neuralgym.ops

layers

layers

neuralgym.ops.layers.HWNC_to_NHWC(x, name='HWNC_to_NHWC')

Convert data format from HWNC to NHWC, may be used for re-indexing.

neuralgym.ops.layers.NCHW_to_NHWC(x, name='NCHW_to_NHWC')

Convert data format from NCHW to NHWC.

neuralgym.ops.layers.NHWC_to_HWNC(x, name='NHWC_to_HWNC')

Convert data format from NHWC to HWNC, may be used for re-indexing.

neuralgym.ops.layers.NHWC_to_NCHW(x, name='NHWC_to_NCHW')

Convert data format from NHWC to NCHW.

neuralgym.ops.layers.apply_activation(x, relu, activation_fn, name='activation')

Wrapper for apply activation.

Note activation_fn has higher execution level.

neuralgym.ops.layers.avg_pool(x, ksize=2, stride=2, padding='SAME', name='avg_pool')

Average pooling wrapper.

neuralgym.ops.layers.batch_transformer(U, thetas, out_size, name='BatchSpatialTransformer')

Batch Spatial Transformer Layer

Parameters:
  • U (float) –
  • of inputs [num_batch,height,width,num_channels] (tensor) –
  • thetas (float) –
  • set of transformations for each input [num_batch,num_transforms,6] (a) –
  • out_size (int) –
  • size of the output [out_height,out_width] (the) –
Returns: float
Tensor of size [num_batch*num_transforms,out_height,out_width,num_channels]
neuralgym.ops.layers.bilinear_upsample(x, scale=2)

Bilinear upsample.

Caffe bilinear upsample forked from https://github.com/ppwwyyxx/tensorpack Deterministic bilinearly-upsample the input images.

Parameters:
  • x (tf.Tensor) – a NHWC tensor
  • scale (int) – the upsample factor
Returns:

a NHWC tensor.
Return type:

tf.Tensor
neuralgym.ops.layers.concatenated_relu(x, name='concatenated_relu')

Concatenated relu wrapper.

neuralgym.ops.layers.flatten(x, name='flatten')

Flatten wrapper.

neuralgym.ops.layers.get_variable(name, shape, initializer, weight_decay=0.0, dtype='float', trainable=True, freeze_weights=False)

Simple wrapper for get_variable.

neuralgym.ops.layers.max_pool(x, ksize=2, stride=2, padding='SAME', name='max_pool')

Max pooling wrapper.

neuralgym.ops.layers.moving_average_var(x, decay=0.99, initial_value=0.0, name='moving_average_var')

Moving_average_var.

neuralgym.ops.layers.pixel_flow(x, offset, interpolation='bilinear', name='pixel_flow')

pixel_flow: an operation to reorder pixels according to offsets.

Parameters:
  • x (tf.Tensor) – NHWC
  • offset (tf.Tensor) – NHW2, 2 indicates (h, w) coordinates offset
  • interpolation – bilinear, softmax
  • name – name of module

References

[1] Spatial Transformer Networks: https://arxiv.org/abs/1506.02025 [2] https://github.com/ppwwyyxx/tensorpack

neuralgym.ops.layers.scaled_elu(x, name='scaled_elu')

Scaled elu wrapper.

neuralgym.ops.layers.transformer(U, theta, out_size=None, name='SpatialTransformer')

Spatial Transformer Layer.

Forked from tensorflow/models transformer.

summary_ops

summary ops.

neuralgym.ops.summary_ops.scalar_summary(name, value, sess=None, summary_writer=None, step=None)

Add scalar summary.

In addition to summary tf.Tensor and tf.Variable, this function supports summary of constant values by creating placeholder.

Example usage:

>>> scalar_summary('lr', lr)
Parameters:
  • name – name of summary variable
  • value – numpy or tensorflow tensor
  • summary_writer – if summary writer is provided, write to summary instantly
  • step – if summary writer is provided, write to summary with step
Returns:

None
neuralgym.ops.summary_ops.filters_summary(kernel, rescale=True, name='kernel')

Visualize filters and write to image summary.

Parameters:
  • kernel – kernel tensor
  • rescale – rescale weights to [0, 1]
Returns:

None
neuralgym.ops.summary_ops.images_summary(images, name, max_outs, color_format='BGR')

Summary images.

Note that images should be scaled to [-1, 1] for ‘RGB’ or ‘BGR’, [0, 1] for ‘GREY’.

Parameters:
  • images – images tensor (in NHWC format)
  • name – name of images summary
  • max_outs – max_outputs for images summary
  • color_format – ‘BGR’, ‘RGB’ or ‘GREY’
Returns:

None
neuralgym.ops.summary_ops.gradients_summary(y, x, norm=<function abs>, name='gradients_y_wrt_x')

Summary gradients w.r.t. x.

Sum of norm of \(\nabla_xy\).

Parameters:
  • y – y
  • x – w.r.t x
  • norm – norm function, default is tf.abs
  • name – name of gradients summary
Returns:

None

loss_ops

loss related functions

neuralgym.ops.loss_ops.huber_loss(x, delta=1.0, name='huber_loss')

Huber loss: https://en.wikipedia.org/wiki/Huber_loss.

Deprecated. Please use tensorflow huber loss implementation.

neuralgym.ops.loss_ops.l1_loss(x, y, name='l1_loss')

L1 loss: mean(abs(x-y)).

neuralgym.ops.loss_ops.l2_loss(x, y, name='l2_loss')

L2_loss: mean((x-y) ** 2).

neuralgym.ops.loss_ops.tv_loss(x, name='tv_loss')

tv_loss.

Deprecated. Please use tensorflow total_variation loss implementation.

image_ops

image related ops.

neuralgym.ops.image_ops.np_random_crop(image, shape, align=True)

Random crop.

shape from image.

Parameters:
  • image – numpy image, 2d or 3d
  • shape – (height, width)
Returns:

numpy image
neuralgym.ops.image_ops.np_scale_to_shape(image, shape, align=True)

Scale the image.

The minimum side of height or width will be scaled to or larger than shape.

Parameters:
  • image – numpy image, 2d or 3d
  • shape – (height, width)
Returns:

numpy image