Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
tags:
|
| 4 |
+
- video-classification
|
| 5 |
+
- timesformer
|
| 6 |
+
library_name: transformers
|
| 7 |
+
---
|
| 8 |
+
|
| 9 |
+
# TimeSformer fine-tuned model
|
| 10 |
+
|
| 11 |
+
**Checkpoint:** `timesformer_max_full_even_6.pt`
|
| 12 |
+
|
| 13 |
+
## Training configuration
|
| 14 |
+
- Model: TimeSformer-Base (Kinetics-400 pretrained)
|
| 15 |
+
- Frame sampling: even
|
| 16 |
+
- Data balancing: max_full
|
| 17 |
+
- Unfrozen last layers: 6
|
| 18 |
+
- Task: workout action classification (plank, push-up, squat, russian-twist)
|
| 19 |
+
|
| 20 |
+
## Notes
|
| 21 |
+
This repository provides a raw PyTorch checkpoint.
|
| 22 |
+
To use it, load the checkpoint into a `TimesformerForVideoClassification`
|
| 23 |
+
model with the same configuration used during training.
|