Emotion Classification

This model is a fine-tuned version of google/vit-base-patch16-224-in21k on the FastJobs/Visual_Emotional_Analysis dataset.

In theory, the accuracy for a random guess on this dataset is 0.125 (8 labels and you need to choose one).

It achieves the following results on the evaluation set:

• Loss: 1.0511
• Accuracy: 0.6687
• Precision: 0.7104
• F1: 0.6713

Model description

The Vision Transformer base version trained on ImageNet-21K released by Google. Further details can be found on their repo.

Training and evaluation data

Data Split

Trained on FastJobs/Visual_Emotional_Analysis dataset. Used a 4:1 ratio for training and development sets and a random seed of 42. Also used a seed of 42 for batching the data, completely unrelated lol.

Pre-processing Augmentation

The main pre-processing phase for both training and evaluation includes:

• Bilinear interpolation to resize the image to (224, 224, 3) because it uses ImageNet images to train the original model
• Normalizing images using a mean and standard deviation of [0.5, 0.5, 0.5] just like the original model

Other than the aforementioned pre-processing, the training set was augmented using:

• Random horizontal & vertical flip
• Color jitter
• Random resized crop

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 64
• eval_batch_size: 64
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine_with_restarts
• lr_scheduler_warmup_steps: 150
• num_epochs: 300

Training results

Framework versions

• Transformers 4.33.0
• Pytorch 2.0.0
• Datasets 2.1.0
• Tokenizers 0.13.3