add: initial files.

app.py

@@ -0,0 +1,69 @@
+"""
+Thanks to Freddy Boulton (https://github.com/freddyaboulton) for helping with this.
+"""
+
+
+import pickle
+
+import gradio as gr
+from datasets import load_dataset
+from transformers import AutoModel
+
+from similarity_utils import BuildLSHTable
+
+seed = 42
+
+# Only runs once when the script is first run.
+with open("lsh.pickle", "rb") as handle:
+    loaded_lsh = pickle.load(handle)
+
+# Load model for computing embeddings.
+model_ckpt = "nateraw/vit-base-beans"
+model = AutoModel.from_pretrained(model_ckpt)
+lsh_builder = BuildLSHTable(model)
+lsh_builder.lsh = loaded_lsh
+
+# Candidate images.
+dataset = load_dataset("beans")
+candidate_dataset = dataset["train"].shuffle(seed=seed)
+
+
+def query(image, top_k):
+    results = lsh_builder.query(image)
+
+    # Should be a list of string file paths for gr.Gallery to work
+    images = []
+    # List of labels for each image in the gallery
+    labels = []
+
+    candidates = []
+    overlaps = []
+
+    for idx, r in enumerate(sorted(results, key=results.get, reverse=True)):
+        if idx == top_k:
+            break
+        image_id, label = r.split("_")[0], r.split("_")[1]
+        candidates.append(candidate_dataset[int(image_id)]["image"])
+        labels.append(label)
+        overlaps.append(results[r])
+
+    candidates.insert(0, image)
+    labels.insert(0, label)
+
+    for i, candidate in enumerate(candidates):
+        filename = f"{i}.png"
+        candidate.save(filename)
+        images.append(filename)
+
+    # The gallery component can be a list of tuples, where the first element is a path to a file
+    # and the second element is an optional caption for that image
+    return list(zip(images, labels))
+
+
+# You can set the type of gr.Image to be PIL, numpy or str (filepath)
+# Not sure what the best for this demo is.
+gr.Interface(
+    query,
+    inputs=[gr.Image(), gr.Slider(value=5, minimum=1, maximum=10, step=1)],
+    outputs=gr.Gallery(),
+).launch()

lsh.pickle

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:caa1727832f2279a4026b03b9f17638ff4a4deffa0a28586e74db59332dce732
+size 136667

requirements.txt

@@ -0,0 +1,3 @@
+transformers==4.25.1
+datasets==2.7.1
+numpy==1.21.6

similarity_utils.py

@@ -0,0 +1,175 @@
+from typing import List, Union
+
+import datasets
+import numpy as np
+import torch
+import torchvision.transforms as T
+from PIL import Image
+from tqdm.auto import tqdm
+from transformers import AutoFeatureExtractor, AutoModel
+
+seed = 42
+hash_size = 8
+hidden_dim = 768  # ViT-base
+np.random.seed(seed)
+
+
+# Device.
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+# Load model for computing embeddings..
+model_ckpt = "nateraw/vit-base-beans"
+extractor = AutoFeatureExtractor.from_pretrained(model_ckpt)
+
+# Data transformation chain.
+transformation_chain = T.Compose(
+    [
+        # We first resize the input image to 256x256 and then we take center crop.
+        T.Resize(int((256 / 224) * extractor.size["height"])),
+        T.CenterCrop(extractor.size["height"]),
+        T.ToTensor(),
+        T.Normalize(mean=extractor.image_mean, std=extractor.image_std),
+    ]
+)
+
+
+# Define random vectors to project with.
+random_vectors = np.random.randn(hash_size, hidden_dim).T
+
+
+def hash_func(embedding, random_vectors=random_vectors):
+    """Randomly projects the embeddings and then computes bit-wise hashes."""
+    if not isinstance(embedding, np.ndarray):
+        embedding = np.array(embedding)
+    if len(embedding.shape) < 2:
+        embedding = np.expand_dims(embedding, 0)
+
+    # Random projection.
+    bools = np.dot(embedding, random_vectors) > 0
+    return [bool2int(bool_vec) for bool_vec in bools]
+
+
+def bool2int(x):
+    y = 0
+    for i, j in enumerate(x):
+        if j:
+            y += 1 << i
+    return y
+
+
+def compute_hash(model: Union[torch.nn.Module, str]):
+    """Computes hash on a given dataset."""
+    device = model.device
+
+    def pp(example_batch):
+        # Prepare the input images for the model.
+        image_batch = example_batch["image"]
+        image_batch_transformed = torch.stack(
+            [transformation_chain(image) for image in image_batch]
+        )
+        new_batch = {"pixel_values": image_batch_transformed.to(device)}
+
+        # Compute embeddings and pool them i.e., take the representations from the [CLS]
+        # token.
+        with torch.no_grad():
+            embeddings = model(**new_batch).last_hidden_state[:, 0].cpu().numpy()
+
+        # Compute hashes for the batch of images.
+        hashes = [hash_func(embeddings[i]) for i in range(len(embeddings))]
+        example_batch["hashes"] = hashes
+        return example_batch
+
+    return pp
+
+
+class Table:
+    def __init__(self, hash_size: int):
+        self.table = {}
+        self.hash_size = hash_size
+
+    def add(self, id: int, hashes: List[int], label: int):
+        # Create a unique indentifier.
+        entry = {"id_label": str(id) + "_" + str(label)}
+
+        # Add the hash values to the current table.
+        for h in hashes:
+            if h in self.table:
+                self.table[h].append(entry)
+            else:
+                self.table[h] = [entry]
+
+    def query(self, hashes: List[int]):
+        results = []
+
+        # Loop over the query hashes and determine if they exist in
+        # the current table.
+        for h in hashes:
+            if h in self.table:
+                results.extend(self.table[h])
+        return results
+
+
+class LSH:
+    def __init__(self, hash_size, num_tables):
+        self.num_tables = num_tables
+        self.tables = []
+        for i in range(self.num_tables):
+            self.tables.append(Table(hash_size))
+
+    def add(self, id: int, hash: List[int], label: int):
+        for table in self.tables:
+            table.add(id, hash, label)
+
+    def query(self, hashes: List[int]):
+        results = []
+        for table in self.tables:
+            results.extend(table.query(hashes))
+        return results
+
+
+class BuildLSHTable:
+    def __init__(
+        self,
+        model: Union[torch.nn.Module, None],
+        batch_size: int = 48,
+        hash_size: int = hash_size,
+        dim: int = hidden_dim,
+        num_tables: int = 10,
+    ):
+        self.hash_size = hash_size
+        self.dim = dim
+        self.num_tables = num_tables
+        self.lsh = LSH(self.hash_size, self.num_tables)
+
+        self.batch_size = batch_size
+        self.hash_fn = compute_hash(model.to(device))
+
+    def build(self, ds: datasets.DatasetDict):
+        dataset_hashed = ds.map(self.hash_fn, batched=True, batch_size=self.batch_size)
+
+        for id in tqdm(range(len(dataset_hashed))):
+            hash, label = dataset_hashed[id]["hashes"], dataset_hashed[id]["labels"]
+            self.lsh.add(id, hash, label)
+
+    def query(self, image, verbose=True):
+        if isinstance(image, str):
+            image = Image.open(image).convert("RGB")
+
+        # Compute the hashes of the query image and fetch the results.
+        example_batch = dict(image=[image])
+        hashes = self.hash_fn(example_batch)["hashes"][0]
+
+        results = self.lsh.query(hashes)
+        if verbose:
+            print("Matches:", len(results))
+
+        # Calculate Jaccard index to quantify the similarity.
+        counts = {}
+        for r in results:
+            if r["id_label"] in counts:
+                counts[r["id_label"]] += 1
+            else:
+                counts[r["id_label"]] = 1
+        for k in counts:
+            counts[k] = float(counts[k]) / self.dim
+        return counts