🔥CNN Structured Pruning Leaderboard
Welcome to our dedicated site for the survey paper: "Structured Pruning for Deep Convolutional Neural Networks: A Survey". Our survey is accepted by IEEE T-PAMI. Links include arXiv and IEEE Xplore.
Github Repo: Awesome Pruning: A curated list of neural network pruning resources.
This platform serves as a repository and visual representation of the benchmarks from studies covered in our survey.
Here, you can explore the reported accuracy and FLOPs metrics from various papers, providing an at-a-glance view of the advancements and methodologies in the domain of structured pruning.
If you find this website helpful, please consider citing our paper 😊
Guide to use this leaderboard
Sections
We divide the webpage into below sections:
- Dataset Tabs
- Query Section
- Data Plotting
- Data Table
More detailed functions are explained in the following sections.
Dataset Tabs
- Click the corresponding tabs to view the results of different datasets.
- We currently support three datasets: CIFAR-10, CIFAR-100, and ImageNet-1K.
- Results are 'isolated' for each dataset, i.e., the results of different datasets are not mixed together.
Query Section
The query box includes two parts
- red box: query by paper attributes
- blue box: query by experimental results
Press [Enter] key to update.
- update both plotting and table.
Example: Here, we provide a use case and show how query works.
If a user wants to find methods that satisfy the followings:
- Select Dataset: ImageNet-1K
- Select Model: ResNet-50
- Select Pruning Method: Regularization-based Pruning
- Target 1: Accuracy after pruning > 75%
- Target 2: Pruned FLOPs > 40%
- Target 3: Model size after pruning < 30M
By entering the requirements to the corresponding query box, we can narrow down the results and compare the remaining ones.
Data Plotting
The data plotting section can be split into two parts:
- red box: contains two radio buttons to select:
- (left) Group colors by ‘model’, ‘section’, or ‘year’.
- (right) Change x-axis of the plots to ‘FLOPs drop (%)’ or ‘FLOPs after pruning (M)’.
- blue box: interactive plots
Group by Model (default)
X-axis: FLOPs drop (%) (default)
Group by Section
X-axis: FLOPs drop (%) (default)
Group by Year
X-axis: FLOPs drop (%) (default)
Group by Model (default)
X-axis: FLOPs after pruning (M)
Default Figure
- Shift the graph by dragging.
- Zoom-in/out by scrolling.
- Hover over the data point to see the details.
- Click any legend to filter out others.
- Click any legend to filter out others.
- Click white spaces/Double Click to restore to default scaling and legends.
Data Table
Click to the expand the table
- The expanded table contains the baseline FLOPs and Parameters for each model.
Click the sort button:
- Sort in ascending order.
- click more than once to toggle ascending/descending.
Click any method name (highlighted in the red box) to show details of the paper (blue box).
The details include:
- detailed section
- link of paper
- venue of publication
- released code (if any)
- the BibTex used in our paper
Search by below options:
Section | Year | Method | Model | Acc | Acc Pruned | Acc ↓ (%) | FLOPs (M) | FLOPs Pruned (M) | FLOPs ↓ (%) | Params (M) | Params Pruned (M) | Params ↓ (%) |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
2.6.1 | 2022 | AutoCompress | PreResNet-101 | 94.81 | 94.85 | -0.04 | 253.15 | 1232.04 | 19.07 | 44.54 | 31.18 | 20.19 |
sec | year | method | model | acc | acc-pruned | acc-change | flops | flops-pruned | flops-drop | param | param-pruned | param-drop | dataset |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
2.6.1 | 2022 | AutoCompress | PreResNet-101 | 94.81 | 94.85 | -0.04 | 253.15 | 1232.04 | 19.07 | 44.54 | 31.18 | 20.19 | cifar10 |
Section | Year | Method | Model | Acc | Acc Pruned | Acc ↓ (%) | FLOPs (M) | FLOPs Pruned (M) | FLOPs ↓ (%) | Params (M) | Params Pruned (M) | Params ↓ (%) | Dataset |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
2.6.1 | 2022 | AutoCompress | PreResNet-101 | 94.81 | 94.85 | -0.04 | 253.15 | 1232.04 | 19.07 | 44.54 | 31.18 | 20.19 | cifar10 |
Search by below options:
Section | Year | Method | Model | Acc | Acc Pruned | Acc ↓ (%) | FLOPs (M) | FLOPs Pruned (M) | FLOPs ↓ (%) | Params (M) | Params Pruned (M) | Params ↓ (%) |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
2.4.2 | 2019 | EarlyCroP | PreResNet-101 | 74.64 | 72.19 | -0.08 | 253.15 | 6329.17 | 28.23 | 44.54 | 31.18 | 37.74 |
sec | year | method | model | acc | acc-pruned | acc-change | flops | flops-pruned | flops-drop | param | param-pruned | param-drop | dataset |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
2.4.2 | 2019 | EarlyCroP | PreResNet-101 | 74.64 | 72.19 | -0.08 | 253.15 | 6329.17 | 28.23 | 44.54 | 31.18 | 37.74 | cifar100 |
Section | Year | Method | Model | Acc | Acc Pruned | Acc ↓ (%) | FLOPs (M) | FLOPs Pruned (M) | FLOPs ↓ (%) | Params (M) | Params Pruned (M) | Params ↓ (%) | Dataset |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
2.4.2 | 2019 | EarlyCroP | PreResNet-101 | 74.64 | 72.19 | -0.08 | 253.15 | 6329.17 | 28.23 | 44.54 | 31.18 | 37.74 | cifar100 |
Search by below options:
Section | Year | Method | Model | Acc | Acc Pruned | Acc ↓ (%) | FLOPs (M) | FLOPs Pruned (M) | FLOPs ↓ (%) | Params (M) | Params Pruned (M) | Params ↓ (%) |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
2.3.3 | 2019 | SuperTickets | ProxylessNet-Mobile | 56.98 | 57.87 | -0.89 | 11514 | 5568.82 | 31.03 | 132.86 | 17.18 | 47.09 |
sec | year | method | model | acc | acc-pruned | acc-change | flops | flops-pruned | flops-drop | param | param-pruned | param-drop | dataset |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
2.3.3 | 2019 | SuperTickets | ProxylessNet-Mobile | 56.98 | 57.87 | -0.89 | 11514 | 5568.82 | 31.03 | 132.86 | 17.18 | 47.09 | imagenet |
Section | Year | Method | Model | Acc | Acc Pruned | Acc ↓ (%) | FLOPs (M) | FLOPs Pruned (M) | FLOPs ↓ (%) | Params (M) | Params Pruned (M) | Params ↓ (%) | Dataset |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
2.3.3 | 2019 | SuperTickets | ProxylessNet-Mobile | 56.98 | 57.87 | -0.89 | 11514 | 5568.82 | 31.03 | 132.86 | 17.18 | 47.09 | imagenet |