Compression UG edits. #3352

dwelsch-esi · 2024-09-20T21:52:12Z

Style and grammar edits of the Compression User Guide.

quic-hitameht · 2024-10-01T18:53:06Z

Docs/user_guide/spatial_svd.rst


-Spatial SVD is a tensor decomposition technique which decomposes one large layer (in terms of mac or memory) into two smaller layers. SVD stands for Singular Value Decomposition.
+Spatial singular value decomposition (SSVD) is a technique that decomposes one large convolution (Conv) MAC or memory layer into two smaller layers.


Can we not abbreviate Spatial SVD to SSVD?

quic-hitameht · 2024-10-01T18:53:33Z

Docs/user_guide/spatial_svd.rst

+- ℎ is the height of the kernel
+- 𝑤 is the width of the kernel 
+
+SSVD decomposes the kernel into two kernels, one of size (𝑚,𝑘,ℎ,1) and one of size (𝑘,𝑛,1,𝑤), where 𝑘 is called the `rank`. The smaller the value of 𝑘, the larger the degree of compression.


Same comment as above. Please replace SSVD w/ Spatial SVD.

quic-hitameht · 2024-10-01T18:53:42Z

Docs/user_guide/spatial_svd.rst

+
+SSVD decomposes the kernel into two kernels, one of size (𝑚,𝑘,ℎ,1) and one of size (𝑘,𝑛,1,𝑤), where 𝑘 is called the `rank`. The smaller the value of 𝑘, the larger the degree of compression.
+
+The following figure illustrates how SSVD decomposes both the output channel dimension and the size of the Conv kernel itself. 


Same comment as above. Please replace SSVD w/ Spatial SVD.

quic-hitameht · 2024-10-01T18:59:56Z

Docs/user_guide/greedy_compression_ratio_selection.rst


 Overview
 ========
-The model compression methods, Spatial SVD and Channel Pruning work on per layer basis. Not all the layers in the given model are equally compressible. Compression of individual layers of a given model can have varying impact on the final accuracy of the model. Greedy Per Layer Compression Ratio Selection Algorithm is used to assess the sensitivity of applicable layers to compression and find appropriate compression-ratio for each individual layers. The algorithm makes sure that the entire model has highest remaining accuracy and also meets the given target compression-ratio.
+
+Spatial SVD (SSVD) and channel pruning (CP) work on individual layers of a model. Not all the layers are equally compressible, so compression of a given layer has a variable impact on the final model accuracy. The greedy per-layer compression ratio selection algorithm assesses the sensitivity of layers to compression and finds an appropriate compression ratio for each layer. The algorithm ensures that the model maintains the highest possible accuracy while meeting the target compression ratio.


Same comment here. replace Spatial SVD (SSVD) w/ Spatial SVD.

…tats * Adding boxplots for advanced stats and modifying the backend for the same * Cleanup JavaScript callbacks by packaging repetitive code as a function in utils * Generalize the code in callbacks by adaptively iterating over columns present in datasources instead of manually listing them * Dynamically adjust boxplot width according to the number of boxplots * Update docstrings --------- Signed-off-by: Ishan Pendse <quic_ipendse@quicinc.com> Signed-off-by: Dave Welsch <dwelsch@expertsupport.com>

* Implement pyproject.toml to build AIMET package There are 3 dynamic fields in metadata: - version - dependencies - name (not PEP compatible!) A plugin of scikit-build-core build system generates dependencies and package name based on `CMAKE_ARGS` environment variable. Signed-off-by: Evgeny Mironov <quic_emironov@quicinc.com> * Build all docker images from the single Dockerfile Docker images contain dependencies to build and run tests. Signed-off-by: Evgeny Mironov <quic_emironov@quicinc.com> * Build docker images on a CI Signed-off-by: Evgeny Mironov <quic_emironov@quicinc.com> --------- Signed-off-by: Evgeny Mironov <quic_emironov@quicinc.com> Co-authored-by: Evgeny Mironov <quic_emironov@quicinc.com> Signed-off-by: Dave Welsch <dwelsch@expertsupport.com>

Signed-off-by: Raj Gite <quic_rgite@quicinc.com> Signed-off-by: Dave Welsch <dwelsch@expertsupport.com>

Signed-off-by: Kyunggeun Lee <quic_kyunggeu@quicinc.com> Signed-off-by: Dave Welsch <dwelsch@expertsupport.com>

Signed-off-by: Kevin Hsieh <quic_klhsieh@quicinc.com> Signed-off-by: Dave Welsch <dwelsch@expertsupport.com>

* Edited Quantization User Guide. Signed-off-by: Dave Welsch <dwelsch@expertsupport.com> * Quantization Guide - more edits. Signed-off-by: Dave Welsch <dwelsch@expertsupport.com> * Corrected TOC errors introduced by Quant UG edits. Signed-off-by: Dave Welsch <dwelsch@expertsupport.com> * Review changes PR quic#3348 - Quantization UG edits. Signed-off-by: Dave Welsch <dwelsch@expertsupport.com> * More review changes PR quic#3348 - Quantization UG edits. Signed-off-by: Dave Welsch <dwelsch@expertsupport.com> --------- Signed-off-by: Dave Welsch <dwelsch@expertsupport.com>

Signed-off-by: Kyunggeun Lee <quic_kyunggeu@quicinc.com> Signed-off-by: Dave Welsch <dwelsch@expertsupport.com>

Signed-off-by: Kevin Hsieh <quic_klhsieh@quicinc.com> Signed-off-by: Dave Welsch <dwelsch@expertsupport.com>

Signed-off-by: Sai Chaitanya Gajula <quic_gsaichai@quicinc.com> Signed-off-by: Dave Welsch <dwelsch@expertsupport.com>

- Backend doesn't expect non-float tensors to have encodings. Compare op's output is a tensor of bools for which the encoding shouldn't have been present. This change fixes this issue by disabling the quantizer. Signed-off-by: yathindra kota <quic_ykota@quicinc.com> Signed-off-by: Dave Welsch <dwelsch@expertsupport.com>