Update README.md
Browse files
README.md
CHANGED
|
@@ -1,7 +1,5 @@
|
|
| 1 |
---
|
| 2 |
license: mit
|
| 3 |
-
datasets:
|
| 4 |
-
- Francesco/uno-deck
|
| 5 |
language:
|
| 6 |
- en
|
| 7 |
metrics:
|
|
@@ -69,12 +67,14 @@ pip install torch transformers torchvision
|
|
| 69 |
```
|
| 70 |
Then go to the files of this HF page and simply use model.py in your python scripts, and that is it.
|
| 71 |
The current model.py will get an image from the internet, perform the processing, then make a heatmap of the attention (what the model cares about) and return its guess and confidence.
|
| 72 |
-
On that animated image, it gets a pretty underestimating confidence of 30%, but that is because the dataset is insanely limited and i had to resort to augmentation, which is a game changer.
|
| 73 |
-
|
| 74 |
## Limitations
|
| 75 |
|
| 76 |
- **Limited Dataset:** The model was trained with a small dataset and may not be robust to unseen variations. (56 cards only, but due to augmentation, it is quite robust.)
|
| 77 |
- **Not that good:** It trained for just about 130 epochs (about 5 minutes), but it still isn't enough, since it hasn't seen that many augmentations
|
| 78 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 79 |
## Conclusion
|
| 80 |
This is a basic attempt at creating an uno card recognizer, and is only made for fun and for learning experience. The model can only recognize one card at a time, and understands wild cards the best because of their ease of style.
|
|
|
|
| 1 |
---
|
| 2 |
license: mit
|
|
|
|
|
|
|
| 3 |
language:
|
| 4 |
- en
|
| 5 |
metrics:
|
|
|
|
| 67 |
```
|
| 68 |
Then go to the files of this HF page and simply use model.py in your python scripts, and that is it.
|
| 69 |
The current model.py will get an image from the internet, perform the processing, then make a heatmap of the attention (what the model cares about) and return its guess and confidence.
|
|
|
|
|
|
|
| 70 |
## Limitations
|
| 71 |
|
| 72 |
- **Limited Dataset:** The model was trained with a small dataset and may not be robust to unseen variations. (56 cards only, but due to augmentation, it is quite robust.)
|
| 73 |
- **Not that good:** It trained for just about 130 epochs (about 5 minutes), but it still isn't enough, since it hasn't seen that many augmentations
|
| 74 |
|
| 75 |
+
Datasets:
|
| 76 |
+
- Natural Images - https://www.kaggle.com/datasets/prasunroy/natural-images - For projecting images onto random backgrounds
|
| 77 |
+
- Uno Cards - https://www.kaggle.com/datasets/vatsalparsaniya/uno-cards - For the cards obviously.
|
| 78 |
+
|
| 79 |
## Conclusion
|
| 80 |
This is a basic attempt at creating an uno card recognizer, and is only made for fun and for learning experience. The model can only recognize one card at a time, and understands wild cards the best because of their ease of style.
|