DL_CV

Practiced basic Deeplearning skills, especially in field of Computer Vision, mainly using Keras

Conv_Practice_01

• Basic practice of using Conv2D layer, Pooling
• Creating basic CNN Model
• Preprocessing Fashion MNIST data and model training
• O = (I - F + 2P)/2 + 1 Formula

CIFAR10_Custom

• Visualisation of CIFAR10 dataset
• Data preprocessing
• Recognised the difference of 'sparse_categorical_crossentropy' and 'categorical_crossentropy'
• Tried He Normal instead of the basic weight initialisation of Keras, glorot_uniform
• Tried Batch Normalisation and shuffle

CIFAR10_Custom_CB_GAP_WR

• Tried to use some Callbacks; ReduceLROnPlateau for the dynamic change in Learning rate and EarlyStopping to avoid overfitting
• Tried Global Average Pooling, instead of Flatten. (A stage of Flatten layer -> Classification Dense Layer requires a lot of parameters, and it increases a chance of overfitting and increase of training time)
• Tried l1, l2, l1_l2 weight regularisations per each layer by using tensorflow.keras.regularizers

ImageDataGenerator_Aug

• Practiced some augmentations with ImageDataGenerator
• Including H,V Flips, Rotation, Shift, Zoom, Shear, Bright, Channel Shift, ZCA Whiteing, Normalisation

CIFAR10_Custom_Aug

• Included data augmentations in the previous CIFAR10_Custom to check whether there is any difference in a performance metric

CIFAR10_Pretrained

• Practiced a way of using pretrained model in Keras Framework.

CatnDog_Gen

• Downloaded cat-and-dog dataset from a kaggle dataset
• Practiced how to read the directory and jpeg file names to create their absolute directory and made them as a dataframe
• Practiced flow_from_directory(), flow_from_dataframe()

Albumentations_Aug

• Practiced how to use albumentation to created an augmented image.
• Including Flip, Rotation, ShiftScaleRotation, Compose, Crop, RandomBrightnessContrast, HueSaturationValue, RGBShift, ChannelShuffle, ColorJitter, Gaussian Noise, Cutout, CoarseDropout, CLAHE, Blur, GaussianBlur

CatnDog_Sequence

• Created a Dataset instance that inherits Keras Sequence

AlexNet_Practice

• Created the layer structure of AlexNet

VGG_Practice

• Created the layer structure of VGG16
• Created a function conv_block() which creates the consecutive Conv layers as a block.

Inception_Practice

• Created the layer structure of GoogLeNet
• Created a function inception_module to create the characteristical 'Inception module', including 1x1 Convolution.

ResNet_Practice

• Created the layer structure of ResNet
• Created a function identity_block() to create the characteristical 'Identity block', including shortcuts

CatnDog_Fine_Tuning

• Learned the 'trainable' attribute of layers
• Learned how to freeze the feature extractor parts of a pretrained model and how to unfreeze it after a certain number of epochs

Learning_Rate_Scheduler

• Learned how to build a scheduler function which will be inserted into the LearningRateScheduler callback object
• Learned Step Decay, Cosine Decay, Cosine Annealing, and Ramp Up and Step Down Decay