csx4337-project

Student Name: Soumyendu Sarkar Student ID: X123160 Course: COMPSCI X433.7 Machine Learning With TensorFlow

Project Proposal: Implement DenseNet 121 DenseNet convolutional neural network model and train both from Scratch and with Imagenet dataset pretrained Transfer Learning to classify the different flower species from images of flowers. The Transfer Learned method will reuse the lower convolution layers of the image classifier for their feature extraction capabilities and train a fully connected new classification layer on top to detect different species of flowers. In this project the various hyper parameters like learning rate, batch size, and regularization have been tuned to improve model training.

This project will also include visualization of the training with Tensorboard. Data Source : http://download.tensorflow.org/example_images/flower_photos.tgz This project has been chosen to demonstrate the power of transfer learning and to show how smaller image datasets can be effectively used by convolutional neural network with limited computation to create a highly accurate image classifier.

Requirements

Tensorflow 1.8 Python 3.5+

Install the following Packages : pip3 install pandas pip3 install sklearn pip3 install scipy pip3 install matplotlib pip3 install seaborn pip3 install pyprind pip3 install pillow

Pre-trained Models

The top-1/5 accuracy rates by using single center crop (crop size: 224x224, image size: 256xN)

Network	Top-1	Top-5	Checkpoints
DenseNet 121 (k=32)	74.91	92.19	model

Usage

Step-by-step Example of training on flowers dataset.

Downloading ans pre-processing flowers dataset

DATA_DIR=../data/flowers

python download_and_convert_data.py
--dataset_name=flowers
--dataset_dir="${DATA_DIR}"

#### Training of DenseNet 121 model from scratch.

DATASET_DIR=../data/flowers TRAIN_NTL_DIR=./train_scratch_logs NUM_CLONES=1 python train_image_classifier.py
--train_dir=${TRAIN_NTL_DIR}
--dataset_name=flowers
--dataset_split_name=train
--dataset_dir=${DATASET_DIR}
--num_clones=${NUM_CLONES}
--model_name=densenet121

#### Training of Pre-Trained DenseNet 121 model with Transfer Learning

DATASET_DIR=../data/flowers TRAIN_TL_DIR=./train_transfer_learning_logs NUM_CLONES=1 CHECKPOINT_TL_PATH=./transferlearning/tf-densenet121.ckpt

python train_image_classifier.py
--train_dir=${TRAIN_TL_DIR}
--dataset_name=flowers
--dataset_split_name=train
--dataset_dir=${DATASET_DIR}
--num_clones=${NUM_CLONES}
--model_name=densenet121
--checkpoint_path=${CHECKPOINT_TL_PATH}
--checkpoint_exclude_scopes=global_step,densenet121/logits
--trainable_scopes=densenet121/logits

Validation of Densenet 121 Model Trained from Scratch

DATASET_DIR=../data/flowers EVAL_DIR=./eval_scratch_logs CHECKPOINT_M_PATH=./train_scratch_logs NUM_CLONES=1

python eval_image_classifier.py
--eval_dir=${EVAL_DIR}
--dataset_name=flowers
--dataset_split_name=validation
--dataset_dir=${DATASET_DIR}
--model_name=densenet121
--num_clones=${NUM_CLONES}
--checkpoint_path=${CHECKPOINT_M_PATH}

Training of Pre-Trained DenseNet 121 model with Transfer Learning

DATASET_DIR=../data/flowers EVAL_TL_DIR=./eval_transfer_learning_logs CHECKPOINT_TLM_PATH=./train_transfer_learning_logs NUM_CLONES=1

python eval_image_classifier.py
--eval_dir=${EVAL_TL_DIR}
--dataset_name=flowers
--dataset_split_name=validation
--dataset_dir=${DATASET_DIR}
--model_name=densenet121
--num_clones=${NUM_CLONES}
--checkpoint_path=${CHECKPOINT_TLM_PATH}

Results

Results are all organized under Tensorboard

Tensorboard

Issue the command in the main project directory

tensorboard --logdir=

All folders for Training and Validation with or without Transfer Learning for the DenseNet 121 model are organized under the main folder with self discovery.

GPU Usage

For use with GPUs set the following flags :

clone_on_cpu to False num_clones to the number of GPUs