MobilenetV1

My project is using the TensorFlow framework to implement the MobilenetV1 model to classify flower images. Give me a star if you like this repo.

Model Architecture

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

We see that the model has 30 layers with following characteristics:

• Layer 1: Convolution layer with stride size 2

• Layer 2: Depthwise layer

• Layer 3: Pointwise layer

• Layer 4: Depthwise layer with stride size 2 (different from layer 2, layer 2 dw has stride size 1

• Layer 5: Pointwise layer

• Layer 30: Softmax, used for classification

Depthwise Separable Concolution

The improvement point of the model is to use a convolution method called Depthwise Separable Convolution to reduce the model size and reduce the computational complexity.

Depthwise separable convolution is a depthwise convolution followed by a pointwise convolution as follows:

A small note about the architecture here, is that after each convolution MobileNet will use Batch Normalization (BN) and ReLU as shown below:

Standard Convolution on the left, Depthwise separable convolution with BN and ReLU on the right

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 MobilenetV1 and use command line

conda env create -f environment.yml

• Step 3: Run conda environment using the command

conda activate MobilenetV1

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
• Mobilenetv1-folder: 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
• image-size: The image size of the dataset
• alpha: Width Multiplier. It was mentioned in the paper on page 4
• rho: Resolution Multiplier, It was mentioned in the paper on page 4

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 00097: val_acc improved from 0.87534 to 0.87805, saving model to MobilenetV1
Epoch 98/100
207/207 [==============================] - 46s 220ms/step - loss: 0.2158 - acc: 0.9421 - val_loss: 0.4410 - val_acc: 0.8862

Epoch 00098: val_acc improved from 0.87805 to 0.88618, saving model to MobilenetV1
Epoch 99/100
207/207 [==============================] - 45s 217ms/step - loss: 0.1981 - acc: 0.9488 - val_loss: 0.4763 - val_acc: 0.8753

Epoch 00099: val_acc did not improve from 0.88618
Epoch 100/100
207/207 [==============================] - 45s 218ms/step - loss: 0.2038 - acc: 0.9470 - val_loss: 0.4322 - val_acc: 0.8726

VI. Feedback

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