pszemraj/dinov2-small-film-shot-classifier

dinov2-small: types of film shots

Model description

This model is a fine-tuned version of facebook/dinov2-small on the szymonrucinski/types-of-film-shots dataset. It achieves the following results on the evaluation set:

• Loss: 0.9864
• Accuracy: 0.6259

class labels

The dataset contains the following labels:

"id2label": {
    "0": "ambiguous",
    "1": "closeUp",
    "2": "detail",
    "3": "extremeLongShot",
    "4": "fullShot",
    "5": "longShot",
    "6": "mediumCloseUp",
    "7": "mediumShot"
  },

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 4
• eval_batch_size: 4
• seed: 17480
• gradient_accumulation_steps: 8
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.05
• num_epochs: 12.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
1.6177	0.97	24	1.5501	0.4101
1.3029	1.99	49	1.2448	0.5108
1.1785	2.96	73	1.0556	0.5252
1.2146	3.98	98	1.2316	0.5396
0.8389	4.99	123	1.0235	0.5971
0.7883	5.97	147	0.9960	0.6259
0.7899	6.98	172	1.1354	0.5540
0.663	8.0	197	1.0971	0.5827
0.6013	8.97	221	0.9864	0.6259
0.6276	9.99	246	1.0182	0.6115
0.5196	10.96	270	1.0074	0.6547
0.4761	11.7	288	0.9956	0.6763

Framework versions

• Transformers 4.38.0.dev0
• Pytorch 2.2.0+cu121
• Datasets 2.17.1
• Tokenizers 0.15.2