MobilenetV2

My project is using the TensorFlow framework to implement the MobilenetV2 model to classify flower images. Give me a star 🌟 if you like this repo.

Model Architecture

The MobileNetV2 network architecture is shown below. The image below is excerpted from the author's original article

MobileNet V2 still uses depthwise separable convolutions, but its main building block now looks like this:

There are two type of Convolution layers in MobileNet V2 architecture:

• 1x1 Convolution
• 3x3 Depthwise Convolution

These are the two different components in MobileNet V2 model:

Stride 1 Block:

• Input
• 1x1 Convolution with Relu6
• Depthwise Convolution with Relu6
• 1x1 Convolution without any linearity
• Add

Stride 2 Block:

• Input
• 1x1 Convolution with Relu6
• Depthwise Convolution with stride=2 and Relu6
• 1x1 Convolution without any linearity

Author

• Github: Nguyendat-bit
• Email: nduc0231@gmail
• Facebook: Nguyễn Đạt
• Linkedin: Đạt Nguyễn Tiến

I. Set up environment

• Step 1: Make sure you have installed Miniconda. If not yet, see the setup document here

• Step 2: cd into MobilenetV2 and use command line

conda env create -f environment.yml

• Step 3: Run conda environment using the command

conda activate MobilenetV2

II. Set up your dataset

1. Download the data:

• Download dataset here

2. Extract file and put folder train and validation to ./data by using splitfolders

• train folder was used for the training process
• validation folder was used for validating training result after each epoch

This library use ImageDataGenerator API from Tensorflow 2.0 to load images. Make sure you have some understanding of how it works via its document Structure of these folders in ./data

train/
...daisy/
......daisy0.jpg
......daisy1.jpg
...dandelion/
......dandelion0.jpg
......dandelion1.jpg
...roses/
......roses0.jpg
......roses1.jpg
...sunflowers/
......sunflowers0.jpg
......sunflowers1.jpg
...tulips/
......tulips0.jpg
......tulips1.jpg

validation/
...daisy/
......daisy2000.jpg
......daisy2001.jpg
...dandelion/
......dandelion2000.jpg
......dandelion2001.jpg
...roses/
......roses2000.jpg
......roses2001.jpg
...sunflowers/
......sunflowers2000.jpg
......sunflowers2001.jpg
...tulips/
......tulips2000.jpg
......tulips2001.jpg

III. Train your model by running this command line

Review training on colab:

Training script:

python train.py --train-folder ${link_to_train_folder} --valid-folder ${link_to_valid_folder} --classes ${num_classes} --epochs ${epochs}

Example:

python train.py  --train-folder ./data/train --valid-folder ./data/val --classes 5 --epochs 100 

There are some important arguments for the script you should consider when running it:

• train-folder: The folder of training data
• valid-folder: The folder of validation data
• Mobilenetv2-save: Where the model after training saved
• classes: The number of your problem classes.
• batch-size: The batch size of the dataset
• lr: The learning rate
• droppout: The droppout
• label-smoothing: The label smoothing
• expansion: The expansion factor
• image-size: The image size of the dataset
• alpha: Width Multiplier. It was mentioned in the paper on page 4 - Mobilenet V1
• rho: Resolution Multiplier, It was mentioned in the paper on page 4 - Mobilenet V1

IV. Predict Process

If you want to test your single image, please run this code:

python predict.py --test-file ${link_to_test_image}

V. Result and Comparision

My implementation

Epoch 88/90
207/207 [==============================] - 78s 377ms/step - loss: 0.2962 - acc: 0.9082 - val_loss: 0.3822 - val_acc: 0.8726

Epoch 00088: val_acc did not improve from 0.88889
Epoch 89/90
207/207 [==============================] - 78s 376ms/step - loss: 0.2930 - acc: 0.9115 - val_loss: 0.3681 - val_acc: 0.8780

Epoch 00089: val_acc did not improve from 0.88889
Epoch 90/90
207/207 [==============================] - 78s 375ms/step - loss: 0.3002 - acc: 0.9103 - val_loss: 0.3803 - val_acc: 0.8862

Epoch 00090: val_acc did not improve from 0.88889

VI. Feedback

If you meet any issues when using this library, please let us know via the issues submission tab.