Skip to content

Latest commit

 

History

History
205 lines (144 loc) · 14.5 KB

readme.md

File metadata and controls

205 lines (144 loc) · 14.5 KB

Meesho Visual Taxonomy Data Challenge - Team ML Maverick

Members:

This repository contains the code and methodologies developed for the Meesho Visual Taxonomy Data Challenge. The challenge focused on predicting attributes from supplier-uploaded images as part of cataloging for an e-commerce platform. The task is crucial for ensuring accurate and complete attribute information, especially in the Indian e-commerce context, where incorrect or incomplete details are a common issue.

Challenge Overview

Problem Statement

Have you ever encountered a product on an e-commerce website where the image shows a short-sleeve shirt, but the description states full-sleeve? Predicting such attributes accurately from supplier-uploaded images is critical for any e-commerce platform. Suppliers are required to fill in various attribute details when uploading their products for listing. However, this often results in incorrect or incomplete information.

This challenge motivated the release of a subset of the Meesho product dataset to encourage innovative solutions for attribute prediction in the Indian market.

Dataset Overview

Data Set Overview

Approach and Methodology

Our approach used a deep learning pipeline with the following highlights:

  1. Model Selection:

    • We employed facebook/convnext-base-384-22k-1k as the base feature extractor, pre-trained on ImageNet-22k and fine-tuned on ImageNet-1k for robust feature extraction.

  2. Architecture:

    • Shared feature backbone using ConvNeXt.
    • Attribute-specific classifier heads tailored to each attribute.
    • Integrated spatial and channel attention mechanisms.

  3. Training Strategy:

    • Two-phase training:
      • Initial end-to-end training of the full model.
      • Fine-tuning classifier heads for individual attributes.
    • Utilized combined Cross-Entropy and MMD loss for initial training and Cross-Entropy loss for fine-tuning.
    • Augmented training images (e.g., resizing, flipping) to improve generalization.

  4. Efficiency:

    • The model demonstrated memory-efficient deployment, requiring minimal computational resources.
    • Fast inference times of 9.4–10.4 images per second across categories.

Detailed Report: Report

Bug Found During Training

During the third iteration of training, we encountered a bug in the following code statement:

model.module.load_state_dict(torch.load(BEST_MODEL_FROM_BASE_FIRST_TRAINING, map_location=device), strict=False)

Initially, the strict parameter was set to True, which caused an error while loading the model state. To solve the issue, we set strict=False, allowing mismatched keys between the loaded weights and the model architecture to bypass the error. However, this introduced unexpected changes to the model architecture.

We suspect that these changes may have affected the training process slightly, potentially causing a 1–2% deviation in the final model's performance during the third iteration. This issue was not present in the second iteration of training, which was error-free and robust. Despite this, the overall model performance remained competitive, as evident from the results.

How to Run the code?

Instructions

  1. If you want to set up the custom environment as that of Kaggle used by us, then you can use the Docker file gcr.io/kaggle-gpu-images/python.
  2. Change the input_path and working_path variables according to your environment.
  3. Change the test_c_name variable from the options [c1, c2, c3, c4, c5]. The notebook is configured to train a single model at a time, where c1 represents Men Tshirt, c2 represents Sarees, c3 represents Kurtis, c4 represents Women Tshirt, and c5 represents Women Tops and Tunics.
  4. Change the NUM_EPOCH as per the available GPUs for the initial stage training (T1). By default, it is set to 4. The results attached are for 12 epochs for each category (c1, c2, c3, c4, and c5), based on the 12-hour limit of the Kaggle notebook. Adjust it according to your available GPUs.
  5. Change the NUM_ATTR_EPOCHS as per the available GPUs for the second stage training (T2). By default, it is set to 1. The results attached are for 3 epochs for each category (c1, c2, c3, c4, and c5), based on the 12-hour limit of the Kaggle notebook. Adjust it according to your available GPUs.
  6. You are ready to go.

Training Code

The link to the main notebook for the first half of training: T1 Training Code

The link to the main notebook for the second half of training: T2 Training Code

Infernce Code Including the post-processing

  1. If you want to set up the custom environment as that of Kaggle used by us, then you can use the Docker file gcr.io/kaggle-gpu-images/python.
  2. Download the weights from the google drive (link: https://drive.google.com/drive/folders/1tRd7yV-L63Wv4T_8sogffr410eRfIGmS?usp=sharing).
  3. setup the input_path, working_path accordingly as per the instructions written in the above sections, and make sure that the csv files and model weights are available at that path.
  4. The link to the inference notebook which also includes the post-processing and directly gives output in csv file: Inference Code

Results

Our model achieved a 23rd position on the private leaderboard with a final F1-score of 76.93%, trailing the top-performing team by only 3.27%. Despite this, the solution excelled in deployment-friendly performance, requiring minimal GPU resources while maintaining high inference speeds.

Evaluation Metrics per Attribute as per Validation Dataset

Evaluation:

Category Micro-F1-score Macro-F1-score Harmonic mean of micro and macro f1 score
Men Tshirts 0.972 0.967 0.969
Sarees 0.754 0.534 0.6252
Kurtis 0.929 0.888 0.908
Women Tshirts 0.901 0.766 0.828
Women Tops & Tunics 0.913 0.854 0.882

Detailed metrics for Category 1 (Men Tshirts):

Attribute Val Loss Val Accuracy Weighted F1-score Micro F1-score Macro F1-score Harmonic Mean of F1
attr_1 0.2893 0.9100 0.9102 0.9100 0.9133 0.9116
attr_2 0.0152 0.9984 0.9984 0.9984 0.9984 0.9984
attr_3 0.0518 0.9897 0.9897 0.9897 0.9896 0.9896
attr_4 0.1097 0.9645 0.9645 0.9645 0.9431 0.9537
attr_5 0.0140 0.9983 0.9983 0.9983 0.9919 0.9951
average 0.0962 0.9721 0.9722 0.9721 0.9672 0.9696

Detailed metrics for Category 2 (Sarees):

Attribute Val Loss Val Accuracy Weighted F1-Score Micro F1-Score Macro F1-Score Harmonic Mean of F1
attr_1 0.5766 0.7335 0.7399 0.7335 0.5821 0.6491
attr_2 0.7649 0.7002 0.6942 0.7002 0.6440 0.6709
attr_3 0.2415 0.9021 0.8894 0.9021 0.6299 0.7418
attr_4 0.9826 0.5839 0.5093 0.5839 0.4575 0.5131
attr_5 0.6890 0.7398 0.7202 0.7398 0.5924 0.6580
attr_6 0.3215 0.8940 0.8754 0.8940 0.5969 0.7158
attr_7 0.8439 0.6182 0.5556 0.6182 0.4530 0.5229
attr_8 0.4434 0.8315 0.7804 0.8315 0.3240 0.4663
attr_9 0.9015 0.7052 0.6804 0.7052 0.5938 0.6447
attr_10 0.3686 0.8343 0.7626 0.8343 0.4711 0.6021
average 0.6133 0.7542 0.7207 0.7542 0.5344 0.6184

Detailed metrics for Category 3 (Kurtis):

Attribute Val Loss Val Accuracy Weighted F1-Score Micro F1-Score Macro F1-Score Harmonic Mean of F1
attr 1 0.4018 0.8712 0.8678 0.8712 0.7490 0.8055
attr 2 0.2888 0.9105 0.9101 0.9105 0.9021 0.9063
attr 3 0.4849 0.8471 0.8466 0.8471 0.8395 0.8432
attr 4 0.1976 0.9581 0.9544 0.9581 0.8179 0.8825
attr 5 0.1535 0.9537 0.9527 0.9537 0.9429 0.9483
attr 6 0.2128 0.9349 0.9350 0.9349 0.9349 0.9349
attr 7 0.2329 0.9323 0.9322 0.9323 0.9318 0.9321
attr 8 0.1457 0.9686 0.9672 0.9686 0.9463 0.9573
attr 9 0.0310 0.9925 0.9928 0.9925 0.9340 0.9624
average 0.2387 0.9298 0.9287 0.9298 0.8887 0.9080

Detailed metrics for Category 4 (Women Tshirts):

Attribute Val Loss Val Accuracy Weighted F1-Score Micro F1-Score Macro F1-Score Harmonic Mean of F1
attr 1 0.4894 0.8374 0.8372 0.8374 0.8164 0.8432
attr 2 0.3434 0.8858 0.8664 0.8858 0.6241 0.7323
attr 3 0.3181 0.8841 0.8736 0.8841 0.7815 0.8296
attr 4 0.0911 0.9722 0.9724 0.9722 0.9129 0.9416
attr 5 0.7400 0.7482 0.7449 0.7482 0.7269 0.7374
attr 6 0.1510 0.9638 0.9598 0.9638 0.8620 0.9101
attr 7 0.0406 0.9845 0.9847 0.9845 0.9236 0.9531
attr 8 0.2613 0.9375 0.9073 0.9375 0.4839 0.6383
average 0.3043 0.9016 0.8932 0.9016 0.7664 0.8232

Detailed metrics for Category 5 (Women Tops & Tunics):

Attribute Val Loss Val Accuracy Weighted F1-Score Micro F1-Score Macro F1-Score Harmonic Mean of F1
attr 1 0.4914 0.8711 0.8690 0.8711 0.8347 0.8525
attr 2 0.4670 0.8589 0.8560 0.8589 0.8297 0.8441
attr 3 0.2462 0.9195 0.9192 0.9195 0.9176 0.9185
attr 4 0.3858 0.8802 0.8807 0.8802 0.8596 0.8698
attr 5 0.0818 0.9745 0.9776 0.9745 0.7078 0.8200
attr 6 0.1812 0.9477 0.9462 0.9477 0.8696 0.9070
attr 7 0.2818 0.9235 0.9239 0.9235 0.8383 0.8788
attr 8 0.1994 0.9601 0.9603 0.9601 0.9526 0.9564
attr 9 0.2584 0.9395 0.9386 0.9395 0.9153 0.9272
attr 10 0.5723 0.8603 0.8621 0.8603 0.8219 0.8407
average 0.31653 0.91353 0.91336 0.91353 0.85471 0.8815

Iterative Development

We performed multiple experiments and iterative refinements to achieve these results. All code versions, intermediate approaches, and experimental outcomes are documented in this repository.

Future Work

One of the promising experiments we plan to revisit is Vision-Language Modeling. While this approach demonstrated potential, it required approximately 4 seconds per image during inference, far exceeding the competition's time constraint of 5 microseconds per image. Optimization and efficient scaling of such models remain a key area for future exploration.

Data Preprocessing and Augmentation

  • Addressed missing values and class imbalance by fine-tuning classifiers with all available data.
  • Applied image augmentations dynamically, with 50% probability per image during training.
  • Resized images to 512x512 pixels for enhanced feature extraction.

Conclusion

This repository demonstrates our journey through experimentation, architectural innovations, and deployment-friendly solutions for the Meesho Visual Taxonomy Data Challenge. Feel free to explore the code, contribute, or provide feedback to enhance this project further.