from datasets import load_dataset from collections import Counter from sklearn.metrics import pairwise_distances_argmin_min from scipy.spatial.distance import cdist import re import numpy as np import pandas as pd from sklearn.decomposition import PCA from sklearn.cluster import Birch import plotly.express as px from sklearn.metrics import silhouette_score from sklearn.preprocessing import MinMaxScaler class fashionTrendClusterer: def __init__(self, dataset_name="tonyassi/vogue-runway-top15-512px-nobg-embeddings2", top_k_clusters=10): self.dataset_name = dataset_name self.top_k_clusters = top_k_clusters self.dataset = None self.embeddings = None self.labels = None self.seasons = [] self.years = [] self.cluster_labels = None self.reduced_2d = None self.df = None self.top_clusters = None self.rep_indices = [] self.image_paths = [] #option #LOAD VOGUE DATASET FROM HUGGINGFACE def load_dataset(self): dataset = load_dataset(self.dataset_name)["train"] label_mapping = dataset.features["label"].int2str dataset = dataset.filter(lambda example: "menswear" not in label_mapping(example["label"]).lower()) self.dataset = dataset self.embeddings = np.array([item["embeddings"] for item in dataset]) self.labels = [label_mapping(item["label"]) for item in dataset] #EXTRACT LABELS DESIGNER, SEASON, YEAR pattern = r'(\b(?:pre\s)?\w+\b)\s(\d{4})' for label in self.labels: match = re.search(pattern, label) if match: self.seasons.append(match.group(1)) self.years.append(match.group(2)) #BIRCH CLUSTERING PARAMETERS def cluster_embeddings(self, threshold=12): pca = PCA(n_components=150) embeddings_pca = pca.fit_transform(self.embeddings) explained_variance = sum(pca.explained_variance_ratio_) print(f"Cumulative Explained Variance Ratio (PCA): {explained_variance:.2f}") birch = Birch(branching_factor=50, threshold=threshold, n_clusters=None) self.cluster_labels = birch.fit_predict(embeddings_pca) #calc silohuttte score silhouette_avg = silhouette_score(embeddings_pca, self.cluster_labels) print(f"Silhouette Score: {silhouette_avg:.2f}") print(len(set(self.cluster_labels))) cluster_counts = Counter(self.cluster_labels) self.top_clusters = [c[0] for c in cluster_counts.most_common(self.top_k_clusters)] self.reduced_2d = PCA(n_components=2).fit_transform(embeddings_pca) self.df = pd.DataFrame({ "x": self.reduced_2d[:, 0], "y": self.reduced_2d[:, 1], "Cluster": self.cluster_labels, "Designer": [l.split("-")[0].strip() for l in self.labels], "Season": self.seasons, "Year": self.years }) def visualize_clusters(self, interactive=True): """plot cluster scatter graph using plot.ly- dash""" #dataframe labels df_top = self.df[self.df["Cluster"].isin(self.top_clusters)].copy() df_top["Label"] = df_top["Designer"] + " - " + df_top["Season"] + " " + df_top["Year"] # Compute 2D centroids in PCA space centroids_2d = [] for cluster_id in self.top_clusters: cluster_points = self.reduced_2d[self.df["Cluster"] == cluster_id] centroid_2d = cluster_points.mean(axis=0) centroids_2d.append({ "x": centroid_2d[0], "y": centroid_2d[1], "Cluster": str(cluster_id) }) centroids_df = pd.DataFrame(centroids_2d) #scatter graph that plots tha pca points of the top ten clusters if interactive: fig = px.scatter( df_top, x="x", y="y", color=df_top["Cluster"].astype(str), hover_name="Label", title="Top 10 BIRCH Trends from 1995 - 2024", width=1200, height=850 ) fig.update_layout( paper_bgcolor="#FAF6F7", plot_bgcolor="#FFFFFF" ) return fig.to_html(full_html=False) else: return None def get_trend_dataframe(self): """ creates a dataframe of year, season, cluster and returns this dataframe that is for the top ten clusters where the frequecny and year is scaled for the trend forecaster """ trend_df = pd.DataFrame({ "Year": [int(y) for y in self.years], "Season": self.seasons, "Cluster": self.cluster_labels }) trend_df = trend_df.groupby(["Year", "Season", "Cluster"]).size().reset_index(name="Frequency") top_clusters = trend_df.groupby("Cluster")["Frequency"].sum().nlargest(10).index top_clusters_df = trend_df[trend_df["Cluster"].isin(top_clusters)].copy() # Save original values for plotting top_clusters_df["OriginalYear"] = top_clusters_df["Year"] top_clusters_df["OriginalFrequency"] = top_clusters_df["Frequency"] # Scale both Year and Frequency scaler = MinMaxScaler() scaled_values = scaler.fit_transform(top_clusters_df[["Year", "Frequency"]]) top_clusters_df["Year"] = scaled_values[:, 0] top_clusters_df["ScaledFrequency"] = scaled_values[:, 1] return top_clusters_df def get_designer_dataframe(self): """ Returns a DataFrame of Designer, Season, and Year along with the corresponding embeddings. """ designer_names = [label.split("-")[0].strip() for label in self.labels] years_int = [int(y) for y in self.years] meta_df = pd.DataFrame({ "Designer": designer_names, "Season": self.seasons, "Year": years_int }) return meta_df, self.embeddings def find_representative_images(self): """ For each of the top 10 clusters, find the two images closest to the cluster centroid. Save their DataFrame indices in a dictionary: {cluster_id: [idx1, idx2]}. """ self.rep_indices = {} for cluster_id in self.top_clusters: # Filter the DataFrame for current cluster cluster_df = self.df[self.df['Cluster'] == cluster_id].copy() # Get indices and corresponding embeddings cluster_indices = cluster_df.index.tolist() cluster_embeddings = self.embeddings[cluster_indices] # Get centroid of the cluster centroid = np.mean(cluster_embeddings, axis=0) # Compute Euclidean distance from each point to the centroid distances = np.linalg.norm(cluster_embeddings - centroid, axis=1) # Get indices of 2 closest points closest_indices = np.argsort(distances)[:2] # Save original DataFrame indices self.rep_indices[cluster_id] = [cluster_indices[i] for i in closest_indices]