This repository forms the solution by Anton Drasbæk Schiønning (202008161) to assignment 3 in the course "Visual Analytics" at Aarhus University.
It trains a convelutional neural network, based on VGG16, for the task of classifying garnement types from the Indofasion dataset. This dataset contains over 106,000 images of 15 cloth categories originating from Indian ethnic clothes.
.
├── README.md
├── assign_desc.md <---- original assignment description
├── images <---- the indofashion data (SHOULD BE DOWNLOADED AND PLACED HERE)
│ ├── metadata <---- metadata containing json files with labels for images
│ ├── test <---- 7500 testing images
│ ├── train <---- 91166 training images
│ └── val <---- 7500 validation images
├── out
│ ├── model_card.txt <---- model architecture
│ ├── classification_report.txt <---- classification report for model
│ └── loss.png <---- training/validation loss plot
├── requirements.txt
├── run.sh
└── src
└── main.py <---- script for building, training and testing model.
Due to the size of the dataset, the computational demand is heavy and the usage pipeline has therefore only been tested on cloud computing. Specifically, it has been tested on Ucloud which used Ubuntu v22.10 as the operating system. Please note that venv from Python must also be installed if you wish to run the analysis on Ucloud.
Firstly, the Indofashion dataset should be downloaded from Kaggle here. This dataset is 3 GB so this make take some time.
The downloaded folder must be named images and contain four subdirectories: metadata, train, test and val.
These should be the default names, but please ensure that they are correct and place this folder directly into the root directory (see Repository Tree if in doubt where).
With the folder placed, you can run the full analysis with the bash script as such:
bash run.sh
This achieves the following:
- Sets up a virtual environment
- Installs requirements to this environment
- Builds, trains and tests pretrained CNN model on the data (
main.py) - Deactivates virtual environment
The classification report and training/validation loss plot will be saved to the out directory.
The model was constructed using VGG16 without its top layer to gain embeddings of images. This was fed to a 128 node hidden layer, then to a 64 node layer (both relu activation) and finally the output layer representing the 15 classes. For the full model architecture, please check model_card.txt in the out directory.
The training duration was set for a maximum of 10 epochs, but an early callback function was defined that would restore best weights from earlier if no improvements to the validation loss occured for two consequtive epochs.
As seen from the loss plot, the model training was stopped after 8 epochs, restoring weights from the sixth epoch as those proved the highest.
This final model achieved an overall F1-score of 0.84 on the test data.
| Class | Precision | Recall | F1-Score | Support |
|---|---|---|---|---|
| Blouse | 0.95 | 0.97 | 0.96 | 500 |
| Dhoti_Pants | 0.89 | 0.66 | 0.76 | 500 |
| Dupattas | 0.81 | 0.71 | 0.76 | 500 |
| Gowns | 0.75 | 0.52 | 0.62 | 500 |
| Kurta_men | 0.88 | 0.90 | 0.89 | 500 |
| Leggings_and_salwars | 0.69 | 0.88 | 0.77 | 500 |
| Lehenga | 0.93 | 0.92 | 0.93 | 500 |
| Mojaris_men | 0.87 | 0.88 | 0.88 | 500 |
| Mojaris_women | 0.88 | 0.87 | 0.87 | 500 |
| Nehru_jackets | 0.94 | 0.94 | 0.94 | 500 |
| Palazzos | 0.96 | 0.76 | 0.84 | 500 |
| Petticoats | 0.95 | 0.89 | 0.92 | 500 |
| Saree | 0.84 | 0.94 | 0.89 | 500 |
| Sherwanis | 0.90 | 0.91 | 0.91 | 500 |
| Women_kurta | 0.58 | 0.89 | 0.70 | 500 |
| Accuracy | 0.84 | 7500 | ||
| Macro Avg | 0.85 | 0.84 | 0.84 | 7500 |
| Weighted Avg | 0.85 | 0.84 | 0.84 | 7500 |
The high F1-score reveal that the pretrained CNN was able to fit to the data and identify patterns related to garnment type. This result is just slightly lower than the F1-score of 0.88 for the best model in the original paper on the dataset (Rajput & Aneja, 2021).
Interestingly, blouses were almost identified correctly on all instances with an F1 of 0.96. Contrarily, the model struggled with Women Kurtas. The low precision for this class suggests that there were many false positives, i.e. many garnements were suggested to be Woman Kurtas when they were not. As noted by the Rajput & Aneja (2021), this likely pertains to the fact that Woman Kurtas and Gowns are visually similar.
- Singh Rajput, P., & Aneja, S. (2021). IndoFashion: Apparel Classification for Indian Ethnic Clothes. arXiv e-prints, arXiv-2104.