[quant] Implement PTQ for APoT FakeQuant#81040
Closed
asl3 wants to merge 16 commits intogh/asl3/40/basefrom
Closed
[quant] Implement PTQ for APoT FakeQuant#81040asl3 wants to merge 16 commits intogh/asl3/40/basefrom
asl3 wants to merge 16 commits intogh/asl3/40/basefrom
Conversation
[ghstack-poisoned]
Contributor
🔗 Helpful links
✅ No Failures (0 Pending)As of commit 0530d26 (more details on the Dr. CI page): Expand to see more💚 💚 Looks good so far! There are no failures yet. 💚 💚 This comment was automatically generated by Dr. CI (expand for details).Please report bugs/suggestions to the (internal) Dr. CI Users group. |
[ghstack-poisoned]
### Summary: This PR implements FX Graph Mode QAT for APoT FakeQuant. ### Test Plan: Run models with: `python test/quantization/core/experimental/fx_graph_mode_apot.py` Accuracy Stats: Model #1: uniform activation, uniform weight (FX Graph Mode quantized) Size of model (MB): 46.801265 Evaluation accuracy on test dataset: 69.76%, 89.08% Model #2: uniform activation, APoT weight (FX Graph Mode quantized) Size of model (MB): 46.820369 Evaluation accuracy on test dataset: 69.00%, 88.66% Model #3: APoT activation and weight (FX Graph Mode quantized) Size of model (MB): 46.801431 Evaluation accuracy on test dataset: 69.76%, 89.08% Eager mode quantized model Resnet18 Size of model (MB): 11.839989 Evaluation accuracy on test dataset: 69.49%, 88.90% [ghstack-poisoned]
### Summary: This PR implements FX Graph Mode QAT for APoT FakeQuant. ### Test Plan: Run models with: `python test/quantization/core/experimental/fx_graph_mode_apot.py` ### Accuracy Stats: Uniform: int8 APoT: 8-bit (b = 8, k = 2) **Model #1:** Uniform activation, uniform weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 69.54%, 88.99% **Model #2:** Uniform activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 69.76%, 89.08% **Model #3:** APoT activation and weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 69.54%, 89.04% **Model #4:** Eager mode quantized model Resnet18 Evaluation accuracy on test dataset: 69.49%, 88.90% [ghstack-poisoned]
### Summary: This PR implements FX Graph Mode QAT for APoT FakeQuant. ### Test Plan: Run models with: `python test/quantization/core/experimental/fx_graph_mode_apot.py` ### Accuracy Stats: Uniform: int8 APoT: 8-bit (b = 8, k = 2) **Model #1:** Uniform activation, uniform weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 69.54%, 88.99% **Model #2:** Uniform activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 69.76%, 89.08% **Model #3:** APoT activation and weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 69.54%, 89.04% **Model #4:** Eager mode quantized model Resnet18 Evaluation accuracy on test dataset: 69.49%, 88.90% [ghstack-poisoned]
### Summary: This PR implements FX Graph Mode QAT for APoT FakeQuant. ### Test Plan: Run models with: `python test/quantization/core/experimental/fx_graph_mode_apot.py` ### Accuracy Stats: Uniform: int8 APoT: 8-bit (b = 8, k = 2) **Model #1:** Uniform activation, uniform weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 69.54%, 88.99% **Model #2:** Uniform activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 69.76%, 89.08% **Model #3:** APoT activation and weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 69.54%, 89.04% **Model #4:** Eager mode quantized model Resnet18 Evaluation accuracy on test dataset: 69.49%, 88.90% [ghstack-poisoned]
### Summary: This PR implements FX Graph Mode QAT for APoT FakeQuant. ### Test Plan: Run models with: `python test/quantization/core/experimental/fx_graph_mode_apot.py` ### Accuracy Stats: Uniform: int8 APoT: 8-bit (b = 8, k = 2) **Model #1:** Uniform activation, uniform weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 69.54%, 88.99% **Model #2:** Uniform activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 69.76%, 89.08% **Model #3:** APoT activation and weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 69.54%, 89.04% **Model #4:** Eager mode quantized model Resnet18 Evaluation accuracy on test dataset: 69.49%, 88.90% [ghstack-poisoned]
Contributor
|
this PR is for PTQ and not QAT? |
### Summary: This PR implements FX Graph Mode QAT for APoT FakeQuant. ### Test Plan: Run models with: `python test/quantization/core/experimental/fx_graph_mode_apot.py` ### Accuracy Stats: Uniform: int8 APoT: 8-bit (b = 8, k = 2) **Model #1:** Uniform activation, uniform weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 69.54%, 88.99% **Model #2:** Uniform activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 69.76%, 89.08% **Model #3:** APoT activation and weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 69.54%, 89.04% **Model #4:** Eager mode quantized model Resnet18 Evaluation accuracy on test dataset: 69.49%, 88.90% [ghstack-poisoned]
jerryzh168
reviewed
Jul 26, 2022
jerryzh168
reviewed
Jul 26, 2022
jerryzh168
reviewed
Jul 26, 2022
jerryzh168
reviewed
Jul 26, 2022
jerryzh168
reviewed
Jul 26, 2022
jerryzh168
reviewed
Jul 26, 2022
jerryzh168
reviewed
Jul 26, 2022
### Summary: This PR implements PTQ for APoT FakeQuant. ### Test Plan: Run models with: `python test/quantization/core/experimental/fx_graph_mode_apot.py` ### Accuracy Stats: 8-bit (Uniform int8, APoT b = 8 k = 2) **Model #1:** Uniform activation, uniform weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.43%, 85.62% **Model #2:** Uniform activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.51%, 85.78% **Model #3:** APoT activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.32%, 85.78% 4-bit (Uniform int4, APoT b = 8 k = 2) **Model #1:** Uniform activation, uniform weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 45.63%, 71.96% **Model #2:** Uniform activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.24%, 85.56% **Model #3:** APoT activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 45.40%, 76.21% **Full Precision model (FX Graph Mode quantized)** Evaluation accuracy on test dataset: 69.76%, 89.08% **Eager mode quantized model** Evaluation accuracy on test dataset: 69.49%, 88.90% [ghstack-poisoned]
### Summary: This PR implements PTQ for APoT FakeQuant. ### Test Plan: Run models with: `python test/quantization/core/experimental/fx_graph_mode_apot.py` ### Accuracy Stats: 8-bit (Uniform int8, APoT b = 8 k = 2) **Model #1:** Uniform activation, uniform weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.43%, 85.62% **Model #2:** Uniform activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.51%, 85.78% **Model #3:** APoT activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.32%, 85.78% 4-bit (Uniform int4, APoT b = 8 k = 2) **Model #1:** Uniform activation, uniform weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 45.63%, 71.96% **Model #2:** Uniform activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.24%, 85.56% **Model #3:** APoT activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 45.40%, 76.21% **Full Precision model (FX Graph Mode quantized)** Evaluation accuracy on test dataset: 69.76%, 89.08% **Eager mode quantized model** Evaluation accuracy on test dataset: 69.49%, 88.90% [ghstack-poisoned]
Contributor
Author
|
@pytorchbot merge -g |
Collaborator
|
@pytorchbot successfully started a merge job. Check the current status here |
Collaborator
|
Merge failed due to Refusing to merge as mandatory check(s) pull failed for rule superuser |
### Summary: This PR implements PTQ for APoT FakeQuant. It runs models (Resnet-18 pre-trained model, ImageNet dataset) to compare accuracy metrics for different qconfig settings of uniform vs. APoT quantized activation and weight. According to the collected accuracy stats, model #2 (uniform activation and APoT weight) appears to have a slight improvement in accuracy compared to model #1 (uniform activation and uniform weight) for 8-bit and significant improvement for 4-bit (see "Accuracy Stats" section below). ### Test Plan: Run models with: `python test/quantization/core/experimental/fx_graph_mode_apot.py` ### Accuracy Stats: 8-bit (Uniform int8, APoT b = 8 k = 2) **Model #1:** Uniform activation, uniform weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.43% (Top-1), 85.62% (Top-5) **Model #2:** Uniform activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.51% (Top-1), 85.78% (Top-5) **Model #3:** APoT activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.32% (Top-1), 85.78% (Top-5) 4-bit (Uniform int4, APoT b = 4 k = 2) **Model #1:** Uniform activation, uniform weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 45.63% (Top-1), 71.96% (Top-5) **Model #2:** Uniform activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.24% (Top-1), 85.56% (Top-5) **Model #3:** APoT activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 45.40% (Top-1), 76.21% (Top-5) **Full Precision model (FX Graph Mode quantized)** Evaluation accuracy on test dataset: 69.76% (Top-1), 89.08% (Top-5) **Eager mode quantized model** Evaluation accuracy on test dataset: 69.49% (Top-1), 88.90% (Top-5) [ghstack-poisoned]
### Summary: This PR implements PTQ for APoT FakeQuant. It runs models (Resnet-18 pre-trained model, ImageNet dataset) to compare accuracy metrics for different qconfig settings of uniform vs. APoT quantized activation and weight. According to the collected accuracy stats, model #2 (uniform activation and APoT weight) appears to have a slight improvement in accuracy compared to model #1 (uniform activation and uniform weight) for 8-bit and significant improvement for 4-bit (see "Accuracy Stats" section below). ### Test Plan: Run models with: `python test/quantization/core/experimental/fx_graph_mode_apot.py` ### Accuracy Stats: 8-bit (Uniform int8, APoT b = 8 k = 2) **Model #1:** Uniform activation, uniform weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.43% (Top-1), 85.62% (Top-5) **Model #2:** Uniform activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.51% (Top-1), 85.78% (Top-5) **Model #3:** APoT activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.32% (Top-1), 85.78% (Top-5) 4-bit (Uniform int4, APoT b = 4 k = 2) **Model #1:** Uniform activation, uniform weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 45.63% (Top-1), 71.96% (Top-5) **Model #2:** Uniform activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.24% (Top-1), 85.56% (Top-5) **Model #3:** APoT activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 45.40% (Top-1), 76.21% (Top-5) **Full Precision model (FX Graph Mode quantized)** Evaluation accuracy on test dataset: 69.76% (Top-1), 89.08% (Top-5) **Eager mode quantized model** Evaluation accuracy on test dataset: 69.49% (Top-1), 88.90% (Top-5) [ghstack-poisoned]
Contributor
Author
|
@pytorchbot merge -g |
Collaborator
|
@pytorchbot successfully started a merge job. Check the current status here |
Collaborator
|
Merge failed due to Command Raised by https://github.com/pytorch/pytorch/actions/runs/2750364309 |
### Summary: This PR implements PTQ for APoT FakeQuant. It runs models (Resnet-18 pre-trained model, ImageNet dataset) to compare accuracy metrics for different qconfig settings of uniform vs. APoT quantized activation and weight. According to the collected accuracy stats, model #2 (uniform activation and APoT weight) appears to have a slight improvement in accuracy compared to model #1 (uniform activation and uniform weight) for 8-bit and significant improvement for 4-bit (see "Accuracy Stats" section below). ### Test Plan: Run models with: `python test/quantization/core/experimental/fx_graph_mode_apot.py` ### Accuracy Stats: 8-bit (Uniform int8, APoT b = 8 k = 2) **Model #1:** Uniform activation, uniform weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.43% (Top-1), 85.62% (Top-5) **Model #2:** Uniform activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.51% (Top-1), 85.78% (Top-5) **Model #3:** APoT activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.32% (Top-1), 85.78% (Top-5) 4-bit (Uniform int4, APoT b = 4 k = 2) **Model #1:** Uniform activation, uniform weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 45.63% (Top-1), 71.96% (Top-5) **Model #2:** Uniform activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.24% (Top-1), 85.56% (Top-5) **Model #3:** APoT activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 45.40% (Top-1), 76.21% (Top-5) **Full Precision model (FX Graph Mode quantized)** Evaluation accuracy on test dataset: 69.76% (Top-1), 89.08% (Top-5) **Eager mode quantized model** Evaluation accuracy on test dataset: 69.49% (Top-1), 88.90% (Top-5) [ghstack-poisoned]
Contributor
Author
|
@pytorchbot merge -g |
Collaborator
|
@pytorchbot successfully started a merge job. Check the current status here |
### Summary: This PR implements PTQ for APoT FakeQuant. It runs models (Resnet-18 pre-trained model, ImageNet dataset) to compare accuracy metrics for different qconfig settings of uniform vs. APoT quantized activation and weight. According to the collected accuracy stats, model #2 (uniform activation and APoT weight) appears to have a slight improvement in accuracy compared to model #1 (uniform activation and uniform weight) for 8-bit and significant improvement for 4-bit (see "Accuracy Stats" section below). ### Test Plan: Run models with: `python test/quantization/core/experimental/fx_graph_mode_apot.py` ### Accuracy Stats: 8-bit (Uniform int8, APoT b = 8 k = 2) **Model #1:** Uniform activation, uniform weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.43% (Top-1), 85.62% (Top-5) **Model #2:** Uniform activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.51% (Top-1), 85.78% (Top-5) **Model #3:** APoT activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.32% (Top-1), 85.78% (Top-5) 4-bit (Uniform int4, APoT b = 4 k = 2) **Model #1:** Uniform activation, uniform weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 45.63% (Top-1), 71.96% (Top-5) **Model #2:** Uniform activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.24% (Top-1), 85.56% (Top-5) **Model #3:** APoT activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 45.40% (Top-1), 76.21% (Top-5) **Full Precision model (FX Graph Mode quantized)** Evaluation accuracy on test dataset: 69.76% (Top-1), 89.08% (Top-5) **Eager mode quantized model** Evaluation accuracy on test dataset: 69.49% (Top-1), 88.90% (Top-5) [ghstack-poisoned]
Contributor
Author
|
@pytorchbot merge -g |
Collaborator
|
@pytorchbot successfully started a merge job. Check the current status here |
facebook-github-bot
pushed a commit
that referenced
this pull request
Jul 29, 2022
Summary: ### Summary: This PR implements PTQ for APoT FakeQuant. It runs models (Resnet-18 pre-trained model, ImageNet dataset) to compare accuracy metrics for different qconfig settings of uniform vs. APoT quantized activation and weight. According to the collected accuracy stats, model #2 (uniform activation and APoT weight) appears to have a slight improvement in accuracy compared to model #1 (uniform activation and uniform weight) for 8-bit and significant improvement for 4-bit (see "Accuracy Stats" section below). ### Test Plan: Run models with: `python test/quantization/core/experimental/fx_graph_mode_apot.py` ### Accuracy Stats: 8-bit (Uniform int8, APoT b = 8 k = 2) **Model #1:** Uniform activation, uniform weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.43% (Top-1), 85.62% (Top-5) **Model #2:** Uniform activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.51% (Top-1), 85.78% (Top-5) **Model #3:** APoT activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.32% (Top-1), 85.78% (Top-5) 4-bit (Uniform int4, APoT b = 4 k = 2) **Model #1:** Uniform activation, uniform weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 45.63% (Top-1), 71.96% (Top-5) **Model #2:** Uniform activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 64.24% (Top-1), 85.56% (Top-5) **Model #3:** APoT activation, APoT weight (FX Graph Mode quantized) Evaluation accuracy on test dataset: 45.40% (Top-1), 76.21% (Top-5) **Full Precision model (FX Graph Mode quantized)** Evaluation accuracy on test dataset: 69.76% (Top-1), 89.08% (Top-5) **Eager mode quantized model** Evaluation accuracy on test dataset: 69.49% (Top-1), 88.90% (Top-5) Pull Request resolved: #81040 Approved by: https://github.com/jerryzh168 Test Plan: contbuild & OSS CI, see https://hud.pytorch.org/commit/pytorch/pytorch/13ad4739a6e9402e2039a1ce521b9aed595760b3 Reviewed By: osalpekar Differential Revision: D38252390 Pulled By: asl3 fbshipit-source-id: 86ff2f3928fb1fc2b57867d6abcac998d17306e4
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stack from ghstack (oldest at bottom):
Summary:
This PR implements PTQ for APoT FakeQuant. It runs models (Resnet-18 pre-trained model, ImageNet dataset) to compare accuracy metrics for different qconfig settings of uniform vs. APoT quantized activation and weight.
According to the collected accuracy stats, model #2 (uniform activation and APoT weight) appears to have a slight improvement in accuracy compared to model #1 (uniform activation and uniform weight) for 8-bit and significant improvement for 4-bit (see "Accuracy Stats" section below).
Test Plan:
Run models with:
python test/quantization/core/experimental/fx_graph_mode_apot.pyAccuracy Stats:
8-bit (Uniform int8, APoT b = 8 k = 2)
Model #1: Uniform activation, uniform weight (FX Graph Mode quantized)
Evaluation accuracy on test dataset: 64.43% (Top-1), 85.62% (Top-5)
Model #2: Uniform activation, APoT weight (FX Graph Mode quantized)
Evaluation accuracy on test dataset: 64.51% (Top-1), 85.78% (Top-5)
Model #3: APoT activation, APoT weight (FX Graph Mode quantized)
Evaluation accuracy on test dataset: 64.32% (Top-1), 85.78% (Top-5)
4-bit (Uniform int4, APoT b = 4 k = 2)
Model #1: Uniform activation, uniform weight (FX Graph Mode quantized)
Evaluation accuracy on test dataset: 45.63% (Top-1), 71.96% (Top-5)
Model #2: Uniform activation, APoT weight (FX Graph Mode quantized)
Evaluation accuracy on test dataset: 64.24% (Top-1), 85.56% (Top-5)
Model #3: APoT activation, APoT weight (FX Graph Mode quantized)
Evaluation accuracy on test dataset: 45.40% (Top-1), 76.21% (Top-5)
Full Precision model (FX Graph Mode quantized)
Evaluation accuracy on test dataset: 69.76% (Top-1), 89.08% (Top-5)
Eager mode quantized model
Evaluation accuracy on test dataset: 69.49% (Top-1), 88.90% (Top-5)