Intermediate Tensor Compression Profiles for Low-Latency Bitstreams
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video coding systems face challenges in efficiently compressing and transmitting intermediate data tensors, particularly in terms of bandwidth and latency, without compromising on accuracy and processing efficiency.
Innovation Solution
A system and method for compressing intermediate data tensors by generating a compressed set of tensors along with metadata that includes compression profiles, split point configurations, and tensor characteristics, allowing for efficient decompression and data processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If intermediate data tensors are compressed using compression profiles, then bandwidth usage is optimized and transmission efficiency is improved, but latency may increase due to compression and decompression processing
Solution Approach 1:
The system dynamically selects compression profiles based on tensor characteristics (shape, data type, sparsity) and transmission requirements. Different compression techniques are applied adaptively: low-rank approximation for certain tensor types, sparse representation for others, and configurable compression levels that balance bandwidth savings against processing time. This dynamic adaptation resolves the contradiction by optimizing the compression-decompression tradeoff for each specific case.
Solution Approach 2:
The compression profile includes multiple parameters such as compression level, approximation error thresholds, and format specifications that can be adjusted to balance bandwidth efficiency and processing speed. By changing these parameters based on the specific tensor characteristics and network conditions, the system can optimize the tradeoff between compression ratio (bandwidth savings) and processing time (latency).
2Loss of energy
If compression profiles are applied to intermediate data tensors, then transmission bandwidth is reduced, but processing complexity increases due to compression and decompression operations
Solution Approach 1:
The system segments the compression process into distinct stages: tensor analysis phase (determining tensor characteristics), profile selection phase (choosing appropriate compression technique), and compression execution phase (applying the selected profile). Similarly, decompression is segmented into verification, decompression, and reconstruction stages. This segmentation allows each stage to be optimized independently and facilitates parallel processing where applicable, reducing overall processing complexity.
Solution Approach 2:
The system uses metadata copies and proxy representations during compression profile application. Instead of directly compressing the full tensor data, the system first creates and processes compact metadata descriptions that guide the compression process. This copying approach reduces the complexity of direct tensor manipulation while maintaining compression effectiveness.
3Productivity
If compression is applied to intermediate data tensors, then transmission efficiency is improved, but accuracy may be compromised due to approximation errors
Solution Approach 1:
The compression profile includes configurable accuracy parameters such as maximum approximation error thresholds, precision levels, and tolerance ranges. By adjusting these parameters, the system can guarantee that compression-induced errors remain within acceptable bounds for the specific application. Different accuracy levels can be selected based on the sensitivity of the downstream processing tasks.
Solution Approach 2:
The system incorporates accuracy verification mechanisms where the compressed tensor representation is evaluated against original data characteristics. If the approximation error exceeds predefined thresholds, the system provides feedback to adjust compression parameters or select alternative compression techniques. This feedback loop ensures that transmission efficiency gains do not come at the cost of unacceptable accuracy degradation.
Data Source
Figure 1A
Figure 1B
Figure 1C
AI summary
System, methods, and instrumentalities are disclosed for carrying compressed intermediate data tensors. An example device for media encoding may receive media data. The device may generate intermediate data based on the media data. A portion of the intermediate data may be associated with an intermediate data tensor. The device may determine a set of intermediate data tensors based on a tensor characteristic associated with the intermediate data tensor. The set of intermediate data tensors may include the intermediate data tensor. The device may generate a compressed set of intermediate data tensors based on the set of intermediate data tensors and a compression profile associated with the set of intermediate data tensors. The device may include the compressed set of intermediate data tensors and metadata associated with the compressed set of intermediate data tensors in a bitstream.