Adds EdgeNeXt-Base model and pretrained weights

It achieves SOTA performance of 83.7% top-1 ImageNet-1K accuracy with just 18.5M parameters and 3.8G MAdds.

README.md

@@ -15,6 +15,10 @@ and [Fahad Shahbaz Khan](https://scholar.google.es/citations?user=zvaeYnUAAAAJ&h
 [![slides](https://img.shields.io/badge/Presentation-Slides-B762C1)](https://mbzuaiac-my.sharepoint.com/:b:/g/personal/muhammad_maaz_mbzuai_ac_ae/EaFA4bSPEMBNlJuHMbKDD3UBHmwXrmpijSRqZITk2l1-wQ?e=b7ruLV)
 
 ## :rocket: News
+* **(Jul 28, 2022)**
+  * EdgeNeXt-B model is released. It achieves 82.5% top-1 ImageNet-1K accuracy with 18.51M parameters and 3.84G MAdds.
+  Further, using USI (https://arxiv.org/abs/2204.03475) training recipe, the same model achieves 83.7% accuracy.
+
 * **(Jun 28, 2022)**
   * EdgeNeXt-S model trained using USI (https://arxiv.org/abs/2204.03475) is released. 
   It achieves 81.1% top-1 ImageNet-1K accuracy with only 5.59M parameters and 1.26G MAdds.
@@ -32,6 +36,8 @@ and [Fahad Shahbaz Khan](https://scholar.google.es/citations?user=zvaeYnUAAAAJ&h
 
 | Name |Acc@1 | #Params | MAdds | Model |
 |---|:---:|:---:| :---:|:---:|
+| edgenext_base_usi | 83.68 | 18.51M | 3.84G  | [model](https://github.com/mmaaz60/EdgeNeXt/releases/download/v1.2/edgenext_base_usi.pth)
+| edgenext_base | 82.47 | 18.51M | 3.84G | [model](https://github.com/mmaaz60/EdgeNeXt/releases/download/v1.2/edgenext_base.pth)
 | edgenext_small_usi | 81.07 | 5.59M | 1.26G | [model](https://github.com/mmaaz60/EdgeNeXt/releases/download/v1.1/edgenext_small_usi.pth)
 | edgenext_small | 79.41 | 5.59M | 1.26G | [model](https://github.com/mmaaz60/EdgeNeXt/releases/download/v1.0/edgenext_small.pth)
 | edgenext_x_small | 74.96 | 2.34M | 538M | [model](https://github.com/mmaaz60/EdgeNeXt/releases/download/v1.0/edgenext_x_small.pth)
@@ -102,7 +108,7 @@ This should give,
 Acc@1 79.412 Acc@5 94.512 loss 0.881
 ```
 
-###### Note: For evaluating the USI model, please set `usi_eval True`.
+##### Note: For evaluating the USI model, please set `usi_eval True`.
 
 <hr />
 

models/model.py

@@ -66,6 +66,24 @@ def edgenext_small(pretrained=False, **kwargs):
     return model
 
 
+@register_model
+def edgenext_base(pretrained=False, **kwargs):
+    # 18.51M & 3840.93M @ 256 resolution
+    # 82.5% (normal) 83.7% (USI) Top-1 accuracy
+    # AA=True, Mixup & Cutmix, DropPath=0.1, BS=4096, lr=0.006, multi-scale-sampler
+    # Jetson FPS=xx.xx versus xx.xx for MobileViT_S
+    # For A100: FPS @ BS=1: xxx.xx & @ BS=256: xxxx.xx
+    model = EdgeNeXt(depths=[3, 3, 9, 3], dims=[80, 160, 288, 584], expan_ratio=4,
+                     global_block=[0, 1, 1, 1],
+                     global_block_type=['None', 'SDTA', 'SDTA', 'SDTA'],
+                     use_pos_embd_xca=[False, True, False, False],
+                     kernel_sizes=[3, 5, 7, 9],
+                     d2_scales=[2, 2, 3, 4],
+                     **kwargs)
+
+    return model
+
+
 """
     Using BN & HSwish instead of LN & GeLU
 """