MohamedSebaie/Tahaluf-Deep-Learning-Engineer-Task

Tahaluf-Deep-Learning-Engineer-Task

Tahaluf-Deep-Learning-Engineer-Task

• Clothing Article Classifier

• YOLOX

• Weights and Biases Dashboard for YOLOX training on Custom Data

• Wandb Dashboard

• Google Drive Link For The Weights

Clothing Article Classifier

This repository contains the code and resources for a clothing article classifier using machine learning techniques. The classifier is designed to identify and categorize different types of clothing items from high-resolution images.

Table of Contents

1. Introduction
2. Data Handling
3. Model Architecture and Approach
4. Metrics
5. Results and Discussion
6. Computational Complexity
7. Conclusion
8. Future Works
9. Report

Introduction

In this project, we aim to build a clothing article classifier that can accurately identify and categorize different types of clothing items from high-resolution images. This classifier can be used in various applications, such as e-commerce platforms, inventory management systems, and more.

Data Handling

Dataset

We used the Clothing dataset (full, high resolution) from Kaggle, which contains 5,000 images of clothes from 20 different classes. The dataset is released under CC0. Some of the items are labeled "Not sure", "Others", or "Skip" and may contain labeling errors. We removed these three classes and trained our models on the remaining 17 classes, which are imbalanced.

Data Loading

We used Keras for data loading and preprocessing. The data was divided into training, validation, and testing sets, with a suitable split.

Data Preprocessing

The images were resized and normalized and we removed labeled "Not sure", "Others", or "Skip" and may contain labeling errors. We also applied data augmentation techniques, such as horizontal flipping, to increase the training dataset's diversity and improve the classifier's generalization capabilities.

Model Architecture and Approach

We explored four different approaches:

1. Custom CNN without any pre-trained models.
2. ResNet50 with 17 classes.
3. ResNet50 with only the top 10 classes to handle the imbalanced dataset.
4. ConvNeXt tiny feature extraction model from TensorFlow Hub.

After comparing the performance and efficiency of these models, we chose the ConvNeXt tiny feature extraction model from TensorFlow Hub as our final approach, as it performed well on the imbalanced dataset. We then applied 5-fold cross-validation to the ConvNeXt tiny feature extraction model to ensure its highest accuracy and robustness.

Metrics

We selected F1-score as our primary metric since it is suitable for imbalanced datasets. Additionally, we monitored accuracy as a secondary metric.

Results and Discussion

The final ConvNeXt tiny feature extraction model achieved an average weighted F1-score of 88% using 5-fold cross-validation. This model performed well on the imbalanced dataset and demonstrated good generalization capabilities.

We compared the performance of the four approaches, The table includes the Weighted F1 Score for each approach:

Approach	F1 Score
Custom CNN	66
ResNet50_17_Class	61
ResNet50_10_Class	70
ConvNext-T	90

This table provides a clear comparison of the performance metrics for each approach. We then applied 5-fold cross-validation to the ConvNeXt tiny feature extraction model to ensure its highest accuracy and robustness.

The ConvNeXt tiny feature extraction model with 5-fold cross-validation outperformed the other approaches, making it the best choice for this problem despite of the imbalance data which shows a very good results in the imbalance classes. But we should take into our account the performance because in the deployment we have to make optimization between the accuracy and performance.

Computational Complexity

We calculated the number of Floating Point Operations (FLOPS) and Multiply-accumulate operations (MACCs) per layer for the main convolutional and fully connected layers in the three models: Custom CNN and ResNet50. we used this GitHub Repo. The results are presented in the table below the most expensive layers:

Model	Layer Name	FLOPSx 10^6	MACCsx 10^6
Custom CNN	Conv2	9,912.32	4,956.16
ResNet50	ResBlock1_Conv1	248.84	124.42

Total FLOPS and MACCs for each Model:

Model	Total FLOPSx 10^6	Total MACCsx 10^6
Custom CNN	29,891.94	14,945.97
ResNet50	7,735.46	3,867.73
ConvNext-T	4,500.46	2,250.23

Note: The Flops of ConvNext-T couldn't get it from the calculations because the tensorflowHub model is colsed, but in future i will try to revert it to its archiecturte, or use the PyTorch model. So the Flops in the table is from its benchmark

Decreasing FLOPS and MACCs

To decrease the FLOPS and MACCs, we can apply various techniques, such as:

1. Reduce the number of filters or neurons: By reducing the number of filters in convolutional layers or neurons in fully connected layers, we can decrease the number of operations performed, thereby lowering FLOPs and MACCs.

2. Apply network pruning: Network pruning involves removing less important connections or weights from the network based on some criteria, such as their magnitude. By pruning the network, we can reduce its complexity without sacrificing much performance.

3. Reduce input size: Decreasing the spatial dimensions of the input image can lower the number of operations performed in each layer, reducing FLOPs and MACCs.

4. Employ model compression techniques: Techniques like quantization, knowledge distillation, and weight sharing can help reduce FLOPs and MACCs. Quantization reduces the precision of weights and activations, leading to a decrease in computational complexity. Knowledge distillation trains a smaller student network to mimic the behavior of a larger teacher network. Weight sharing involves grouping similar weights together, reducing the number of unique weights and operations.

By applying these techniques may impact the overall performance of the model. It's essential to find a balance between reducing FLOPs and MACCs and maintaining acceptable accuracy and performance for our specific task.

Most Computationally Expensive Layers

I have added a table with the most expensive layer in each model, along with the time it takes to compute.

Model	Most Expensive Layer	Time (ms)
Custom CNN	Batch-Normalization	0.055
ResNet50	ResBlock3_Conv3	0.45
ConvNext-T	ConvNext_Block_FE	0.07

In the table above, the most computationally expensive layers are typically the convolutional layers with the highest number of filters and the largest filter sizes. In the case of ResNet50, the most expensive layers is the ones within the residual blocks.

By focusing on optimizing these layers, we can significantly decrease the overall computational complexity of the models.

Conclusion

In this project, we successfully built a clothing article classifier using machine learning techniques. The classifier achieved satisfactory performance on the imbalanced dataset using the ConvNeXt tiny feature extraction model from TensorFlow Hub with 5-fold cross-validation. We also compared different approaches and demonstrated the advantages of using pre-trained models for feature extraction in handling imbalanced datasets and improving classification performance.

Future Works

• Combine a larger dataset to make the model more generalized and solve the imbalanced classes.
• Apply more data augmentatoin techniques.
• Use light models like EfficientNet to improve the performace and reduce the FLOPs and MACCs.
• Breakdown the ConvNext model to get the most optimized version from it and calculates its RF and use the PyTorch Version.
• Applying Hyperparameter Tuning Optimization Techniques like Bayesian Optimizer.

Report

The attached Report discuss Receptive field and discuss FLOPs and MACCs Results.