Ambarella/SwinV2

Swin Transformer V2 extends the original Swin Transformer with improved attention scaling and positional encoding, enabling stable training and strong performance on very large and high-resolution vision datasets.

Original paper: Swin Transformer V2: Scaling Up Capacity and Resolution

Swin Transformer V2-Tiny

This model uses the Swin Transformer V2-Tiny variant, a compact hierarchical transformer that applies shifted window self-attention for efficient computation. It is well suited for high-resolution image classification and as a backbone for dense vision tasks such as detection and segmentation.

Model Configuration:

• Reference implementation: torchvision.models.swin_v2_t
• Original Weight: Swin_V2_T_Weights.IMAGENET1K_V1
• Resolution: 3x256x256
• Support Cooper version:
  • Cooper SDK: [2.5.3]
  • Cooper Foundry: [2.2]

Model	Device	compression	Model Link
SwinV2-Tiny	N1-655	Amba_optimized	Model_Link
SwinV2-Tiny	N1-655	Activation_fp16	Model_Link
SwinV2-Tiny	CV7	Amba_optimized	Model_Link
SwinV2-Tiny	CV7	Activation_fp16	Model_Link
SwinV2-Tiny	CV72	Amba_optimized	Model_Link
SwinV2-Tiny	CV72	Activation_fp16	Model_Link
SwinV2-Tiny	CV75	Amba_optimized	Model_Link
SwinV2-Tiny	CV75	Activation_fp16	Model_Link