CoAtNet-3 backbone (Dai et al., 2021).

Larger variant: blocks_per_stage=(2, 6, 14, 2), dims=(192, 384, 768, 1536), stem_width=192, attn_heads=(24, 48). Approximately 168M parameters, 84.5% ImageNet-1k top-1 at 224×224 (Table 5). With ImageNet-21k or JFT-3B pretraining the paper reports 86.5% / 87.8% at 384x384.