tensorpack.models package¶
Relevant tutorials: Symbolic Layers.

tensorpack.models.
BatchNorm
(variable_scope_name, inputs, axis=None, *, training=None, momentum=0.9, epsilon=1e05, center=True, scale=True, beta_initializer=<tf.python.ops.init_ops.Zeros object>, gamma_initializer=<tf.python.ops.init_ops.Ones object>, virtual_batch_size=None, data_format='channels_last', ema_update='default', sync_statistics=None)[source]¶ A more powerful version of tf.layers.batch_normalization. It differs from the offical one in the following aspects:
Accepts an alternative
data_format
option whenaxis
is None. For 2D input, this argument will be ignored.Default value for
momentum
andepsilon
is different.Default value for
training
is automatically obtained from tensorpack’sTowerContext
. Userprovided value can overwrite this behavior.Support the
ema_update
option, which covers broader use cases than the standard EMA update.Support the
sync_statistics
option, which implements “SyncBN” and is very useful in smallbatch models.Better support of the
virtual_batch_size
option that does not have the bugs intf.layers
.
 Parameters
training (bool) – if True, use perbatch statistics to normalize. Otherwise, use stored EMA to normalize. By default, it is equal to get_current_tower_context().is_training. This is not a good argument name, but it is what the Tensorflow layer uses.
virtual_batch_size (int) –
implement “Ghost BatchNorm” that normalizes the data with a smaller batch size than the input. Only effective when training is True. The value has to be a divisor of the actual batch size.
It does not use the buggy TensorFlow implementation which has the problems of (1) wrong behavior at inference; (2) create variables with unnecessary size=1 dimensions. Corresponding TF issue: https://github.com/tensorflow/tensorflow/issues/23050
ema_update (str) –
Only effective when
training=True
. It has the following options:”default”: same as “collection”. Because this is the default behavior in TensorFlow.
”skip”: do not update EMA. This can be useful when you reuse a batch norm layer in several places but do not want them to all update your EMA.
”collection”: Add EMA update ops to collection tf.GraphKeys.UPDATE_OPS in the first training tower. The ops in the collection will be run automatically by the callback
RunUpdateOps
, along with your training iterations. This can waste compute if your training iterations do not always depend on the BatchNorm layer.”internal”: EMA is updated in the first training tower inside this layer itself by control dependencies. In standard scenarios, it has similar speed to “collection”. But it supports more scenarios:
BatchNorm is used inside dynamic control flow. The collectionbased update does not support dynamic control flows.
BatchNorm layer is sometimes unused (e.g., in GANs you have two networks to train alternatively). Putting all update ops into a single collection will waste a lot of compute.
Other part of the model relies on the “updated” EMA. The collectionbased method does not update EMA immediately.
It has less chance to cause TensorFlow bugs in a graph with complicated control flow.
Therefore this option is preferred over TensorFlow default. Corresponding TF issue: https://github.com/tensorflow/tensorflow/issues/14699
sync_statistics (str or None) –
one of None, “nccl”, or “horovod”. It determines how to compute the “perbatch statistics” when
training==True
.None: it uses statistics of the input tensor to normalize during training. This is the standard way BatchNorm was implemented in most frameworks.
”nccl”: this layer must be used under tensorpack’s multiGPU trainers. It uses the aggregated statistics of the whole batch (across all GPUs) to normalize.
”horovod”: this layer must be used under tensorpack’s
HorovodTrainer
. It uses the aggregated statistics of the whole batch (across all MPI ranks) to normalize. Note that on a single machine this is found to be slower than the “nccl” implementation.
When not None, each GPU computes its own E[x] and E[x^2], which are then averaged among all GPUs to compute global mean & variance. Therefore each GPU needs to have the same batch size.
The synchronization is based on the current variable scope + the name of the layer (BatchNorm(‘name’, input)). Therefore, you need to make sure that:
The BatchNorm layer on different GPUs needs to have the same name, so that statistics can be synchronized. If names do not match, this layer will hang.
A BatchNorm layer cannot be reused within one tower.
A BatchNorm layer needs to be executed for the same number of times by all GPUs. If different GPUs execute one BatchNorm layer for different number of times (e.g., if some GPUs do not execute it), this layer may hang.
This option is also known as “SyncBN” or “CrossGPU BatchNorm” as mentioned in: MegDet: A Large MiniBatch Object Detector. Corresponding TF issue: https://github.com/tensorflow/tensorflow/issues/18222.
When sync_statistics is enabled, ema_update is set to “internal” automatically. This is to avoid running UPDATE_OPS, which requires synchronization.
Variable Names:
beta
: the bias term. Will be zeroinited by default.gamma
: the scale term. Will be oneinited by default.mean/EMA
: the moving average of mean.variance/EMA
: the moving average of variance.
Note
This layer is more flexible than the standard “BatchNorm” layer and provides more features:
No matter whether you’re doing training or not, you can set the
training
argument to use batch statistics or EMA statistics. i.e., you can use batch statistics during inference, or use EMA statistics during training. Using EMA statistics in training is useful when you load a pretrained BN and don’t want to update it.As long as training=True, sync_statistics and ema_update option will take effect.

tensorpack.models.
BatchRenorm
(variable_scope_name, x, rmax, dmax, *, momentum=0.9, epsilon=1e05, center=True, scale=True, gamma_initializer=None, data_format='channels_last')[source]¶ Batch Renormalization layer, as described in the paper: Batch Renormalization: Towards Reducing Minibatch Dependence in BatchNormalized Models. This implementation is a wrapper around tf.layers.batch_normalization.
 Parameters
 Returns
tf.Tensor – a tensor named
output
with the same shape of x.
Variable Names:
beta
: the bias term.gamma
: the scale term. Input will be transformed byx * gamma + beta
.moving_mean, renorm_mean, renorm_mean_weight
: See TF documentation.moving_variance, renorm_stddev, renorm_stddev_weight
: See TF documentation.

tensorpack.models.
layer_register
(log_shape=False, use_scope=True)[source]¶  Parameters
 Returns
A decorator used to register a layer.
Example:
@layer_register(use_scope=True) def add10(x): return x + tf.get_variable('W', shape=[10]) # use it: output = add10('layer_name', input) # the function will be called under variable scope "layer_name".

class
tensorpack.models.
VariableHolder
(**kwargs)[source]¶ Bases:
object
A proxy to access variables defined in a layer.

tensorpack.models.
rename_tflayer_get_variable
()[source]¶ Rename all
tf.get_variable()
with rules that transforms tflayer style to tensorpack style. Returns
A context where the variables are renamed.
Example:
with rename_tflayer_get_variable(): x = tf.layer.conv2d(input, 3, 3, name='conv0') # variables will be named 'conv0/W', 'conv0/b'

tensorpack.models.
disable_layer_logging
()[source]¶ Disable the shape logging for all layers from this moment on. Can be useful when creating multiple towers.

tensorpack.models.
Conv2D
(variable_scope_name, inputs, filters, kernel_size, strides=(1, 1), padding='same', data_format='channels_last', dilation_rate=(1, 1), activation=None, use_bias=True, kernel_initializer=None, bias_initializer=<tf.python.ops.init_ops.Zeros object>, kernel_regularizer=None, bias_regularizer=None, activity_regularizer=None, split=1)[source]¶ Similar to tf.layers.Conv2D, but with some differences:
Default kernel initializer is variance_scaling_initializer(2.0).
Default padding is ‘same’.
Support ‘split’ argument to do group convolution.
Variable Names:
W
: weightsb
: bias

tensorpack.models.
Conv2DTranspose
(variable_scope_name, inputs, filters, kernel_size, strides=(1, 1), padding='same', data_format='channels_last', activation=None, use_bias=True, kernel_initializer=None, bias_initializer=<tf.python.ops.init_ops.Zeros object>, kernel_regularizer=None, bias_regularizer=None, activity_regularizer=None)[source]¶ A wrapper around tf.layers.Conv2DTranspose. Some differences to maintain backwardcompatibility:
Default kernel initializer is variance_scaling_initializer(2.0).
Default padding is ‘same’
Variable Names:
W
: weightsb
: bias

tensorpack.models.
FullyConnected
(variable_scope_name, inputs, units, activation=None, use_bias=True, kernel_initializer=None, bias_initializer=<tf.python.ops.init_ops.Zeros object>, kernel_regularizer=None, bias_regularizer=None, activity_regularizer=None)[source]¶ A wrapper around tf.layers.Dense. One difference to maintain backwardcompatibility: Default weight initializer is variance_scaling_initializer(2.0).
Variable Names:
W
: weights of shape [in_dim, out_dim]b
: bias

tensorpack.models.
LayerNorm
(variable_scope_name, x, epsilon=1e05, *, center=True, scale=True, gamma_initializer=<tf.python.ops.init_ops.Ones object>, data_format='channels_last')[source]¶ Layer Normalization layer, as described in the paper: Layer Normalization.
 Parameters
x (tf.Tensor) – a 4D or 2D tensor. When 4D, the layout should match data_format.
epsilon (float) – epsilon to avoid dividebyzero.
scale (center,) – whether to use the extra affine transformation or not.

tensorpack.models.
InstanceNorm
(variable_scope_name, x, epsilon=1e05, *, center=True, scale=True, gamma_initializer=<tf.python.ops.init_ops.Ones object>, data_format='channels_last', use_affine=None)[source]¶ Instance Normalization, as in the paper: Instance Normalization: The Missing Ingredient for Fast Stylization.
 Parameters
x (tf.Tensor) – a 4D tensor.
epsilon (float) – avoid dividebyzero
scale (center,) – whether to use the extra affine transformation or not.
use_affine – deprecated. Don’t use.

class
tensorpack.models.
LinearWrap
(tensor)[source]¶ Bases:
object
A simple wrapper to easily create “linear” graph, consisting of layers / symbolic functions with only one input & output.

apply
(func, *args, **kwargs)[source]¶ Apply a function on the wrapped tensor.
 Returns
LinearWrap –
LinearWrap(func(self.tensor(), *args, **kwargs))
.

apply2
(func, *args, **kwargs)[source]¶ Apply a function on the wrapped tensor. The tensor will be the second argument of func.
This is because many symbolic functions (such as tensorpack’s layers) takes ‘scope’ as the first argument.
 Returns
LinearWrap –
LinearWrap(func(args[0], self.tensor(), *args[1:], **kwargs))
.

print_tensor
()[source]¶ Print the underlying tensor and return self. Can be useful to get the name of tensors inside
LinearWrap
. Returns
self


tensorpack.models.
Maxout
([variable_scope_name, ]x, num_unit)[source]¶ Maxout as in the paper Maxout Networks.
 Parameters
x (tf.Tensor) – a NHWC or NC tensor. Channel has to be known.
num_unit (int) – a int. Must be divisible by C.
 Returns
tf.Tensor – of shape NHW(C/num_unit) named
output
.

tensorpack.models.
PReLU
(variable_scope_name, x, init=0.001, name=None)[source]¶ Parameterized ReLU as in the paper Delving Deep into Rectifiers: Surpassing HumanLevel Performance on ImageNet Classification.
 Parameters
Variable Names:
alpha
: learnable slope.

tensorpack.models.
BNReLU
([variable_scope_name, ]x, name=None)[source]¶ A shorthand of BatchNormalization + ReLU.
 Parameters
x (tf.Tensor) – the input
name – deprecated, don’t use.

tensorpack.models.
MaxPooling
(variable_scope_name, inputs, pool_size, strides=None, padding='valid', data_format='channels_last')[source]¶ Same as tf.layers.MaxPooling2D. Default strides is equal to pool_size.

tensorpack.models.
FixedUnPooling
(variable_scope_name, x, shape, unpool_mat=None, data_format='channels_last')[source]¶ Unpool the input with a fixed matrix to perform kronecker product with.
 Parameters
x (tf.Tensor) – a 4D image tensor
shape – int or (h, w) tuple
unpool_mat – a tf.Tensor or np.ndarray 2D matrix with size=shape. If is None, will use a matrix with 1 at topleft corner.
 Returns
tf.Tensor – a 4D image tensor.

tensorpack.models.
AvgPooling
(variable_scope_name, inputs, pool_size, strides=None, padding='valid', data_format='channels_last')[source]¶ Same as tf.layers.AveragePooling2D. Default strides is equal to pool_size.

tensorpack.models.
GlobalAvgPooling
(variable_scope_name, x, data_format='channels_last')[source]¶ Global average pooling as in the paper Network In Network.
 Parameters
x (tf.Tensor) – a 4D tensor.
 Returns
tf.Tensor – a NC tensor named
output
.

tensorpack.models.
regularize_cost
(regex, func, name='regularize_cost')[source]¶ Apply a regularizer on trainable variables matching the regex, and print the matched variables (only print once in multitower training). In replicated mode, it will only regularize variables within the current tower.
If called under a TowerContext with is_training==False, this function returns a zero constant tensor.
 Parameters
regex (str) – a regex to match variable names, e.g. “conv.*/W”
func – the regularization function, which takes a tensor and returns a scalar tensor. E.g.,
tf.nn.l2_loss, tf.contrib.layers.l1_regularizer(0.001)
.
 Returns
tf.Tensor – a scalar, the total regularization cost.
Example
cost = cost + regularize_cost("fc.*/W", l2_regularizer(1e5))

tensorpack.models.
regularize_cost_from_collection
(name='regularize_cost')[source]¶ Get the cost from the regularizers in
tf.GraphKeys.REGULARIZATION_LOSSES
. If in replicated mode, will only regularize variables created within the current tower. Parameters
name (str) – the name of the returned tensor
 Returns
tf.Tensor – a scalar, the total regularization cost.

tensorpack.models.
Dropout
([variable_scope_name, ]x, *args, **kwargs)[source]¶ Same as tf.layers.dropout. However, for historical reasons, the first positional argument is interpreted as keep_prob rather than drop_prob. Explicitly use rate= keyword arguments to ensure things are consistent.

tensorpack.models.
ConcatWith
([variable_scope_name, ]x, tensor, dim)[source]¶ A wrapper around
tf.concat
to cooperate withLinearWrap
.