Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,32 @@
|
|
1 |
---
|
2 |
-
|
|
|
|
|
|
|
|
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
+
language : en
|
3 |
+
tags : cv
|
4 |
+
license : mit
|
5 |
+
dataset : cifar10
|
6 |
+
metrics : accuracy (https://hf.co/metrics/accuracy)
|
7 |
---
|
8 |
+
## Model description
|
9 |
+
|
10 |
+
**Upside down detector**: Model to detect if images are upside down
|
11 |
+
|
12 |
+
* Picked a dataset of natural images - cifar10
|
13 |
+
* Synthetically turned some of images upside down. Created a training and test set.
|
14 |
+
* Trained it to classify image orientation ie if the image is upside down or not.
|
15 |
+
|
16 |
+
## Intended uses & limitations
|
17 |
+
Intended to showcase skill set of being able to train a simple CNN classifier.
|
18 |
+
|
19 |
+
## How to use
|
20 |
+
n/a
|
21 |
+
|
22 |
+
## Limitations and bias
|
23 |
+
Trained on a relatively small dataset, hence it's hard to derive conclusions.
|
24 |
+
|
25 |
+
## Training data
|
26 |
+
cifar10
|
27 |
+
|
28 |
+
## Training procedure
|
29 |
+
Trained using Keras with Nadam classifier with ReduceLROnPlateau which halves the learning rate when the validation loss doesn't improve for 5 iterations
|
30 |
+
|
31 |
+
## Evaluation results
|
32 |
+
The classifier was able to achieve 90% validation accuracy
|