Kiffaz11/Videomae_v2_base-hmdb51-finetuned

VideoMAE-v2 (Base) — HMDB51 (51 classes) Fine-tuned Checkpoint

This repository contains a fine-tuned PyTorch checkpoint exported from Kaggle for reuse and future fine-tuning.

Contents

• best_videomaev2_base_51cls.pt — raw checkpoint (torch.load)
• checkpoint_meta.json — metadata (best epoch, best val accuracy, etc.)
• training_results.csv — epoch-level train/val accuracy
• label2id.json / id2label.json — label maps (if available)

Training results

• Best validation accuracy: 0.8626 (epoch 10)

Epoch	Train Acc	Val Acc	Saved Best
1	0.2614	0.7125	✅
2	0.5320	0.8147	✅
3	0.6024	0.8275	✅
4	0.6108	0.8259
5	0.6350	0.8450	✅
6	0.6444	0.8530	✅
7	0.6387	0.8530
8	0.6328	0.8578	✅
9	0.6583	0.8610	✅
10	0.6664	0.8626	✅

Download & load

from huggingface_hub import hf_hub_download
import torch

repo_id = "Kiffaz11/Videomae_v2_base-hmdb51-finetuned"
ckpt_path = hf_hub_download(repo_id=repo_id, filename="best_videomaev2_base_51cls.pt")
ckpt = torch.load(ckpt_path, map_location="cpu")

Notes

• This repo stores a raw PyTorch checkpoint (not necessarily a save_pretrained() Transformers folder).
• Use the included metadata + label maps to reconstruct the training/inference code you used in Kaggle.