Improve test crop for ViT models. Small now 77.85, added base weights at 79.35 top-1.

Author: rwightman

Diff for: README.md

@@ -2,6 +2,9 @@
 
 ## What's New
 
+### Oct 21, 2020
+* Weights added for Vision Transformer (ViT) models. 77.86 top-1 for 'small' and 79.35 for 'base'. Thanks to [Christof](https://www.kaggle.com/christofhenkel) for training the base model w/ lots of GPUs.
+
 ### Oct 13, 2020
 * Initial impl of Vision Transformer models. Both patch and hybrid (CNN backbone) variants. Currently trying to train...
 * Adafactor and AdaHessian (FP32 only, no AMP) optimizers

Diff for: timm/models/vision_transformer.py

@@ -39,7 +39,7 @@ def _cfg(url='', **kwargs):
     return {
         'url': url,
         'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': None,
-        'crop_pct': 1.0, 'interpolation': 'bicubic',
+        'crop_pct': .9, 'interpolation': 'bicubic',
         'mean': IMAGENET_DEFAULT_MEAN, 'std': IMAGENET_DEFAULT_STD,
         'first_conv': '', 'classifier': 'head',
         **kwargs
@@ -51,7 +51,9 @@ def _cfg(url='', **kwargs):
     'vit_small_patch16_224': _cfg(
         url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/vit_small_p16_224-15ec54c9.pth',
     ),
-    'vit_base_patch16_224': _cfg(),
+    'vit_base_patch16_224': _cfg(
+        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/vit_base_p16_224-4e355ebd.pth'
+    ),
     'vit_base_patch16_384': _cfg(input_size=(3, 384, 384)),
     'vit_base_patch32_384': _cfg(input_size=(3, 384, 384)),
     'vit_large_patch16_224': _cfg(),
@@ -283,6 +285,9 @@ def vit_small_patch16_224(pretrained=False, **kwargs):
 def vit_base_patch16_224(pretrained=False, **kwargs):
     model = VisionTransformer(patch_size=16, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4, **kwargs)
     model.default_cfg = default_cfgs['vit_base_patch16_224']
+    if pretrained:
+        load_pretrained(
+            model, num_classes=kwargs.get('num_classes', 0), in_chans=kwargs.get('in_chans', 3))
     return model