w+",
-        dtype="float32",
-        shape=(total, dim),
-    )
-
-    # 4) Inference + save loop
+    # 3) Inference + save loop
     start = perf_counter()
-    idx = 0
+    list_embeds = []
     with torch.inference_mode():
@@ -128,8 +152,7 @@ for batch in tqdm(loader):
             emb = normed.cpu().numpy().astype("float32")  # (B, dim)
-            bs = emb.shape[0]
-            mmap[idx : idx + bs] = emb  # write into .npy
-            idx += bs
+            list_embeds.append(emb)
 
+    embeds = np.concatenate(list_embeds, axis=0)
+    np.save(path / f"{model}_embeds.npy", embeds)
     durations = perf_counter() - start
-    mmap.flush()  # ensure all data is on disk
 
@@ -147,3 +170,4 @@ result.append((model, durations))
 query = f"""
-select *
+select * exclude(file_path)
+, REPLACE(file_path, 'data_4t', '75b198db-809a-4bd2-a97c-e52daa6b3a2d') AS file_path
 from read_parquet('{file}')
@@ -156,8 +180,9 @@ """
 result = []
+# result = fast_img_inference(result=result, file_paths=file_paths, model="siglip2")
 result = fast_img_inference(result=result, file_paths=file_paths, model="siglip")
 result = fast_img_inference(result=result, file_paths=file_paths, model="dinov2")
 
-# result
-df_result = pl.DataFrame(result, orient="row", schema=["name", "duration"])
-df_result.write_csv(path / "img_embed_benchmark.csv")
-print(df_result)
+# # result
+# df_result = pl.DataFrame(result, orient="row", schema=["name", "duration"])
+# df_result.write_csv(path / "img_embed_benchmark.csv")
+# print(df_result)

PKG-INFO

 Metadata-Version: 2.4
 Name: item_matching
-Version: 0.0.106
+Version: 0.0.107
 Summary: A name matching package
@@ -5,0 +5,0 @@ Project-URL: Homepage, https://github.com/kevinkhang2909/item_matching

pyproject.toml

@@ -7,3 +7,3 @@ [build-system]
 name = "item_matching"
-version = "0.0.106"
+version = "0.0.107"
 authors = [
@@ -10,0 +10,0 @@ { name="Kevin Khang", email="[e-mail]" },

src/item_matching/pipeline/data_loading.py

@@ -14,3 +14,3 @@ from PIL import Image
 from FlagEmbedding import BGEM3FlagModel
-from transformers import Dinov2WithRegistersModel, Siglip2VisionModel
+from transformers import Dinov2WithRegistersModel, SiglipVisionModel
 from .func import _create_folder
@@ -53,3 +53,3 @@
                 # ),
-                # transforms.CenterCrop(img_size),
+                transforms.CenterCrop(img_size),
                 transforms.ConvertImageDtype(torch.float32),  # to [0,1]
@@ -68,9 +68,10 @@ transforms.Normalize(
         tensor = transforms.ToTensor()(img)  # HWC→CHW float32
-        return self.transform(tensor)
+        tensor = self.transform(tensor)
+        return tensor
 
 
-def get_img_model():
-    pretrain_name = "google/siglip2-base-patch16-224"
+def setup_dinov2():
+    pretrain_name = "facebook/dinov2-with-registers-base"
     img_model = (
-        Siglip2VisionModel.from_pretrained(
+        Dinov2WithRegistersModel.from_pretrained(
             pretrain_name,
@@ -82,12 +83,16 @@ torch_dtype=torch.bfloat16,
     )
+    # return torch.compile(img_model)
+    return img_model
 
-    # pretrain_name = "facebook/dinov2-with-registers-base"
-    # img_model = (
-    #     Dinov2WithRegistersModel.from_pretrained(
-    #         pretrain_name,
-    #         torch_dtype=torch.bfloat16,
-    #     )
-    #     .to(device)
-    #     .eval()
-    # )
+
+def setup_siglip():
+    pretrain_name = "google/siglip-base-patch16-224"
+    img_model = (
+        SiglipVisionModel.from_pretrained(
+            pretrain_name,
+            torch_dtype=torch.bfloat16,
+        )
+        .to(device)
+        .eval()
+    )
     # return torch.compile(img_model)
@@ -140,2 +145,3 @@ return img_model
         SHARD_SIZE: int = 1_500_000,
+        model_name: str = "dinov2",
     ):
@@ -146,2 +152,3 @@ # Config
         self.SHARD_SIZE = SHARD_SIZE
+        self.model_name = model_name
 
@@ -166,3 +173,6 @@ # Path
             self.col_embedding = f"{self.MATCH_BY}_embed"
-            self.img_model = get_img_model()
+            if self.model_name == "siglip":
+                self.img_model = setup_siglip()
+            else:
+                self.img_model = setup_dinov2()
 
@@ -169,0 +179,0 @@ def load(self, data: pl.DataFrame):

Alert delta unavailable

Currently unable to show alert delta for PyPI packages.

Improved metrics

Lines of code

1300

2.44%

Worsened metrics

Total package byte prevSize

3254300

-9.55%