TensorFlow Compression

description: Data compression in TensorFlow.

Module: tfc

Data compression in TensorFlow.

Classes

class ContinuousBatchedEntropyModel: Batched entropy model for continuous random variables.

class ContinuousIndexedEntropyModel: Indexed entropy model for continuous random variables.

class DeepFactorized: Fully factorized distribution based on neural network cumulative.

class GDN: Generalized divisive normalization layer.

class GDNParameter: Nonnegative parameterization as needed for GDN parameters.

class IdentityInitializer: Initialize to the identity kernel with the given shape.

class LocationScaleIndexedEntropyModel: Indexed entropy model for location-scale family of random variables.

class MonotonicAdapter: Adapt a continuous distribution via an ascending monotonic function.

class NoisyDeepFactorized: DeepFactorized that is convolved with uniform noise.

class NoisyLogistic: Logistic distribution with additive i.i.d. uniform noise.

class NoisyLogisticMixture: Mixture of logistic distributions with additive i.i.d. uniform noise.

class NoisyMixtureSameFamily: Mixture of distributions with additive i.i.d. uniform noise.

class NoisyNormal: Gaussian distribution with additive i.i.d. uniform noise.

class NoisyNormalMixture: Mixture of normal distributions with additive i.i.d. uniform noise.

class NoisyRoundedDeepFactorized: Rounded DeepFactorized + uniform noise.

class NoisyRoundedNormal: Rounded normal distribution + uniform noise.

class NoisySoftRoundedDeepFactorized: Soft rounded deep factorized distribution + uniform noise.

class NoisySoftRoundedNormal: Soft rounded normal distribution + uniform noise.

class PackedTensors: Packed representation of compressed tensors.

class Parameter: Reparameterized Layer variable.

class RDFTParameter: RDFT reparameterization of a convolution kernel.

class Round: Applies rounding.

class RoundAdapter: Continuous density function + round.

class SignalConv1D: 1D convolution layer.

class SignalConv2D: 2D convolution layer.

class SignalConv3D: 3D convolution layer.

class SoftRound: Applies a differentiable approximation of rounding.

class SoftRoundAdapter: Differentiable approximation to round.

class SoftRoundConditionalMean: Conditional mean of inputs given noisy soft rounded values.

class UniformNoiseAdapter: Additive i.i.d. uniform noise adapter distribution.

class UniversalBatchedEntropyModel: Batched entropy model model which implements Universal Quantization.

class UniversalIndexedEntropyModel: Indexed entropy model model which implements Universal Quantization.

Functions

estimate_tails(...): Estimates approximate tail quantiles.

irdft_matrix(...): Matrix for implementing kernel reparameterization with tf.matmul.

lower_bound(...): Same as tf.maximum, but with helpful gradient for inputs < bound.

lower_tail(...): Approximates lower tail quantile for range coding.

perturb_and_apply(...): Perturbs the inputs of a pointwise function.

pmf_to_quantized_cdf(...): Converts a PMF into a quantized CDF for range coding.

quantization_offset(...): Computes distribution-dependent quantization offset.

range_decode(...): Range-decodes code into an int32 tensor of shape shape.

range_encode(...): Range encodes integer data with a finite alphabet.

same_padding_for_kernel(...): Determine correct amount of padding for same convolution.

soft_round(...): Differentiable approximation to round().

soft_round_conditional_mean(...): Conditional mean of inputs given noisy soft rounded values.

soft_round_inverse(...): Inverse of soft_round().

unbounded_index_range_decode(...): Range decodes encoded using an indexed probability table.

unbounded_index_range_encode(...): Range encodes unbounded integer data using an indexed probability table.

upper_bound(...): Same as tf.minimum, but with helpful gradient for inputs > bound.

upper_tail(...): Approximates upper tail quantile for range coding.