feat(benchmarks): add evaluation framework for DINO-based compressors

mini-nav/benchmarks/__init__.py

@@ -0,0 +1,49 @@
+from typing import Literal, cast
+
+import torch
+from compressors import DinoCompressor, FloatCompressor
+from transformers import AutoImageProcessor, BitImageProcessorFast
+from utils import get_device, get_output_diretory
+
+from .task_eval import task_eval
+
+
+def evaluate(
+    compressor_model: Literal["Dinov2", "Dinov2WithCompressor"],
+    dataset: Literal["CIFAR-10", "CIFAR-100"],
+    benchmark: Literal["Recall@1", "Recall@10"],
+):
+    match compressor_model:
+        case "Dinov2":
+            processor = cast(
+                BitImageProcessorFast,
+                AutoImageProcessor.from_pretrained(
+                    "facebook/dinov2-large", device_map=get_device()
+                ),
+            )
+            model = DinoCompressor().to(get_device())
+        case "Dinov2WithCompressor":
+            processor = cast(
+                BitImageProcessorFast,
+                AutoImageProcessor.from_pretrained(
+                    "facebook/dinov2-large", device_map=get_device()
+                ),
+            )
+
+            compressor = FloatCompressor().load_state_dict(
+                torch.load(get_output_diretory() / "compressor.pt")
+            )
+            model = DinoCompressor(compressor).to(get_device())
+        case _:
+            raise ValueError(f"Unknown compressor: {compressor_model}")
+
+    match benchmark:
+        case "Recall@1":
+            task_eval(processor, model, dataset, 1)
+        case "Recall@10":
+            task_eval(processor, model, dataset, 10)
+        case _:
+            raise ValueError(f"Unknown benchmark: {benchmark}")
+
+
+__all__ = ["task_eval", "evaluate"]

mini-nav/benchmarks/task_eval.py

@@ -0,0 +1,77 @@
+from typing import Literal, cast
+
+import polars as pl
+import torch
+from datasets import Dataset, load_dataset
+from torch import Tensor, nn
+from torch.utils.data import DataLoader
+from tqdm.auto import tqdm
+from transformers import BitImageProcessorFast
+from utils import get_device
+
+
+def establish_database(
+    processor: BitImageProcessorFast,
+    model: nn.Module,
+    dataset: Dataset,
+    batch_size: int = 32,
+) -> pl.DataFrame:
+    df = pl.DataFrame()
+
+    model.eval()
+    dataloader = DataLoader(
+        dataset.with_format("torch"), batch_size=batch_size, shuffle=True, num_workers=4
+    )
+
+    with torch.no_grad():
+        for batch in tqdm(dataloader, desc="Establish Database"):
+            imgs = batch["img"]
+            labels = batch["label"]
+
+            inputs = processor(imgs, return_tensors="pt").to(get_device())
+
+            outputs = cast(Tensor, model(inputs))
+
+    return df
+
+
+def task_eval(
+    processor: BitImageProcessorFast,
+    model: nn.Module,
+    dataset: Literal["CIFAR-10", "CIFAR-100"],
+    top_k: int = 10,
+    batch_size: int = 32,
+):
+    match dataset:
+        case "CIFAR-10":
+            train_dataset = load_dataset("uoft-cs/cifar10", split="train")
+            test_dataset = load_dataset("uoft-cs/cifar10", split="test")
+        case "CIFAR-100":
+            train_dataset = load_dataset("uoft-cs/cifar100", split="train")
+            test_dataset = load_dataset("uoft-cs/cifar100", split="test")
+        case _:
+            raise ValueError(
+                f"Unknown dataset: {dataset}. Only support: 'CIFAR-10', 'CIFAR-100'."
+            )
+
+    # Establish database
+    df = establish_database(processor, model, train_dataset, batch_size)
+
+    # Test
+    dataloader = DataLoader(
+        test_dataset.with_format("torch"),
+        batch_size=batch_size,
+        shuffle=True,
+        num_workers=4,
+    )
+
+    with torch.no_grad():
+        for batch in tqdm(dataloader, desc="Test Evaluation"):
+            imgs = batch["img"]
+            labels = batch["label"]
+
+            inputs = processor(imgs, return_tensors="pt").to(get_device())
+
+            outputs = cast(Tensor, model(inputs))
+            for vec in outputs:
+                pass

mini-nav/compressors/__init__.py

@@ -1,5 +1,6 @@
+from .dino_compressor import DinoCompressor
 from .float_compressor import FloatCompressor
 from .int_compressor import IntCompressor
 from .train import train
 
-__all__ = ["train", "FloatCompressor", "IntCompressor"]
+__all__ = ["train", "FloatCompressor", "IntCompressor", "DinoCompressor"]

mini-nav/compressors/dino_compressor.py

@@ -0,0 +1,29 @@
+from typing import Optional, cast
+
+import torch.nn.functional as F
+from torch import nn
+from transformers import AutoModel, Dinov2Model
+
+
+class DinoCompressor(nn.Module):
+    def __init__(self, compressor: Optional[nn.Module] = None):
+        super().__init__()
+
+        self.dino = cast(
+            Dinov2Model,
+            AutoModel.from_pretrained("facebook/dinov2-large"),
+        )
+
+        self.compressor = compressor
+
+    def forward(self, inputs):
+        teacher_tokens = self.dino(**inputs).last_hidden_state  # [B,N,1024]
+
+        teacher_embed = teacher_tokens.mean(dim=1)
+        teacher_embed = F.normalize(teacher_embed, dim=-1)  # [B,1024]
+
+        if self.compressor is None:
+            return teacher_embed
+
+        feats, recon = self.compressor(teacher_tokens)
+        return feats

mini-nav/compressors/train.py

@@ -93,12 +93,10 @@ def train(
                 with torch.no_grad():
                     inputs = processor(imgs, return_tensors="pt").to(device)
 
-                    teacher_tokens = dino(**inputs).last_hidden_state
-                    # [B,N,1024]
+                    teacher_tokens = dino(**inputs).last_hidden_state  # [B,N,1024]
 
                     teacher_embed = teacher_tokens.mean(dim=1)
-                    teacher_embed = F.normalize(teacher_embed, dim=-1)
-                    # [B,1024]
+                    teacher_embed = F.normalize(teacher_embed, dim=-1)  # [B,1024]
 
                 # ---- student forward ----
                 z512, recon = compressor(teacher_tokens)

mini-nav/feature_retrieval.py

@@ -6,8 +6,14 @@ from database import db_manager
 from datasets import load_dataset
 from PIL import Image
 from PIL.PngImagePlugin import PngImageFile
+from torch import nn
 from tqdm.auto import tqdm
-from transformers import AutoImageProcessor, AutoModel
+from transformers import (
+    AutoImageProcessor,
+    AutoModel,
+    BitImageProcessorFast,
+    Dinov2Model,
+)
 
 
 def pil_image_to_bytes(image: Image.Image, format: str = "PNG") -> bytes:
@@ -31,8 +37,8 @@ class FeatureRetrieval:
     _instance: Optional["FeatureRetrieval"] = None
 
     _initialized: bool = False
-    processor: Any
-    model: Any
+    processor: BitImageProcessorFast
+    model: nn.Module
 
     def __new__(cls, *args, **kwargs) -> "FeatureRetrieval":
         if cls._instance is None:
@@ -40,7 +46,9 @@ class FeatureRetrieval:
         return cls._instance
 
     def __init__(
-        self, processor: Optional[Any] = None, model: Optional[Any] = None
+        self,
+        processor: Optional[BitImageProcessorFast] = None,
+        model: Optional[nn.Module] = None,
     ) -> None:
         """Initialize the singleton with processor and model.
 
@@ -84,10 +92,10 @@ class FeatureRetrieval:
         for i in tqdm(range(0, len(images), batch_size)):
             batch_imgs = images[i : i + batch_size]
 
-            inputs = self.processor(images=batch_imgs, return_tensors="pt")
+            inputs = self.processor(batch_imgs, return_tensors="pt")
 
             # 迁移数据到GPU
-            inputs.to(device, non_blocking=True)
+            inputs.to(device)
 
             outputs = self.model(**inputs)
 
@@ -166,10 +174,14 @@ if __name__ == "__main__":
         "truck",
     ]
 
-    processor = AutoImageProcessor.from_pretrained(
-        "facebook/dinov2-large", device_map="cuda"
+    processor = cast(
+        BitImageProcessorFast,
+        AutoImageProcessor.from_pretrained("facebook/dinov2-large", device_map="cuda"),
+    )
+    model = cast(
+        Dinov2Model,
+        AutoModel.from_pretrained("facebook/dinov2-large", device_map="cuda"),
     )
-    model = AutoModel.from_pretrained("facebook/dinov2-large", device_map="cuda")
 
     feature_retrieval = FeatureRetrieval(processor, model)
 

mini-nav/main.py

@@ -2,14 +2,22 @@ import argparse
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
-    parser.add_argument("train")
+    parser.add_argument(
+        "action",
+        choices=["train", "benchmark", "visualize"],
+        help="Action to perform: train, benchmark, or visualize",
+    )
     args = parser.parse_args()
 
-    if args.train:
+    if args.action == "train":
         from compressors import FloatCompressor, train
 
         train(FloatCompressor(), 1, 32)
-    else:
+    elif args.action == "benchmark":
+        from benchmarks import evaluate
+
+        evaluate("Dinov2", "CIFAR-10", "Recall@10")
+    else:  # visualize
         from visualizer import app
 
         app.run(debug=True)

mini-nav/utils/__init__.py

@@ -0,0 +1,3 @@
+from .common import get_device, get_output_diretory
+
+__all__ = ["get_device", "get_output_diretory"]

mini-nav/utils/common.py

@@ -0,0 +1,23 @@
+from functools import lru_cache
+from pathlib import Path
+
+import torch
+from configs import cfg_manager
+from torch.types import Device
+
+
+@lru_cache(maxsize=1)
+def get_device() -> Device:
+    config = cfg_manager.get()
+    device = config.model.device
+    if device == "auto":
+        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    else:
+        device = torch.device(device)
+    return device
+
+
+@lru_cache(maxsize=1)
+def get_output_diretory() -> Path:
+    config = cfg_manager.get()
+    return config.output.directory