Update README.md
Browse files
README.md
CHANGED
|
@@ -52,7 +52,7 @@ Training a model like this with just a few card images is challenging. Here's ho
|
|
| 52 |
- **Preprocessing:** The images are processed using the `CLIPProcessor` to convert them into the input format that the CLIP model expects, specifically pixel values.
|
| 53 |
- **Optimizer:** AdamW is used to train the network and the cross entropy loss (`nn.CrossEntropyLoss`) for classification is used.
|
| 54 |
- **Training:**
|
| 55 |
-
- The model is trained for 50 epochs. (Reliable enough to counter tons of different states of uno cards)
|
| 56 |
- We use a batch size of 4.
|
| 57 |
- A low learning rate of 1e-5 is used to avoid instability. (Super important! While it is low, it learns extremely well)
|
| 58 |
|
|
|
|
| 52 |
- **Preprocessing:** The images are processed using the `CLIPProcessor` to convert them into the input format that the CLIP model expects, specifically pixel values.
|
| 53 |
- **Optimizer:** AdamW is used to train the network and the cross entropy loss (`nn.CrossEntropyLoss`) for classification is used.
|
| 54 |
- **Training:**
|
| 55 |
+
- The model is generally trained for 50 epochs. (Reliable enough to counter tons of different states of uno cards) But this one has trained for 130 epochs.
|
| 56 |
- We use a batch size of 4.
|
| 57 |
- A low learning rate of 1e-5 is used to avoid instability. (Super important! While it is low, it learns extremely well)
|
| 58 |
|