Adaptive quantization

A computer-implemented decoding method comprising:

• obtaining a compressed tensor code, wherein a tensor is divided into a set of elements, each element including a value;
• obtaining a compressed quantization mask, wherein the quantization mask is associated with the tensor code, the quantization mask indicating, for each of a set of elements into which the tensor is divided, an alphabet size associated with the element, wherein an alphabet size is a number of values that may be used to quantize the data of a tensor element;
• obtaining a compressed probability prediction, wherein the probability prediction comprises a set of probability distributions, each probability distribution associated with an element of the tensor over possible elements of the tensor;
• decoding the compressed quantization mask to obtain the quantization mask;
• decoding the compressed probability prediction to obtain the probability prediction;
• and
• decompressing the compressed tensor code using an entropy decoding algorithm wherein each element of the tensor code is decoded using the associated alphabet indicated by the quantization mask and the associated probability distribution indicated by the probability prediction.

A non-transitory computer-readable storage medium storing computer program instructions executable by one or more processors of a system to perform encoding steps comprising:

• receiving a tensor for compression, the tensor divided into a set of elements wherein each element includes a value;
• obtaining a quantization mask associated with the tensor, the quantization mask indicating, for each of the set of elements into which the tensor is divided, an alphabet size associated with the element, wherein an alphabet size is a number of values that may be used to quantize the data of a tensor element;
• generating a quantized tensor by quantizing the tensor using information from the quantization mask, wherein quantizing the tensor comprises quantizing each element of the tensor using the alphabet size associated with the element, as indicated in the quantization mask;
• generating, based on the received tensor, a probability prediction comprising a probability distribution associated with each element of the tensor over the possible values of elements of the tensor;
• and
• compressing the quantized tensor into a tensor code using entropy coding, wherein a probability used to code each element of the quantized tensor is obtained from the probability distribution.

A non-transitory computer-readable storage medium storing computer program instructions executable by one or more processors of a system to perform decoding steps comprising:

• obtaining a compressed tensor code, wherein a tensor is divided into a set of elements, each element including a value;
• obtaining a compressed quantization mask, wherein the quantization mask is associated with the tensor code, the quantization mask indicating, for each of a set of elements into which the tensor is divided, an alphabet size associated with the element, wherein an alphabet size is a number of values that may be used to quantize the data of a tensor element;
• obtaining a compressed probability prediction, wherein the probability prediction comprises a set of probability distributions, each probability distribution associated with an element of the tensor over possible elements of the tensor;
• decoding the compressed quantization mask to obtain the quantization mask;
• decoding the compressed probability prediction to obtain the probability prediction;
• and
• decompressing the compressed tensor code using an entropy decoding algorithm wherein each element of the tensor code is decoded using the associated alphabet indicated by the quantization mask and the associated probability distribution indicated by the probability prediction.