Usage

• Define a custom GCViT configuration.

from gcvit_tensorflow import GCViT

# Define a custom GCViT configuration
model = GCViT(
    depths=[2, 2, 6, 2],
    num_heads=[2, 4, 8, 16],
    window_size=[7, 7, 14, 7],
    dim=64,
    resolution=224,
    in_chans=3,
    mlp_ratio=3,
    drop_path_rate=0.2,
    data_format="channels_last",
    num_classes=100,
    classifier_activation="softmax",
)

• Use a predefined GCViT configuration.

from gcvit_tensorflow import GCViT

model = GCViT(configuration="xxtiny")
model.build((None, 224, 224, 3))
print(model.summary())

Model: "xxtiny"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 patch_embed (PatchEmbed)    (None, 56, 56, 64)        45632     
                                                                 
 pos_drop (Dropout)          (None, 56, 56, 64)        0         
                                                                 
 levels/0 (GCViTLayer)       (None, 28, 28, 128)       185766    
                                                                 
 levels/1 (GCViTLayer)       (None, 14, 14, 256)       693258    
                                                                 
 levels/2 (GCViTLayer)       (None, 7, 7, 512)         5401104   
                                                                 
 levels/3 (GCViTLayer)       (None, 7, 7, 512)         5400546   
                                                                 
 norm (LayerNorm_)           (None, 7, 7, 512)         1024      
                                                                 
 avgpool (AdaptiveAveragePoo  (None, 512, 1, 1)        0         
 ling2D)                                                         
                                                                 
 head (Linear_)              (None, 1000)              513000    
                                                                 
=================================================================
Total params: 12,240,330
Trainable params: 11,995,428
Non-trainable params: 244,902
_________________________________________________________________

• Train from scratch the model.

# Example
model.compile(
    optimizer="sgd",
    loss="sparse_categorical_crossentropy",
    metrics=["accuracy", "sparse_top_k_categorical_accuracy"],
)
model.fit(x, y)

• Use ported ImageNet pretrained weights

# Example
from gcvit_tensorflow import GCViT

model = GCViT(configuration="base", pretrained=True, classifier_activation="softmax")
y_pred = model(image)

@article{hatamizadeh2022global,
  title={Global Context Vision Transformers},
  author={Hatamizadeh, Ali and Yin, Hongxu and Kautz, Jan and Molchanov, Pavlo},
  journal={arXiv preprint arXiv:2206.09959},
  year={2022}
}