YogiSnap Models

View Report

This project uses a TensorFlow image classification model to categorise the yoga poses. It classifies images using a Keras Sequential model.

Data

In the training dataset, there are 5 yoga poses with 630 images in total. Yoga poses including downdog (179 images), goddess (125 images), plank (107 images), tree (58 images), and warrior (166 images). The data was saved in a tree structure under the project folder. All data is saved in a folder called “data”, underneath which is a "yoga_poses" folder with five subfolders containing photos of five different yoga poses. This dataset is for training and validation purpose.

Training dataset. Image source: Kaggle

There is another dataset for testing the model which is under folder "sample", with 40 images in total, including downdog (11 images), goddess (6 images), plank (8 images), tree (8 images), and warrior (7 images).

Tasting dataset. Image source: Roehampton University yoga society members

Models

There are two model created in the project including basic model and improved model

• Basic Model

Basic model

Model summary

model = Sequential([ # data normalisation layers.Rescaling(1. / 255, input_shape=(img_height, img_width, 3)), # 1st convolution block layers.Conv2D(16, 3, padding='same', activation='relu'), layers.MaxPooling2D(), # 2nd convolution block layers.Conv2D(32, 3, padding='same', activation='relu'), layers.MaxPooling2D(), # 3rd convolution block layers.Conv2D(64, 3, padding='same', activation='relu'), layers.MaxPooling2D(), # fully-connected layer layers.Flatten(), layers.Dense(128, activation='relu'), layers.Dense(num_classes) ])

Model: "sequential" _________________________________________________________________ Layer (type) Output Shape Param ================================================================= rescaling_1 (Rescaling) (None, 180, 180, 3) 0 conv2d (Conv2D) (None, 180, 180, 16) 448 max_pooling2d (MaxPooling2D) (None, 90, 90, 16) 0 conv2d_1 (Conv2D) (None, 90, 90, 32) 4640 max_pooling2d_1 (MaxPooling2D) (None, 45, 45, 32) 0 conv2d_2 (Conv2D) (None, 45, 45, 64) 18496 max_pooling2d_2 (MaxPooling2D) (None, 22, 22, 64) 0 flatten (Flatten) (None, 30976) 0 dense (Dense) (None, 128) 3965056 dense_1 (Dense) (None, 5) 645 ================================================================= Total params: 3,989,285 Trainable params: 3,989,285 Non-trainable params: 0 _________________________________________________________________

• Improved Model

The improved model added data augmentation to increase the data variety & a dropout layer to address the overfitting.

data_augmentation_layer = keras.Sequential( [ layers.Resizing(img_height, img_width), layers.RandomFlip("horizontal"), layers.RandomBrightness((-0.25, 0.25)), layers.RandomContrast((0.25, 0.25)), layers.RandomTranslation((-0.15, 0.15), (-0.15, 0.15), fill_mode='constant'), layers.RandomRotation((-0.025, 0.025), fill_mode='constant'), # The rotation value is low here, # because too much rotation may make the data dislike the original pose layers.RandomZoom((0, 0.25), fill_mode='constant'), # the default fill_mode parameter is 'reflect', # change it to avoid the augmented data look distortion ] ) augmented_ds = train_ds.map(lambda x, y: (data_augmentation_layer(x), y)) reference-RandomRotation

Improved model

Model summary

model = Sequential([ # Data normalisation layers.InputLayer((img_height, img_width, 3)), layers.Rescaling(1. / 255), # 1st convolutional block layers.Conv2D(16, 3, padding='same', activation='relu'), layers.MaxPooling2D(), # 2nd convolutional block layers.Conv2D(32, 3, padding='same', activation='relu'), layers.MaxPooling2D(), # 3nd convolutional block layers.Conv2D(64, 3, padding='same', activation='relu'), layers.MaxPooling2D(), # 4th convolutional block layers.Conv2D(128, 3, padding='same', activation='relu'), layers.MaxPooling2D(), # 5th convolutional block layers.Conv2D(64, 3, padding='same', activation='relu'), layers.MaxPooling2D(), # 6th convolutional block layers.Conv2D(32, 3, padding='same', activation='relu'), layers.MaxPooling2D(), # Add dropout layers.Dropout(0.3), # Fully-connected layer layers.Flatten(), layers.Dense(128, activation='relu'), layers.Dense(num_classes, name="outputs") ])

Model: "sequential_1" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= rescaling (Rescaling) (None, 180, 180, 3) 0 conv2d (Conv2D) (None, 180, 180, 16) 448 max_pooling2d (MaxPooling2D) (None, 90, 90, 16) 0 conv2d_1 (Conv2D) (None, 90, 90, 32) 4640 max_pooling2d_1 (MaxPooling2D) (None, 45, 45, 32) 0 conv2d_2 (Conv2D) (None, 45, 45, 64) 18496 max_pooling2d_2 (MaxPooling2D) (None, 22, 22, 64) 0 conv2d_3 (Conv2D) (None, 22, 22, 128) 73856 max_pooling2d_3 (MaxPooling2D) (None, 11, 11, 128) 0 conv2d_4 (Conv2D) (None, 11, 11, 64) 73792 max_pooling2d_4 (MaxPooling2D) (None, 5, 5, 64) 0 conv2d_5 (Conv2D) (None, 5, 5, 32) 18464 max_pooling2d_5 (MaxPooling2D) (None, 2, 2, 32) 0 dropout (Dropout) (None, 2, 2, 32) 0 flatten (Flatten) (None, 128) 0 dense (Dense) (None, 128) 16512 outputs (Dense) (None, 5) 645 ================================================================= Total params: 206,853 Trainable params: 206,853 Non-trainable params: 0 _________________________________________________________________

Training Result

Comparing the training result between basic and improved models

Basic model	Improved model
batch_size = 32 --------------- epochs=10 history = model.fit( train_ds, validation_data=val_ds, epochs=epochs )	batch_size = 64 --------------- epochs = 100 history = model.fit( augmented_ds, validation_data=val_ds, epochs=epochs )

• The improved model has about 85% accuracy on testing dataset

Confusion Matrix

The heatmaps are based on the result of improved model (both training and test dataset), learning what pose get confuse with other poses.

Training Dataset	Test Dataset