Vision Transformer (base-sized model)
Vision Transformer (ViT) model trained on the
Chaoyang dataset at resolution 384x384, using a fixed 10% of the training set as the validation set and evaluated on the official test set using the best validation model based on the loss
Augmentation pipeline
To address the issue of class imbalance in our training set, we performed oversampling with repetition.
Specifically, we duplicated the minority classes images until we obtained an even distribution across all classes.
This resulted in a larger training set, but ensured that our model was exposed to an equal number of samples from each class during training.
We verified that this approach did not lead to overfitting or other issues by using a validation set with the original class distribution.
We used the following
Albumentationsaugmentation pipeline for our experiments:
- A.Resize(img_size, img_size),
- A.HorizontalFlip(p=0.5),
- A.VerticalFlip(p=0.5),
- A.RandomRotate90(p=0.5),
- A.RandomResizedCrop(img_size, img_size, scale=(0.5, 1.0), p=0.5),
- ToTensorV2(p=1.0)
This pipeline consists of the following transformations:
- Resize: resizes the image to a fixed size of (img_size, img_size).
- HorizontalFlip: flips the image horizontally with a probability of 0.5.
- VerticalFlip: flips the image vertically with a probability of 0.5.
- RandomRotate90: randomly rotates the image by 90, 180, or 270 degrees with a probability of 0.5.
- RandomResizedCrop: randomly crops and resizes the image to a size between 50% and 100% of the original size, with a probability of 0.5.
- ToTensorV2: converts the image to a PyTorch tensor.
These transformations were chosen to augment the dataset with a variety of geometric transformations, while preserving important visual features.
Results
Our model represents the current state-of-the-art in the field, as it outperforms previous state-of-the-art models proposed in papers with code,
based on the dataset's
reference paper.
The results are summarized in the following table using macro avg metrics.
| Model | Accuracy | F1-Score | Precision | Recall |
|---|
| Baseline | 0.83 | 0.77 | 0.78 | 0.75 |
| Vit-384-finetuned | 0.86 ↑3% | 0.81 ↑4% | 0.82 ↑4% | 0.80 ↑5% |
| Vit-384-from-scratch | 0.78 | 0.74 | 0.74 | 0.74 |
| Vit-224-distilled-resnet50 | 0.74 | 0.00 | 0.00 | 0.00 |
How to use
Here is how to use this model to classify an image of the Chaoyang dataset into one of the 4 classes:
1from transformers import ViTFeatureExtractor, ViTForImageClassification
2from PIL import Image
3import requests
4url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
5image = Image.open(requests.get(url, stream=True).raw)
6feature_extractor = ViTFeatureExtractor.from_pretrained('Snarci/ViT-base-patch16-384-Chaoyang-finetuned')
7model = ViTForImageClassification.from_pretrained('Snarci/ViT-base-patch16-384-Chaoyang-finetuned')
8inputs = feature_extractor(images=image, return_tensors="pt")
9outputs = model(**inputs)
10logits = outputs.logits
11# model predicts one of the 4 Chaoyang classes
12predicted_class_idx = logits.argmax(-1).item()
13print("Predicted class:", model.config.id2label[predicted_class_idx])
Currently, both the feature extractor and model support PyTorch. Tensorflow and JAX/FLAX are coming soon, and the API of ViTFeatureExtractor might change.
Training data
The ViT model was tuned on the
Chaoyang dataset at resolution 384x384, using a fixed 10% of the training set as the validation set
Training procedure
Preprocessing
The exact details of preprocessing of images during training/validation can be found
here.
Images are resized/rescaled to the same resolution 384x384 during training and normalized across the RGB channels with mean (0.5, 0.5, 0.5) and standard deviation (0.5, 0.5, 0.5).
License
This model is provided for non-commercial use only and may not be used in any research or publication without prior written consent from the author.
