wwoo / tf_box_classify

A simple TensorFlow example for training CNN models using input queues and labelled JPEGs

Classifying boxes using TensorFlow

This is a simple example of how to build a convolutional neural network (CNN) with TensorFlow. It extends existing MNIST convnet examples with TensorFlow input queues for reading training and validation data in common image formats like JPEG. You should be able to easily change the sample code and text files to use your own training data.

Link to YouTube Video

The sample training data and code classifies images of a toy box into four states:

1. Upright (box is upright)
2. Tilted (box is tilted on its side)
3. Open (box is upright and open)
4. Spilled (box is tilted on it's side, open with contents spilled)

115 images for each class is used to train the model.

tf_convnet_test.ipynb is a Jupyter notebook which runs through the training example.

The main code samples are:

tf_convnet_test.py reads the training and validation data from train.txt and valid.txt, then builds and executes the TensorFlow graph. Run it with no arguments:

python tf_convnet_test.py

tf_convnet_export.py does all the above, and also exports the trained model. You'll need to install or build TensorFlow Serving to run this. Run it with no arguments:

python tf_convnet_export.py

tf_convnet_inference.cc serves the trained TensorFlow Serving. You'll need to install or build TensorFlow Serving to compile it. The code is slightly modified from the MNIST inference sample included with the TensorFlow SDK, most notably changing the shape of the input data and number of output classes. Run it by specifying the port and location of the exported model.

tf_convnet_inference --port=9000 location/to/model/dir/

web_server.py is a simple Flask application that takes images using a webcam and sends them to the inference server for classification. It has only been "tested" on Chrome. Run it with no arguments:

python web_server.py

I don't have a pattern for the box, but if you want to build your own, it is

• 50x50x50mm
• made of balsa wood
• has white electrical tape on the sides to hold it together
• has an articulating "lid"
• filled coloured pom poms on the inside
• marked on four sides with a "this side up" decal

tf_box_classify's Issues

Simple classification using Python

hi,

Nice job!.Could you let me know how to build a Python program todo a simple classification by feeding the input image directly?

How to run the script ?

I run python tf_convnet_test.py correctly

but I dont know how should I run tf_convnet_inference.cc or python tf_convnet_export.py I get many errors

grasp001@grasp001:~/tf_box_classify$ python tf_convnet_export.py 
Traceback (most recent call last):
  File "tf_convnet_export.py", line 20, in <module>
    from tensorflow_serving.session_bundle import exporter
ImportError: No module named tensorflow_serving.session_bundle

target 'manifest_proto' not declared in package 'tensorflow/contrib/session_bundle'

I cloned your project and built the model exporter. It was successful, but when I tried building tf_convnet_inference, I'm facing an issue saying the following

no such target '@org_tensorflow//tensorflow/contrib/session_bundle:manifest_proto': target 'manifest_proto' not declared in package 'tensorflow/contrib/session_bundle'

Value error of images

When i run below script then get following error.

$ python2 tf_convnet_test.py

get_image_label_list: read 460 items
Traceback (most recent call last):
File "tf_convnet_test.py", line 286, in
tf.app.run()
File "/home/test/Tensorflow/local/lib/python2.7/site-packages/tensorflow/python/platform/app.py", line 44, in run
_sys.exit(main(_sys.argv[:1] + flags_passthrough))
File "tf_convnet_test.py", line 154, in main
batch_size=TRAIN_BATCH_SIZE, num_epochs=TRAIN_EPOCHS)
File "tf_convnet_test.py", line 61, in inputs
image, label = read_image_from_disk(input_queue)
File "tf_convnet_test.py", line 53, in read_image_from_disk
rgb_image = tf.image.resize_images(rgb_image, IMAGE_SIZE, IMAGE_SIZE)
File "/home/test/Tensorflow/local/lib/python2.7/site-packages/tensorflow/python/ops/image_ops_impl.py", line 659, in resize_images
raise ValueError(''size' must be a 1-D Tensor of 2 elements: '
ValueError: 'size' must be a 1-D Tensor of 2 elements: new_height, new_width