darshanz/occupation-prediction

This model is ViT base patch16. Which is pretrained on imagenet dataset, then trained on our custom dataset which is based on occupation prediction. This dataset contains facial images of Indian people which are labeled by occupation. This model predicts the occupation of a person from the facial image of a person. This model categorizes input facial images into 5 classes: Anchor, Athlete, Doctor, Professor, and Farmer. This model gives an accuracy of 84.43%.

Training hyperparameters

The following hyperparameters were used during training:

• optimizer: {'inner_optimizer': {'class_name': 'AdamWeightDecay', 'config': {'name': 'AdamWeightDecay', 'learning_rate': {'class_name': 'PolynomialDecay', 'config': {'initial_learning_rate': 0.0001, 'decay_steps': 70, 'end_learning_rate': 0.0, 'power': 1.0, 'cycle': False, 'name': None}}, 'decay': 0.0, 'beta_1': 0.9, 'beta_2': 0.999, 'epsilon': 1e-08, 'amsgrad': False, 'weight_decay_rate': 0.4}}, 'dynamic': True, 'initial_scale': 32768.0, 'dynamic_growth_steps': 2000}
• training_precision: mixed_float16

Training results

Train Loss	Train Accuracy	Train Top-3-accuracy	Validation Loss	Validation Accuracy	Validation Top-3-accuracy	Epoch
1.0840	0.6156	0.8813	0.6843	0.75	0.9700	0
0.4686	0.8406	0.9875	0.5345	0.8100	0.9867	1
0.2600	0.9312	0.9953	0.4805	0.8333	0.9800	2
0.1515	0.9609	0.9969	0.5071	0.8267	0.9733	3
0.0746	0.9875	1.0	0.4853	0.8500	0.9833	4
0.0468	0.9953	1.0	0.5006	0.8433	0.9733	5
0.0378	0.9953	1.0	0.4967	0.8433	0.9800	6

Framework versions

• Transformers 4.18.0
• TensorFlow 2.8.0
• Tokenizers 0.12.1