fine_tuned_image_relevance_model

This model is a fine-tuned version of resnext50_32x4d.fb_swsl_ig1b_ft_in1k on an aggregated dataset of images that were classified as relevant (1.0) or irrelevant (0.0). It achieves the following results on the validation set:

• Loss: 0.1032
• Accuracy: 0.9936

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 4e-06
• train_batch_size: 8
• valid_batch_size: 8
• seed: seed not explicitly set
• optimizer: torch.optim.AdamW(resnet_model.parameters(), lr=lr, eps=0.000001)
• lr_scheduler_type: OneCycleLR
• num_epochs: 6

Training results

Training Loss	Epoch	Validation Loss	Accuracy
0.5536	1	0.3270	0.9856
0.3176	2	0.1720	0.9922
0.1887	3	0.1332	0.9944
0.1280	4	0.1146	0.9938
0.1116	5	0.1236	0.9938
0.1016	6	0.1032	0.9936

Framework versions

• timm 1.0.19
• PyTorch 2.8.0+cpu
• Datasets 4.0.0