Incremental precision networks using residual inference and fine-grain quantization
One embodiment provides for a computing device comprising a parallel processor compute unit to perform a set of parallel integer compute operations; a ternarization unit including a weight ternarization circuit and an activation quantization circuit; wherein the weight ternarization circuit is to convert a weight tensor from a floating-point representation to a ternary representation including a ternary weight and a scale factor; wherein the activation quantization circuit is to convert an activation tensor from a floating-point representation to an integer representation; and wherein the parallel processor compute unit includes one or more circuits to perform the set of parallel integer compute operations on the ternary representation of the weight tensor and the integer representation of the activation tensor.
1. A method of performing ternary quantization on weight data of an artificial neural network, the method comprising:
ternarizing a group of pre-trained weights, wherein ternarizing the group of pre-trained weights includes:
partitioning the weight data along each of one or more input channels of a layer of the artificial neural network;
determining a scale factor and threshold for each partition; and
performing quantization on each partition of the weight data to output quantized ternary weight data;
determining that the group of pre-trained weights is eligible for residual computation to generate residual weights, wherein eligibility is determined based on one or more distribution-based heuristics;
computing the residual weights by determining an arithmetic difference between values of pre-trained weights of the group and values of the ternary weights;
ternarizing the residual weights; and
storing the ternarized residual weights.
2. The method as in claim 1 , wherein the distribution-based heuristics include:
determining the eligibility of the group of pre-trained weights based on the dynamic range of the group of pre-trained weights; and
determining the eligibility of the group based on whether the group has a heavy-tailed distribution.
3. The method as in claim 2 , wherein partitioning the weight data includes partitioning the weight data using one or more static partitions, the one or more static partitions including one or more partitions for each layer of the artificial neural network.
4. The method as in claim 2 , wherein partitioning the weight data includes partitioning the weight data using dynamic partitions within the layer of the artificial neural network.
5. The method as in claim 4 , additionally comprising dynamically determining a set of weight data partitions for the layer of the artificial neural network, wherein dynamically determining the set of weight data partitions for layer includes determining a first dynamic partition for weight data having a first data distribution and determining a second dynamic partition for weight data having a second data distribution.
6. The method as in claim 5 , additionally comprising:
determining a first scale factor for the first dynamic partition; and
determining a second scale factor for the second dynamic partition, the second scale factor different from the first scale factor.
7. A non-transitory machine readable medium storing instructions execution of which by one or more processors of a computing device, cause the computing device to perform operations comprising:
ternarizing a group of pre-trained weights, wherein ternarizing the group of pre-trained weights includes:
partitioning weight data along each of one or more input channels of a layer of an artificial neural network;
determining a scale factor and threshold for each partition; and
performing quantization on each partition of the weight data to output quantized ternary weight data;
determining that the group of pre-trained weights is eligible for residual computation to generate residual weights, wherein eligibility is determined based on one or more distribution-based heuristics;
computing the residual by determining an arithmetic difference between values of pre-trained weights of the group and values of the ternary weights;
ternarizing the residual weights; and
storing the ternarized residual weights.
8. The non-transitory machine-readable medium as in claim 7 , wherein the distribution-based heuristics include:
determining the eligibility of the group of pre-trained weights based on the dynamic range of the group of pre-trained weights; and
determining the eligibility of the group based on whether the group has a heavy-tailed distribution.
9. The non-transitory machine-readable medium as in claim 8 , wherein partitioning the weight data includes partitioning the weight data using one or more static partitions, the one or more static partitions including one or more partitions for each layer.
10. The non-transitory machine-readable medium as in claim 8 , wherein partitioning the weight data includes determining a set of dynamic partitions for the layer of the artificial neural network.
11. The non-transitory machine-readable medium as in claim 10 , wherein determining the set of dynamic partitions for the layer includes determining a first dynamic partition for weight data having a first data distribution and determining a second dynamic partition for weight data having a second data distribution.
12. The non-transitory machine-readable medium as in claim 11 , the operations additionally comprising:
determining a first scale factor for the first dynamic partition; and
determining a second scale factor for the second dynamic partition, the second scale factor different from the first scale factor.
13. A data processing system comprising:
a memory device to store instructions; and
a processor configured to execute the instructions stored in the memory, execution of which cause the processor to perform operations comprising:
ternarizing a group of pre-trained weights, wherein ternarizing the group of pre-trained weights includes:
partitioning weight data along each of one or more input channels of a layer of an artificial neural network;
determining a scale factor and threshold for each partition; and
performing quantization on each partition of the weight data to output quantized ternary weight data;
determining that the group of pre-trained weights is eligible for residual computation to generate residual weights, wherein eligibility is determined based on one or more distribution-based heuristics;
computing the residual weights by determining an arithmetic difference between values of pre-trained weights of the group and values of the ternary weights;
ternarizing the residual weights; and
storing the ternarized residual weights.
14. The data processing system as in claim 13 , wherein the distribution-based heuristics include:
determining the eligibility of the group of pre-trained weights based on the dynamic range of the group of pre-trained weights; and
determining the eligibility of the group based on whether the group has a heavy-tailed distribution.
15. The data processing system as in claim 14 , wherein partitioning the weight data includes partitioning the weight data using one or more static partitions, the one or more static partitions including one or more partitions for each layer of the artificial neural network.
16. The data processing system as in claim 14 , wherein partitioning the weight data includes partitioning the weight data using dynamic partitions within the layer of the artificial neural network.
17. The data processing system as in claim 16 , the operations additionally comprising dynamically determining a set of weight data partitions for the layer of the artificial neural network.
18. The data processing system as in claim 17 , wherein dynamically determining the set of weight data partitions for layer includes determining a first dynamic partition for weight data having a first data distribution and determining a second dynamic partition for weight data having a second data distribution.
19. The data processing system as in claim 18 , the operations additionally comprising:
determining a first scale factor for the first dynamic partition; and
determining a second scale factor for the second dynamic partition, the second scale factor different from the first scale factor.