Using the dataset for flowers
The main goal is to improve the average accuracy of ten class.
Need to experiment various hyperparameters
4. Network topology
Number of neurons per layer (for example, 100 x 200 x 100, 200 x 300 x 100 … )
number of layers (For example, 2 vs 3 vs 4 … )
shape of conv2d
5. While experimenting, need to record the performance such that you can create a bar chart of the performance
I have used VGG-Net for classifying the images. I used data-generator to reshape all images to the same size for training.
In VGGNet, they have used 3 layers in each block, I have used only 2 and got better results. Also, I tried multiple optimizers – Nadam, Adam, AdaGrad, SGD. Also, I used Learning Rate Scheduler and Early Stopping.
The main challenge I faced was finding a good reference. I tried various architectures before deciding on VGG net.
I have attached the training and validation graphs for my model.
I have achieved a test accuracy of 78.79%.
Model Details –
Model: "sequential_18"
Layer (type) Output Shape Param #
================================================================= batch_normalization_118 (Ba (16, 224, 224, 3) 12
tchNormalization)
conv2d_94 (Conv2D) (16, 224, 224, 64) 1792 batch_normalization_119 (Ba (16, 224, 224, 64) 256 tchNormalization)
conv2d_95 (Conv2D) (16, 224, 224, 64) 36928 max_pooling2d_64 (MaxPoolin (16, 111, 111, 64) 0 g2D)
dropout_70 (Dropout) (16, 111, 111, 64) 0 conv2d_96 (Conv2D) (16, 111, 111, 128) 73856 batch_normalization_120 (Ba (16, 111, 111, 128) 512 tchNormalization)
conv2d_97 (Conv2D) (16, 111, 111, 128) 147584 batch_normalization_121 (Ba (16, 111, 111, 128) 512 tchNormalization)
max_pooling2d_65 (MaxPoolin (16, 55, 55, 128) 0 g2D)
dropout_71 (Dropout) (16, 55, 55, 128) 0 conv2d_98 (Conv2D) (16, 55, 55, 256) 295168 batch_normalization_122 (Ba (16, 55, 55, 256) 1024 tchNormalization)
conv2d_99 (Conv2D) (16, 55, 55, 256) 590080 batch_normalization_123 (Ba (16, 55, 55, 256) 1024 tchNormalization)
max_pooling2d_66 (MaxPoolin (16, 27, 27, 256) 0 g2D)
dropout_72 (Dropout) (16, 27, 27, 256) 0 conv2d_100 (Conv2D) (16, 27, 27, 512) 1180160 batch_normalization_124 (Ba (16, 27, 27, 512) 2048 tchNormalization)
dropout_73 (Dropout) (16, 27, 27, 512) 0 conv2d_101 (Conv2D) (16, 27, 27, 512) 2359808 batch_normalization_125 (Ba (16, 27, 27, 512) 2048 tchNormalization)
max_pooling2d_67 (MaxPoolin (16, 13, 13, 512) 0
g2D)
dropout_74 (Dropout) (16, 13, 13, 512) 0
flatten_21 (Flatten) (16, 86528) 0
dense_51 (Dense) (16, 1024) 88605696
batch_normalization_126 (Ba (16, 1024) 4096
tchNormalization)
dropout_75 (Dropout) (16, 1024) 0
dense_52 (Dense) (16, 512) 524800
batch_normalization_127 (Ba (16, 512) 2048
tchNormalization)
dropout_76 (Dropout) (16, 512) 0
dense_53 (Dense) (16, 5) 2565
batch_normalization_128 (Ba (16, 5) 20
tchNormalization)
activation_21 (Activation) (16, 5) 0
================================================================= Total params: 93,832,037
Trainable params: 93,825,237
Non-trainable params: 6,800
Reference –
Very Deep Convolutional Networks for Large-Scale Image Recognition - Karen Simonyan, Andrew Zisserman - https://arxiv.org/abs/1409.1556
https://www.kaggle.com/datasets/alxmamaev/flowers-recognition
Comments