Solutions of projects for students

• Final update: 2019. 05. 20.
• All right reserved @ ModuLabs 2019

Getting Started

Prerequisites

• TensorFlow above version 1.13
• Python 3.6 (recommend Anaconda)
• Python libraries:
  • numpy, matplotblib, pandas
  • PIL, imageio for images
  • fix_yahoo_finance for stock market prediction
• Jupyter notebook
• Ubuntu, OS X and Windows OS

CNN projects

Image Segmentation

Task

• GIANA dataset으로 위내시경 이미지에서 용종을 segmentation 해보자
• Image size: 256으로 변경하여 수행 (baseline code는 image_size: 64)
• 밑에 제시된 여러가지 시도를 해보자
• This code is borrowed from TensorFlow tutorials/Image Segmentation which is made of tf.keras.layers and tf.enable_eager_execution().
• You can see the detail description tutorial link

Dataset

• I use below dataset instead of carvana-image-masking-challenge dataset in TensorFlow Tutorials which is a kaggle competition dataset.
  • carvana-image-masking-challenge dataset: Too large dataset (14GB)
• Gastrointestinal Image ANAlys Challenges (GIANA) Dataset (345MB)
  • Train data: 300 images with RGB channels (bmp format)
  • Train lables: 300 images with 1 channels (bmp format)
  • Image size: 574 x 500

Baseline code

• Dataset: train, test로 split
• Input data shape: (batch_size, 64, 64, 3)
• Output data shape: (batch_size, 64, 64, 1)
• Architecture:
  • 간단한 U-Net 구조
  • tf.keras.layers 사용
• Training
  • tf.data.Dataset 사용
  • tf.GradientTape() 사용 for weight update
• Evaluation
  • MeanIOU: Image Segmentation에서 많이 쓰이는 evaluation measure
  • tf.version 1.12 API: tf.metrics.mean_iou
    • tf.enable_eager_execution()이 작동하지 않음
    • 따라서 예전 방식대로 tf.Session()을 이용하여 작성하거나 아래와 같이 2.0 version으로 작성하여야 함
  • tf.version 2.0 API: tf.keras.metrics.MeanIoU

Try some techniques

• Change model architectures (Custom model)
  • Try another models (DeepLAB, Hourglass, Encoder-Decoder 모델)
• Data augmentation
• Various regularization methods

Image Classification

Task

• Overftting을 피하며, accuracy를 높혀 보자
• Image size: 224 또는 299으로 변경하여 수행 (baseline code는 image_size: 150)
• 밑에 제시된 여러가지 시도를 해보자

Dataset

• Google flower datasets
• 5개의 클래스(daisy, dandelion, roses, sunflowers, tulips)로 이루어진 꽃 이미지 데이터를 분류

Baseline code

• Dataset: train, validation, test로 split
• Input data shape: (batch_size, 150, 150, 3)
• Output data shape: (batch_size, num_classes=5)
• Architecture:
  • Conv2D (x3) - Dense - Softmax
  • tf.keras.layers 사용
• Training
  • model.fit_generator 사용
  • tf.keras.preprocessing.image.ImageDataGenerator 사용 for data augmentation
• Evaluation
  • model.evaluate_generator 사용 for test dataset

Try some techniques

• Change model architectures (Custom model)
  • Or use pretrained models
• Data augmentation
• Various regularization methods

RNN projects

Sentiment Analysis

Task

• 모델 및 하이퍼파라미터들을 바꿔가며 accuracy를 높혀 보자
• 밑에 제시된 여러가지 시도를 해보자
• This code is borrowed from Hvass-Labs/TensorFlow-Tutorials/20_Natural_Language_Processing which is made of tf.keras APIs.
• You can see the detail description or full tutorials: tutorials link

Dataset

• IMDB datasets
• IMDB 영화사이트에서 50000개의 영화평을 가지고 positive/negative인지 구분

Baseline code

• Dataset: train, test로 split (validation data는 model.fit할 때 임의로 배정
• Input data shape: (batch_size, max_sequence_length)
• Output data shape: (batch_size, 1)
• Architecture:
  • Embedding - SimpleRNN - Dense (with Sigmoid)
  • tf.keras.layers 사용
• Training
  • model.fit 사용
• Evaluation
  • model.evaluate 사용 for test dataset

Try some techniques

• Change model architectures (Custom model)
  • Use another cells (LSTM, GRU, etc.)
  • Use dropout layers
  • Change the num_words
  • Change a padding method with 'post'
• Data augmentation (if possible)

Stock Price Prediction

Task

• 모델 및 하이퍼파라미터들을 바꿔가며 accuracy를 높혀 보자
• 밑에 제시된 여러가지 시도를 해보자
• The main flow of this code is referenced in this blog
• LG전자의 주식데이터를 직접 다운 받아서 실제 예측을 해보자
  • Train data: 2017년 1월 1일 ~ 2018년 12월 31일 데이터
  • Test data: 2019년 1월 1일 ~ 2019년 1월 31일 데이터
  • Close price 예측 (baseline은 open price)

Dataset

• Yahoo finance datasets
• 2-3년간 daily stock price 데이터를 이용하여 미래 한달의 주식가격을 예측

Baseline code

• Dataset: train, test로 split
• Input data shape: (batch_size, past_day=60, 1)
• Output data shape: (batch_size, 1)
• Architecture:
  • LSTM - Dense
  • tf.keras.layers 사용
• Training
  • model.fit 사용
• Evaluation
  • model.evaluate 사용 for test dataset

Try some techniques

• Change model architectures (Custom model)
  • Use another cells (LSTM, GRU, etc.)
  • Use dropout layers
  • Change the past_day
• Data augmentation (if possible)
• Try Early stopping
• Use various features (open, high, low, close prices and volume features)

Authors

• Il Gu Yi
• Heedong Yoon