sumit7488/TimesFormer_Baseline

🎬 TimeSformer Fine-Tuned for Video Action Recognition

This model is a fine-tuned version of TimeSformer (Time-Space Transformer) for video action recognition, trained on the UCF101 dataset.

---

📌 Model Overview

• Base Model: facebook/timesformer-base-finetuned-k400
• Task: Video Classification / Action Recognition
• Dataset: UCF101 (101 action classes)
• Framework: PyTorch + Hugging Face Transformers
• Training Environment: Kaggle (GPU)

---

🧠 Training Strategy

Due to Kaggle’s 12-hour session limit, training was performed in multiple stages:

1. Initial training run
2. Checkpoint saving (best model)
3. Resume training from best checkpoint
4. Further fine-tuning across sessions

This approach ensures efficient long training without losing progress.

---

📊 Training Results

🔹 Initial Training

Epoch	Train Loss	Train Acc	Val Loss	Val Acc
1	4.5066	0.0622	4.1089	0.4245
2	3.5721	0.4711	2.5276	0.8007
3	2.3239	0.7323	1.4321	0.8993

---

🔹 Continued Training (Checkpoint Resume)

Epoch	Train Loss	Train Acc	Val Loss	Val Acc
4	1.8289	0.7991	1.1802	0.9199
5	1.7119	0.8094	1.1372	0.9128
6	1.6365	0.8153	1.1085	0.9191
7	1.5982	0.8139	1.0868	0.9218
8	1.5053	0.8194	1.0763	0.9262
9	1.4673	0.8201	1.0824	0.9225

---

🏆 Best Performance

• Best Validation Accuracy: 92.62%
• F1 Score: 0.9244
• Precision: 0.9315
• Recall: 0.9262
• Achieved at Epoch 8

---

📈 Additional Metrics

Metric	Value
Precision	0.9315
Recall	0.9262
F1 Score	0.9244

---

⚙️ Training Details

• Mixed Precision Training (torch.cuda.amp)
• GPU Memory Usage: ~9.3–9.8 GB
• Training Time per Epoch: ~2.5 hours
• Evaluation Time per Epoch: ~20 minutes
• Best model checkpoint saved automatically

---

🚀 Usage

Install Dependencies

pip install torch torchvision transformers