Model card for davit-base-ws4-cav2-in1k-128

DaViT-Base (ws=4, channel_attn_v2) trained from scratch on ImageNet-1K @128x128.

Training Recipe:

• DaViT supervised recipe -- Table 5.1 (https://arxiv.org/pdf/2204.03645)
• BF16 AMP

Model Details

• Model: davit_small
• Model Kwargs:
  • channel_attn_v2: True
  • window_size: 4
• Input Size:
  • 3x128x128
• Top-1 Accuracy: 82.16999997802735

Model Usage

import timm
model = timm.create_model('hf-hub:PRadecki/davit-base-ws4-cav2-in1k-128', pretrained=True)
model.eval()