Image Classification
timm
PDE
ConvNet
liuyao commited on
Commit
d46ec21
1 Parent(s): 0edeeb0

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +6 -4
README.md CHANGED
@@ -10,19 +10,21 @@ library_name: timm
10
 
11
  # Model Card for Model ID
12
 
13
- <!-- Provide a quick summary of what the model is/does. -->
14
 
15
  This modelcard aims to be a base template for new models. It has been generated using [this raw template](https://github.com/huggingface/huggingface_hub/blob/main/src/huggingface_hub/templates/modelcard_template.md?plain=1).
16
 
 
 
17
  ## Model Details
18
 
19
  ### Model Description
20
 
21
- <!-- Provide a longer summary of what this model is. -->
22
 
23
 
24
 
25
- - **Developed by:** [More Information Needed]
26
  - **Funded by [optional]:** [More Information Needed]
27
  - **Shared by [optional]:** [More Information Needed]
28
  - **Model type:** [More Information Needed]
@@ -35,7 +37,7 @@ This modelcard aims to be a base template for new models. It has been generated
35
  <!-- Provide the basic links for the model. -->
36
 
37
  - **Repository:** [More Information Needed]
38
- - **Paper [optional]:** [More Information Needed]
39
  - **Demo [optional]:** [More Information Needed]
40
 
41
  ## Uses
 
10
 
11
  # Model Card for Model ID
12
 
13
+ Based on quaslinear hyperbolic systems of PDEs, the QLNet explores a new model space for ConvNets that uses multiplication (of same-sized tensors) instead of ReLU as the nonlinearity. It achieves comparable accuracy as ResNet50 on ImageNet-1k, demonstrating that it has the same level of capacity/expressivity, and deserves more study (hyper-paremeter tuning) that I alone am not able to do.
14
 
15
  This modelcard aims to be a base template for new models. It has been generated using [this raw template](https://github.com/huggingface/huggingface_hub/blob/main/src/huggingface_hub/templates/modelcard_template.md?plain=1).
16
 
17
+
18
+
19
  ## Model Details
20
 
21
  ### Model Description
22
 
23
+ Instead of the bottleneck of ResNet50 which consists of 1x1, 3x3, 1x1 in succession, we instead of make the 1x1, split into two equal halves and multiply them, then apply a 3x3 (depthwise), and a 1x1. All without activation functions except at the end of the block, where we apply a *radial activation function* that I call `hardball`.
24
 
25
 
26
 
27
+ - **Developed by:** Yao Liu 刘杳
28
  - **Funded by [optional]:** [More Information Needed]
29
  - **Shared by [optional]:** [More Information Needed]
30
  - **Model type:** [More Information Needed]
 
37
  <!-- Provide the basic links for the model. -->
38
 
39
  - **Repository:** [More Information Needed]
40
+ - **Paper [optional]:** [A Novel ConvNet Architecture with a Continuous Symmetry](https://arxiv.org/abs/2308.01621)
41
  - **Demo [optional]:** [More Information Needed]
42
 
43
  ## Uses