Tensorflow docker container with GPU support

Me trying out object detection in various ways..

1 Build docker image

Build the image using regular docker build command.

docker build -t niftitech/tensorflow:gpu . --no-cache=true

2 Run the container

If we don't call /bin/bash the container will exit immediately since there is nothing running.

We share our images and the training folder.

nvidia-docker run -it -v `pwd`/images:/tensorflow/models/research/object_detection/images -v `pwd`/training:/tensorflow/models/research/object_detection/training -v `pwd`/:/shared_with_host -p 6006:6006 niftitech/tensorflow:gpu /bin/bash

3 Get started

When the container is running these commands has to be executed in the models/research folder

protoc object_detection/protos/*.proto --python_out=.

export PYTHONPATH=$PYTHONPATH:pwd:pwd/slim

4 Generate csv files

From inside the object_detection folder run

python xml_to_csv.py

5 Generate TF-record

Edit the (host file) generate_tfrecord.py so that the labels are correct. I edited num_examples to be equal to the total number test.record items.

Copy following file inside the container

cp /shared_with_host/generate_tfrecord.py .

Then run

python generate_tfrecord.py --csv_input=images/train_labels.csv --image_dir=images/train --output_path=train.record

python generate_tfrecord.py --csv_input=images/test_labels.csv --image_dir=images/test --output_path=test.record

6 Train

Update .config file for the model with the correct number of examples.

num_classes: NUMBER_OF_CLASSES
num_examples: NUMBER_OF_TEST_IMAGES

Run this command.

python train.py --logtostderr --train_dir=training/ --pipeline_config_path=training/faster_rcnn_inception_v2_pets.config

Tensorboard

Open another terminal and open a new bash session in the container. Run the following command in the objec detection folder

tensorboard --logdir=training

Evaluate training

python eval.py --logtostderr --pipeline_config_path=training/faster_rcnn_inception_v2_pets.config --checkpoint_dir=training/ --eval_dir=eval/

Open another terminal window and exec bash in container
Run this command in the /tensorflow/models/research/object_detection folder.

tensorboard --logdir=training

5 Export graph

Replace “XXXX” in model.ckpt-XXXX with the highest-numbered .ckpt file in the training folder.

python export_inference_graph.py --input_type image_tensor --pipeline_config_path training/faster_rcnn_inception_v2_pets.config --trained_checkpoint_prefix training/model.ckpt-XXXX --output_directory inference_graph

6 Test detection

Using the file Object_detection_image.py one can try detecting objects.

• Copy a test image to the object_detection folder.
• Open the file Object_detection_image.py with vim.
• Change NUM_CLASSES variable to the desired number.
• Change the image name to the copied image name.
• At the end of the file change the code to the following:

# All the results have been drawn on image. Now display the image.
# cv2.imshow('Object detector', image)

cv2.imwrite('01.png',image)

# Press any key to close the image
# cv2.waitKey(0)

# Clean up
# cv2.destroyAllWindows()

Add the following parameter to visualize_boxes_and_labels_on_image_array method to change how many boxes should be drawn

max_boxes_to_draw=100

• Run the following command:

python Object_detection_image.py

Now a image should be saved to disk. Copy it out of the container and check it out.

7 Save boxes to a file

Add this code to the end of the Object_detection_image.py file.

height, width, channels = image.shape

tempBoxes = np.squeeze(boxes)
tempScores = np.squeeze(scores)
tempClasses = np.squeeze(classes).astype(np.int32)

file = open("results.txt", "w")

file.write("left,right,top,bottom,score,class\n")

for i in range(min(100, tempBoxes.shape[0])):
    if scores is None or tempScores[i] > 0.80:
        box = tuple(tempBoxes[i].tolist())
        ymin, xmin, ymax, xmax = box
        (left, right, top, bottom) = (xmin * width, xmax * width,
                                    ymin * height, ymax * height)

        score = str(int(100*tempScores[i]))
        class_name = category_index[tempClasses[i]]['name']

        file.write(str(left) + "," + str(right) + "," + str(top) + "," + str(bottom) + "," + score  + "," + str(class_name) + "\n")