/CNN-Image-Detection

Neural Network with TensorFlow using its Python API low level to detect objects

Primary LanguagePythonMIT LicenseMIT

CNN-Image-Detection

Neural Network with TensorFlow (Python API low level)

Requirements

  • Install Docker

  • To support GPUs on Linux, install nvidia-docker

  • Build the docker image via the dockerfile present in ./tensorflow/docker/

docker build directory_path_to_the_dockerfile -t give_a_name_to_your_image

  • Launch your container (src : path to the tensorflow folder to be mounted)

nvidia-docker run --runtime=nvidia -it --mount src="$(pwd)/user/tensorflow",target=/mnt,type=bind -p 8888:8888 -p 6006:6006 -e HOST_PERMS="$(id -u):$(id -g)" --shm-size=1g --ulimit memlock=-1 --ulimit stack=67108864 img_id bash

Neural Network Architecture

  • Example : single layer convolutionnal : name, patch(6x6) x input channels(1) ⇒ output channels(6), weight matrix , bias vector bias, output matrix

    1. input_tensor, 1-deep, X [batch, 224, 224, 3]

    2. layer_conv_1, 6x6x3⇒16 stride 1, W_C1 [6, 6, 3, 16], B_C1 [16], Y_C1 [batch, 28, 28, 16]

    3. layer_conv_2, 5x5x16⇒32 stride 2, W_C2 [5, 5, 16, 32], B_C2 [16], Y_C2 [batch, 28, 28, 16]

    4. layer_conv_3, 5x5x32⇒64 stride 2, W_C3 [4, 4, 32, 64], B_C3 [32], Y_C3 [batch, 14, 14, 32]

    5. layer_conv_4, 4x4x64⇒64 stride 2, W_C4 [4, 4, 64, 64], B_C4 [64], Y_C4 [batch, 7, 7, 64]

    6. layer_conv_5, 4x4x64⇒64 stride 2, W_C5 [4, 4, 64, 64], B_C5 [64], Y_C5 [batch, 7, 7, 64]

    7. layer_conv_6, 4x4x64⇒64 stride 2, W_C6 [4, 4, 64, 64], B_C6 [64], Y_C6 [batch, 7, 7, 64]

    8. layer_reshaped, reshaped to YY (flattened) [batch, 7*7*64]

    9. fully connected layer (relu+dropout), W_D1 [7*7*64, 200], B4 [200], Y_D1 [batch, 200]

    10. fully connected layer (softmax), W_D2 [200, 8], B_D2 [8], Y_D2 [batch, 8]

Personnal Neural Network : Order of execution and description of scripts

Augmentation Data

Allows you to increase the size of the image dataset contained in the furniture folder.
Apply different modifications to the images (rotate, zoom, flip, skew).
The newly generated images are put in an output folder in their respective class.

python3 augmentation_data.py

TFRecord Files

Creates the TFRecord Files, three files containing the three sets of images : * Train * Validation * Test

python3 create_TFRecords.py

Neural Network

Creates the neural network and starts the training.

python3 launch_training.py

Freeze

Freeze the model, remember to put the checkpoint number in --input_checkpoint' instead of `NUMBER.

python3 -m tensorflow.python.tools.freeze_graph  --input_graph=graph.pbtxt  --input_checkpoint=model.ckpt-NUMBER  --input_binary=False --output_graph=graph.pb  --output_node_names=input_tensor,action

Usefull command lines

Run a bash in a container
docker exec -it votre_container_id bash
Tensorboard allows you to follow the training of the model in real time, to be executed from the ikea/tensorflow/model/ folder and consult from your browser localhost:6006
tensorboard --logdir .

Notes

  • Installation of TensorFlow via docker

    • Good for testing

    • Facilitates the automation of the installation of the environment

  • Using TensorFlow 1.5.0