Skip to content

Commit 34795fd

Browse files
bump md-dead-link-check to 1.1.0 (#3219)
### Changes Bump md-dead-link-check to 1.1.0 Fix link docs/usage/training_time_compression/other_algorithms/LegacyQuantization.md
1 parent 5ad9bc4 commit 34795fd

File tree

3 files changed

+3
-3
lines changed

3 files changed

+3
-3
lines changed

.github/workflows/nightly.yml

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -21,7 +21,7 @@ jobs:
2121
timeout-minutes: 10
2222
steps:
2323
- uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29 # v4.1.6
24-
- uses: AlexanderDokuchaev/md-dead-link-check@cc3ed55268899a1a6d5fd7068abbc4591eab1f74 # v0.9
24+
- uses: AlexanderDokuchaev/md-dead-link-check@c7210ef8a38c194a119834e39d212387d19b512c # v1.1.0
2525

2626
tensorflow:
2727
runs-on: ubuntu-20.04-8-cores

.github/workflows/pre-commit-linters.yml

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -24,5 +24,5 @@ jobs:
2424
runs-on: ubuntu-22.04
2525
steps:
2626
- uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29 # v4.1.6
27-
- uses: AlexanderDokuchaev/md-dead-link-check@cc3ed55268899a1a6d5fd7068abbc4591eab1f74 # v0.9
27+
- uses: AlexanderDokuchaev/md-dead-link-check@c7210ef8a38c194a119834e39d212387d19b512c # v1.1.0
2828

docs/usage/training_time_compression/other_algorithms/LegacyQuantization.md

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -274,7 +274,7 @@ file.
274274

275275
### AutoQ
276276

277-
NNCF provides an alternate mode, namely AutoQ, for mixed-precision automation. It is an AutoML-based technique that automatically learns the layer-wise bitwidth with explored experiences. Based on [HAQ](https://openaccess.thecvf.com/content_CVPR_2019/papers/Wang_HAQ_Hardware-Aware_Automated_Quantization_With_Mixed_Precision_CVPR_2019_paper.pdf), AutoQ utilizes an actor-critic algorithm, Deep Deterministic Policy Gradient (DDPG) for efficient search over the bitwidth space. DDPG is trained in an episodic fashion, converging to a deterministic mixed-precision policy after a number of episodes. An episode is constituted by stepping, the DDPG transitions from quantizer to quantizer sequentially to predict a precision of a layer. Each quantizer essentially denotes a state in RL framework and it is represented by attributes of the associated layers. For example, a quantizer for 2D Convolution is represented by its quantizer Id (integer), input and output channel size, feature map dimension, stride size, if it is depthwise, number of parameters etc. It is recommended to check out ```_get_layer_attr``` in [```quantization_env.py```](https://github.com/openvinotoolkit/nncf/blob/develop/nncf/automl/environment/quantization_env.py#L333) for the featurization of different network layer types.
277+
NNCF provides an alternate mode, namely AutoQ, for mixed-precision automation. It is an AutoML-based technique that automatically learns the layer-wise bitwidth with explored experiences. Based on [HAQ](https://openaccess.thecvf.com/content_CVPR_2019/papers/Wang_HAQ_Hardware-Aware_Automated_Quantization_With_Mixed_Precision_CVPR_2019_paper.pdf), AutoQ utilizes an actor-critic algorithm, Deep Deterministic Policy Gradient (DDPG) for efficient search over the bitwidth space. DDPG is trained in an episodic fashion, converging to a deterministic mixed-precision policy after a number of episodes. An episode is constituted by stepping, the DDPG transitions from quantizer to quantizer sequentially to predict a precision of a layer. Each quantizer essentially denotes a state in RL framework and it is represented by attributes of the associated layers. For example, a quantizer for 2D Convolution is represented by its quantizer Id (integer), input and output channel size, feature map dimension, stride size, if it is depthwise, number of parameters etc. It is recommended to check out `_get_layer_attr` in [quantization_env.py](/nncf/torch/automl/environment/quantization_env.py#L370) for the featurization of different network layer types.
278278

279279
When the agent enters a state/quantizer, it receives the state features and forward passes them through its network. The output of the forward pass is a scalar continuous action output which is subsequently mapped to the bitwidth options of the particular quantizer. The episode terminates after the prediction of the last quantizer and a complete layer-wise mixed-precision policy is obtained. To ensure a policy fits in the user-specified compression ratio, the policy is post processed by reducing the precision sequentially from the last quantizer until the compression ratio is met.
280280

0 commit comments

Comments
 (0)