You can deploy this model and start earning money today!

imranx654/
Fashion-Recommender-System

Name: Fashion-Recommender-System
Availability: InStock

apache-2.0

Views

No views yet

Fashion Product Recommendation System

Deep Learning + Similarity Search + Cloud Storage

Project Site link

Overview

This project is an AI-powered Fashion Product Recommendation System that suggests visually similar fashion products based on an uploaded image.

The system leverages:

Deep Learning (ResNet50) for image feature extraction
Vector similarity search (KNN / Annoy-ready) for recommendations
AWS S3 for scalable image storage
Flask for backend deployment

It is designed to scale to large image datasets (23GB+) while remaining efficient and maintainable.

Problem Statement

Traditional fashion recommendation systems depend heavily on:

Manual tagging
Text-based filters
User metadata

These approaches fail when:

Metadata is missing or inconsistent
Users want visual similarity, not just category similarity

Solution: Recommend products using visual features extracted by a deep learning model.

System Architecture

User Upload Image ↓ ResNet50 Feature Extraction ↓ Feature Normalization ↓ Similarity Search (KNN / Annoy) ↓ Top-N Similar Fashion Items ↓ Images Served from AWS S3

Download Dataset

python

import kagglehub

# Download latest version
path = kagglehub.dataset_download("paramaggarwal/fashion-product-images-dataset")

print("Path to dataset files:", path)

First train images on Resnet50

python

import tensorflow as tf
import numpy as np
import os
import pickle
from tensorflow.keras.applications.resnet50 import ResNet50, preprocess_input
from tensorflow.keras.layers import GlobalMaxPooling2D
from tensorflow.keras.models import Sequential
from numpy.linalg import norm

# ==============================
# 1. BUILD MODEL
# ==============================

base_model = ResNet50(
    weights='imagenet',
    include_top=False,
    input_shape=(224,224,3)
)
base_model.trainable = False

model = Sequential([
    base_model,
    GlobalMaxPooling2D()
])

# ==============================
# 2. LOAD FILENAMES (FIXED)
# ==============================

IMG_DIR = 'C:/Users/hp/.cache/kagglehub/datasets/paramaggarwal/fashion-product-images-small/versions/1/images/'

image_files = os.listdir(IMG_DIR)

# what we PROCESS (full paths)
image_paths = [os.path.join(IMG_DIR, fname) for fname in image_files]

# what we STORE (portable)
filenames = image_files

print("Total images:", len(filenames))
print("Sample filenames:", filenames[:5])

# ==============================
# 3. FAST tf.data PIPELINE
# ==============================

BATCH_SIZE = 128  # You can try 64 if GPU memory is low

def load_and_preprocess(path):
    img = tf.io.read_file(path)
    img = tf.image.decode_jpeg(img, channels=3)
    img = tf.image.resize(img, (224,224))
    img = preprocess_input(img)
    return img

dataset = tf.data.Dataset.from_tensor_slices(image_paths)
dataset = dataset.map(load_and_preprocess, num_parallel_calls=tf.data.AUTOTUNE)
dataset = dataset.batch(BATCH_SIZE)
dataset = dataset.prefetch(tf.data.AUTOTUNE)

# ==============================
# 4. FEATURE EXTRACTION (FAST)
# ==============================

print("Extracting features...")
features = model.predict(dataset, verbose=1)

# Normalize embeddings
features = features / norm(features, axis=1, keepdims=True)

print("Feature shape:", features.shape)

# ==============================
# 5. SAVE OUTPUTS
# ==============================

pickle.dump(features, open('embeddings.pkl', 'wb'))
pickle.dump(filenames, open('filenames.pkl', 'wb'))

print("✅ Embeddings and filenames saved successfully")

filenames.pkl contains filename and embeddings.pkl contain vector reprentation of images

Cleaning of styles.csv file containing product info

python

import pandas as pd

df = pd.read_csv(r"styles.csv" , on_bad_lines='skip')
df['baseColour'].fillna(df['baseColour'].mode()[0] , inplace=True)
df['season'].fillna(df['season'].mode()[0] , inplace=True)
df['year'].fillna(df['year'].mode()[0] , inplace=True)
df['usage'].fillna(df['usage'].mode()[0] , inplace=True)
df.dropna(subset=['productDisplayName'] , inplace=True)

print(df.head())
print(df.isnull().sum())
print(df.shape)

print(df.columns)

Then create a new folder for project.
Create a folder named model, templates, and a file app.py
In model folder add cleaned styles.csv
I hosted embedding.pkl , filenames.pkl , Resnet50.keras on s3 and then downloading it using code.
In template folder create a html file.
You can get codes in above Fashion Project Folder
So the basic architecture is that we already calculated vectors represntation for each image using Resnet 50.
For new image we again calculate vectors representation using Resnet50
Then we use this vector and predict nearest 6 vectors from whole using K nearest neighbour model usng cosine similarity.
Then we find most nearets 6 vectors filenames and display them.
I also used clous AWS S3 to host images.
Used docker for ease in deployment on AWS EC2
Then Route 53 for domain mapping with EC2 public ip and nginx certbot to make it secure.

Project image

Tech Stack

Deep Learning

TensorFlow / Keras
ResNet50 (ImageNet pretrained)
Global Max Pooling

Machine Learning

KNN (baseline similarity search)
Annoy (for optimized large-scale retrieval)

Backend

Flask
NumPy, Pandas, Pickle

Cloud

AWS S3
IAM (secure access control)

Feature Extraction Pipeline

Input image resized to 224×224
Preprocessed using preprocess_input
Passed through ResNet50 (without top layers)
Global Max Pooling applied
Feature vector normalized using L2 normalization

Cloud Storage with AWS S3

Why AWS S3?

Local hosting of 23GB+ images is impractical
GitHub has strict storage limits
AWS S3 provides:
- Highly durable storage
- Public / private access control
- Cost-efficient scalability

Image Access Strategy

Images are uploaded once to S3
Model returns image filenames
Frontend loads images via S3 public URLs

Example:

https://myfashion-images-recommender.s3.ap-south-1.amazonaws.com/12345.jpg

Dataset used

Fashion product images
Metadata: category, color, season, usage
Dataset link :- https://www.kaggle.com/datasets/paramaggarwal/fashion-product-images-dataset

Contact

Gaohar Imran 📍 India 💼 Aspiring ML / Deep Learning Engineer 🔗 LinkedIn: https://www.linkedin.com/in/gaohar-imran-5a4063379/

imranx654/Fashion-Recommender-System

Fashion Product Recommendation System

Deep Learning + Similarity Search + Cloud Storage

Project Site link

Overview

Problem Statement

System Architecture

Download Dataset

First train images on Resnet50

Cleaning of styles.csv file containing product info

Project image

Tech Stack

Deep Learning

Machine Learning

Backend

Cloud

Feature Extraction Pipeline

Cloud Storage with AWS S3

Why AWS S3?

Image Access Strategy

Dataset used

Contact

imranx654/
Fashion-Recommender-System