# VGGNet on Flowers Recognition

### 1. Introduction

In this problem we will implement the VGGNet with 16 layers to classify five different species of flowers:
- **[VGGNet][VGG from Scratch]**: Construct a new network from scratch.
- **[VGGNet (Pretrained=False)][VGG (pretrained = False)]**: Import vgg16 from torchvision.models, set the "pretrained" parameter to "False".
- **[VGGNet (Pretrained=True)][VGG (pretrained = True)]**: Import vgg16 from torchvision.models, set the "pretrained" parameter to "True".

[VGG from Scratch]: https://www.kaggle.com/hoangnguyen111/flowers-recognition-with-vggnet?scriptVersionId=43485908
[VGG (pretrained = False)]: https://www.kaggle.com/johnlee111/flowers-recognition-with-vggnet-pretrained-false?scriptVersionId=43485856
[VGG (pretrained = True)]: https://www.kaggle.com/hoangnguyen111/flowers-reconigtion-with-vggnet-pretrained-true?scriptVersionId=43551761

### 2. Dataset
- The [dataset][data] contains **4242 images** of flowers.
The data collection is based on the data flicr, google images, yandex images.
- The images are divided into **five classes**: chamomile, tulip, rose, sunflower, dandelion.
For each class there are about 800 photos. Photos are not high resolution, about 320x240 pixels. Photos are not reduced to a single size, they have different proportions!

[data]: https://www.kaggle.com/alxmamaev/flowers-recognition#

*Some examples in the dataset*

### 3. Data Preprocessing
- The data is resized to 256x256 pixels for better performance. Then, each is scaled from range [0,255] to [0,1].
- Data Augmentation:

> transforms.RandomHorizontalFlip(),
transforms.RandomVerticalFlip(),
transforms.RandomRotation((30,60), resample=PIL.Image.BILINEAR),
transforms.RandomRotation((-60,-30), resample=PIL.Image.BILINEAR),
transforms.RandomAffine(degrees=0, translate=(0.1,0.1), scale=None, shear=None, resample=False, fillcolor=0),
transforms.RandomPerspective(distortion_scale=0.2, p=0.5, interpolation=3, fill=0)

*Some examples after augmentation*

- Finally, we normalize the images with these set of parameters from the original model on the ImageNet dataset:
> mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])]

### 4. Architecture
- All nets contain Batch Normalization after each ConV layer.
- In addition, we add another Average Pooling between the ConV Layer and FC Layer.


*The architecture of VGG16-Net*

### 5. Parameters

Train Test Parameters:

> batch_size = 32
valid_split = 0.3
shuffle_dataset = True
random_seed = 42

Loss function and Optimizer:

> criterion = nn.CrossEntropyLoss()
optim = Adam(vgg.parameters(), lr, weight_decay)

| | VGGNet | VGGNet (Pretrained=False) | VGGNet (Pretrained=True) |
|:------------:|:------:|:-------------------------:|:------------------------:|
| lr | 3e-5 | 3e-5 | 1e-5 |
| weight_decay | 1e-3 | 1e-3 | 1e-5 |

### 6. Result
- Max Accuracy on the validation set from three distinct models:

| VGGNet | VGGNet (Pretrained=False) | VGGNet (Pretrained=True) |
|:------:|:-------------------------:|:------------------------:|
| 0.862 | 0.848 | 0.909 |

- Result after 150 epochs on the training set and validation set with three models: