tf_unet Package

unet Module

Created on Jul 28, 2016

author: jakeret

class tf_unet.unet.Trainer(net, batch_size=1, verification_batch_size=4, norm_grads=False, optimizer=u'momentum', opt_kwargs={})[source]

Bases: object

Trains a unet instance

  • net – the unet instance to train
  • batch_size – size of training batch
  • verification_batch_size – size of verification batch
  • norm_grads – (optional) true if normalized gradients should be added to the summaries
  • optimizer – (optional) name of the optimizer to use (momentum or adam)
  • opt_kwargs – (optional) kwargs passed to the learning rate (momentum opt) and to the optimizer
output_epoch_stats(epoch, total_loss, training_iters, lr)[source]
output_minibatch_stats(sess, summary_writer, step, batch_x, batch_y)[source]
store_prediction(sess, batch_x, batch_y, name)[source]
train(data_provider, output_path, training_iters=10, epochs=100, dropout=0.75, display_step=1, restore=False, write_graph=False, prediction_path=u'prediction')[source]

Lauches the training process

  • data_provider – callable returning training and verification data
  • output_path – path where to store checkpoints
  • training_iters – number of training mini batch iteration
  • epochs – number of epochs
  • dropout – dropout probability
  • display_step – number of steps till outputting stats
  • restore – Flag if previous model should be restored
  • write_graph – Flag if the computation graph should be written as protobuf file to the output path
  • prediction_path – path where to save predictions on each epoch
class tf_unet.unet.Unet(channels, n_class, cost=u'cross_entropy', cost_kwargs={}, **kwargs)[source]

Bases: object

A unet implementation

  • channels – number of channels in the input image
  • n_class – number of output labels
  • cost – (optional) name of the cost function. Default is ‘cross_entropy’
  • cost_kwargs – (optional) kwargs passed to the cost function. See Unet._get_cost for more options
predict(model_path, x_test)[source]

Uses the model to create a prediction for the given data

  • model_path – path to the model checkpoint to restore
  • x_test – Data to predict on. Shape [n, nx, ny, channels]
Returns prediction:

The unet prediction Shape [n, px, py, labels] (px=nx-self.offset/2)

restore(sess, model_path)[source]

Restores a session from a checkpoint

  • sess – current session instance
  • model_path – path to file system checkpoint location
save(sess, model_path)[source]

Saves the current session to a checkpoint

  • sess – current session
  • model_path – path to file system location
tf_unet.unet.create_conv_net(x, keep_prob, channels, n_class, layers=3, features_root=16, filter_size=3, pool_size=2, summaries=True)[source]

Creates a new convolutional unet for the given parametrization.

  • x – input tensor, shape [?,nx,ny,channels]
  • keep_prob – dropout probability tensor
  • channels – number of channels in the input image
  • n_class – number of output labels
  • layers – number of layers in the net
  • features_root – number of features in the first layer
  • filter_size – size of the convolution filter
  • pool_size – size of the max pooling operation
  • summaries – Flag if summaries should be created
tf_unet.unet.error_rate(predictions, labels)[source]

Return the error rate based on dense predictions and 1-hot labels.

tf_unet.unet.get_image_summary(img, idx=0)[source]

Make an image summary for 4d tensor image with index idx

image_util Module

author: jakeret

class tf_unet.image_util.BaseDataProvider(a_min=None, a_max=None)[source]

Bases: object

Abstract base class for DataProvider implementation. Subclasses have to overwrite the _next_data method that load the next data and label array. This implementation automatically clips the data with the given min/max and normalizes the values to (0,1]. To change this behavoir the _process_data method can be overwritten. To enable some post processing such as data augmentation the _post_process method can be overwritten.

  • a_min – (optional) min value used for clipping
  • a_max – (optional) max value used for clipping
channels = 1
n_class = 2
class tf_unet.image_util.ImageDataProvider(search_path, a_min=None, a_max=None, data_suffix=u'.tif', mask_suffix=u'_mask.tif', shuffle_data=True)[source]

Bases: tf_unet.image_util.BaseDataProvider

Generic data provider for images, supports gray scale and colored images. Assumes that the data images and label images are stored in the same folder and that the labels have a different file suffix e.g. ‘train/fish_1.tif’ and ‘train/fish_1_mask.tif’ Number of pixels in x and y of the images and masks should be even.

Usage: data_provider = ImageDataProvider(“..fishes/train/*.tif”)

  • search_path – a glob search pattern to find all data and label images
  • a_min – (optional) min value used for clipping
  • a_max – (optional) max value used for clipping
  • data_suffix – suffix pattern for the data images. Default ‘.tif’
  • mask_suffix – suffix pattern for the label images. Default ‘_mask.tif’
  • shuffle_data – if the order of the loaded file path should be randomized. Default ‘True’
class tf_unet.image_util.SimpleDataProvider(data, label, a_min=None, a_max=None)[source]

Bases: tf_unet.image_util.BaseDataProvider

A simple data provider for numpy arrays. Assumes that the data and label are numpy array with the dimensions data [n, X, Y, channels], label [n, X, Y, classes]. Where n is the number of images, X, Y the size of the image.

  • data – data numpy array. Shape=[n, X, Y, channels]
  • label – label numpy array. Shape=[n, X, Y, classes]
  • a_min – (optional) min value used for clipping
  • a_max – (optional) max value used for clipping

util Module

Created on Aug 10, 2016

author: jakeret

tf_unet.util.combine_img_prediction(data, gt, pred)[source]

Combines the data, grouth thruth and the prediction into one rgb image

  • data – the data tensor
  • gt – the ground thruth tensor
  • pred – the prediction tensor
Returns img:

the concatenated rgb image

tf_unet.util.create_training_path(output_path, prefix=u'run_')[source]

Enumerates a new path using the prefix under the given output_path :param output_path: the root path :param prefix: (optional) defaults to run_ :return: the generated path as string in form output_path/prefix_ + <number>

tf_unet.util.crop_to_shape(data, shape)[source]

Crops the array to the given image shape by removing the border (expects a tensor of shape [batches, nx, ny, channels].

  • data – the array to crop
  • shape – the target shape
tf_unet.util.plot_prediction(x_test, y_test, prediction, save=False)[source]
tf_unet.util.save_image(img, path)[source]

Writes the image to disk

  • img – the rgb image to save
  • path – the target path

Converts the given array into a RGB image. If the number of channels is not 3 the array is tiled such that it has 3 channels. Finally, the values are rescaled to [0,255)

Parameters:img – the array to convert [nx, ny, channels]
Returns img:the rgb image [nx, ny, 3]

layers Module

Created on Aug 19, 2016

author: jakeret

tf_unet.layers.bias_variable(shape, name=u'bias')[source]
tf_unet.layers.conv2d(x, W, b, keep_prob_)[source]
tf_unet.layers.crop_and_concat(x1, x2)[source]
tf_unet.layers.cross_entropy(y_, output_map)[source]
tf_unet.layers.deconv2d(x, W, stride)[source]
tf_unet.layers.max_pool(x, n)[source]
tf_unet.layers.weight_variable(shape, stddev=0.1, name=u'weight')[source]
tf_unet.layers.weight_variable_devonc(shape, stddev=0.1, name=u'weight_devonc')[source]