Skip to content

Latest commit

 

History

History
106 lines (91 loc) · 2.59 KB

README.md

File metadata and controls

106 lines (91 loc) · 2.59 KB
Raw

eye-center-detector

Build a CNN to locate eye centers in cropped face images using the facial keypoints dataset available on kaggle

Progress Details -

2018/08/10:

CNN1

Neural Network Architecture:
* (3x3) | 32
* (3x3) | 64
* Dense | 128
* Dense | 128
* Dense | 4

Training Set:
* 7033 : Eye Centers Kaggle

Insights Gained:
* Ran for **1000** epochs
* Model showed clear signs of overfitting -
	train_loss 	~ 1.5e-04
	val_loss 	~ 2.5e-03
* Overfitting = The NeuralNet has the ability to learn complex patterns
* We will train CNN_2 now with data augmentation as well

CNN2

Neural Network Architecture:
* (3x3) | 32
* (3x3) | 64
* Dense | 128
* Dense | 128
* Dense | 4

Training Set:
* 7033 : Eye Centers Kaggle
* 7033 : Data Augmented Eye Centers Kaggle

Insights Gained:
* Ran for **1000** epochs
* Model showed clear signs of overfitting -
	train_loss 	~ 7e-05
	val_loss 	~ 5e-02
* Overfitting = The NeuralNet has the ability to learn complex patterns
* We will train CNN_3 now with dropout to reduce val_loss

2018/08/11:

CNN3

Neural Network Architecture
* (3x3) | 32
* Dropout | 0.1
* (3x3) | 64
* Dropout | 0.2
* Flatten
* Dense | 128
* Dropout | 0.3
* Dense | 128
* Dropout | 0.4
* Dense | 4

Training Set
* 7033 : Eye Centers
* 7033 : Eye Centers; Data Augmented (LR Flipped)

Insights Gained:
* Ran for **2000** epochs
* Model still showed signs of overfitting but this time rate of overfitting was slow
	train_loss 	~ 6.2e-03
	val_los 	~ 7.4e-02
* Total params: 4,000,904. Let's try to reduce this in CNN_4 while improving on val_loss

CNN4

Neural Network Architecture:
* (3x3) | 32
* (3x3) | 64
* Flatten
* Dense | 64
* Dense | 64
* Dense | 4

Training Set:
* 7033 : Eye Centers

Insights Gained:
* Ran for **1000** epochs
* Model still overfits
	train_loss 	~ 1.1e-04
	val_loss 	~ 1.9e-03
* Total params: 2,005,768. Even with half params than before the NN overfits. Therefore reducing the width of our FC was a good idea. We shall continue to reduce the width FC in CNN_5 and see if we still can overfit our training data

CNN5

Neural Network Architecture
* (3x3) | 32
* (3x3) | 64
* Flatten
* Dense | 32
* Dense | 32
* Dense | 4

Training Set
* 7033 : Eye Centers

Insights Gained:
* Ran for **2000** epochs
* Model still overfits
	train_loss 	~ 9.7e-05
	val_loss 	~ 2.2e-03
* Total params: 1,011,272. Reducing the params by half the NN can still overfit. Now let's try to train it with data augmentation to bring down the val_loss. We will do this in CNN_6