CIFAR-100 WideResNets

class deepobs.cifar100.cifar100_wrn.set_up(batch_size, num_residual_units, k, weight_decay, bn_decay)[source]

Class providing the functionality for Wide Residual Networks on CIFAR-100.

The details of the architectures are described in the paper. This test problem is instantiated by the test problems cifar100_wrn404, et cetera.

TensorFlow code is adapted from here.

Parameters:
  • batch_size (int) -- Batch size of the data points. No default value specified.
  • num_residual_units (int) -- Number of residual units in the network. No default value specified.
  • k (int) -- Network width. No default value specified.
  • weight_decay (float) -- Weight decay factor. In this model weight decay is applied to the weights, but not the biases. No default value specified.
  • bn_decay (float) -- Decay factor for the moving average in the batch norm layer. No default value specified.
data_loading

Data loading class for CIFAR-100, cifar100_input.data_loading.

Type:deepobs.data_loading
losses

Tensor of size batch_size containing the individual losses per data point.

Type:tf.Tensor
accuracy

Tensor containing the accuracy of the model.

Type:tf.Tensor
train_init_op

A TensorFlow operation to be performed before starting every training epoch.

Type:tf.Operation
train_eval_init_op

A TensorFlow operation to be performed before starting every training eval epoch.

Type:tf.Operation
test_init_op

A TensorFlow operation to be performed before starting every test evaluation phase.

Type:tf.Operation
batch_norm(x, phase, decay=0.9, name='batch_norm')[source]

Apply batch normalization to tensor x.

Parameters:
  • x (tf.Tensor) -- Input tensor to the batch norm layer.
  • phase (tf.Variable) -- Phase variable switching between train and evaluation mode of the batch norm layer depending on its value ("train", "train_eval", "test").
  • decay (float) -- Decay factor for the moving average in the batch norm layer. Defaults to 0.9.
  • name (str) -- Name for the layer. Defaults to batch_norm.
Returns:

Output after the batch norm layer.
Return type:

tf.Variable
conv(x, filter_size, out_channels, stride, padding='SAME', name='conv')[source]

Apply a convolution to tensor x with a convolution kernel of shape filter_size * filter_size * out_channels, as well as stride and padding as specified. The kernel is created/retrieved via tf.get_variable. No bias is added and no non-linearity is applied.

Parameters:
  • x (tf.Tensor) -- Input tensor to the convolutional layer.
  • filter_size (int) -- Size of the convolution. No default value specified.
  • out_channels (int) -- Number of output channels after the conv layer.
  • stride (int) -- Stride of the convolution. No default value specified.
  • padding (int) -- Padding of the convolution. Can be SAME or VALID. Defaults to SAME.
  • name (str) -- Name of the layer. Defaults to conv.
Returns:

Output after the convolutional layer.
Return type:

tf.Variable
fc(x, out_dim, name='fc')[source]

Apply a affine transformation (fully-connected layer) to tensor x with output dimension out_dim. Weight matrix and bias vector are created/retrieved via tf.get_variable. No non-linearity is applied.

Parameters:
  • x (tf.Tensor) -- Input tensor to the convolutional layer.
  • out_dim (int) -- Number of output dimensions after the fully-connected layer.
  • name (str) -- Name of the layer. Defaults to fc.
Returns:

Output after the fully-connected layer.
Return type:

tf.Variable
get()[source]

Returns the losses and the accuray of the model.

Returns:Tupel consisting of the losses and the accuracy.
Return type:tupel
set_up(num_residual_units, k, weight_decay, bn_decay)[source]

Sets up the test problem.

Parameters:
  • num_residual_units (int) -- Number of residual units in the network. No default value specified.
  • k (int) -- Network width. No default value specified.
  • weight_decay (float) -- Weight decay factor. In this model weight decay is applied to the weights, but not the biases. No default value specified.
  • bn_decay (float) -- Decay factor for the moving average in the batch norm layer. No default value specified.
Returns:

Tupel consisting of the losses and the accuracy.
Return type:

tupel