tensorflow (Python 2.7 / 3.5)

This repo contains my tensorflow codes

MNIST

• mnist training - python mnist_train.py
  Required training and test/validation data is automatically downloaded by code.
  
• mnist test - python mnist_test.py
  Image input as arg (not yet implemented, arg for input image, change in code for now)

EMOTION

• emotion training - python emo_train.py
  Required training and test/validation data can be downloaded from
  
  • https://inclass.kaggle.com/c/facial-keypoints-detector/data --> ./data/ Image input as arg (not yet implemented, arg for input image, change in code for now)

VGG19-STYLE-TRANSFER FOR IMAGE (Python 3.5, Tensorflow LATEST)

• Download model weights from http://www.vlfeat.org/matconvnet/models/beta16/imagenet-vgg-verydeep-19.mat
  
• VGG19 STYLE TRANSFER can be executed by -
  

python style-transfer-main.py  --style ./imgstyle/StyleImage.jpg --content ./imgcontent/ContentImage.jpg --out ./output/ --epochs 100000 --print-iterations 100 --learning-rate 10

• VGG19 Style Transfer can be also executed by vgg19-style-transfer/Style_Transfer.ipynb IPYTHONG NOTEBOOK
  • Please change values for --epochs and --print-iterations as per your requirement.

VGG19-STYLE-TRANSFER FOR VIDEO (Python 3.5, Tensorflow LATEST)

• https://arxiv.org/abs/1708.04538
  
• Download model weights from http://www.vlfeat.org/matconvnet/models/beta16/imagenet-vgg-verydeep-19.mat
  
• Download COCO training images from http://msvocds.blob.core.windows.net/coco2014/train2014.zip
  
• Trained model to download https://drive.google.com/file/d/1cr3S_1GNLLkAtpJ1Cnf0HJi1aZmMUT8i/view?usp=sharing

* Training

python style-transfer-Noise-train.py --style imgstyle\udnie.jpg --chkpnt chkpnt_udnie_batch_16\ --test-image testImage\Kishwar.jpg --print-iterations 500 --chkpnt-iterations 2000 --epochs 3 --learning-rate 0.001 --batch-size 16 --content D:\train2014\train2014\

* Testing (Image)

python style-transfer-Noise-test.py --chkpnt chkpnt_udnie_batch_16\ --cam-url 255 --in-path testImage\ContentImage.jpg --out-path output\Output.jpg

* Testing (Video) - Using IP Camera

python style-transfer-Noise-test.py --chkpnt chkpnt_udnie_batch_16\ --cam-url http://192.168.0.3:8080/video/

* Testing (Video) - Using Laptop Camera

python style-transfer-Noise-test.py --chkpnt chkpnt_udnie_batch_16\ --cam-url 0

TRIGGER WORD LILLY (Python 3.5, Tensorflow LATEST)

• Code generates required training data from given positive, negative and background samples.
• Total number of 2000 samples are generated
• Model output at 97.91% accuracy: rnn-trigger-word-lilly/keras/raw_data/chime/chime_output.wav
• Trained model: rnn-trigger-word-lilly/keras/raw_data/chime/chimModel_loss_0_0715.h5 [loss - 0.0715]
• MODEL

VOICE CONTROLLED CAR (Python 3.5, Tensorflow LATEST, Keras LATEST)

Youtube video can be found at Youtube

• Step 1: Generate New data using Generate_Voice_Data.py
  

  • Before running / executing file please change
    WAVE_OUTPUT_FILENAME = "data/right/file-right-t-" --> Example for keyword "RIGHT"
  • Please create required directories manually as code will not generate any directory.
    

• Step 2: Update config.ini with correct path and keywords.
  

• Step 3: Run Voice-Controlled-Car.ipynb to start training.
  

• Step 4: Save model after training.
  

• Step 5: Change IP to correct IP of car in Test-Voice-Control-Car-Model.py.

Please note that with current data, model doesn't produce good result.

STYLE-TRANSFER FOR MUSIC (Python 3.5, Tensorflow LATEST) - Expected to be finished 30/06/2018

• NEXT IN QUEUE..

STYLE-TRANSFER FOR AUDIO - VOICE CLONING (Python 3.5, Tensorflow LATEST) - Expected to be finished 30/11/2018

• IN QUEUE