Probe Ade20k 40k S512 C12 In21k

Usage

uv add "canvit-pytorch @ git+https://github.com/m2b3/CanViT-PyTorch.git"

import torch
from canvit_pytorch.probes import SegmentationProbe

probe = SegmentationProbe.from_pretrained("canvit/probe-ade20k-40k-s512-c12-in21k").eval()

# [B, H, W, D] canvas features from a CanViT forward pass
features = torch.randn(1, 12, 12, 1024)
with torch.inference_mode():
    logits = probe(features)    # [B, num_classes, H, W]
assert logits.shape == (1, 150, 12, 12)

Training

Architecture: LayerNorm → Dropout → BatchNorm → Conv1×1.

--- license: mit library_name: canvit-pytorch pipeline_tag: image-segmentation tags: - canvit - active-vision - ade20k - segmentation-probe datasets: - scene_parse_150 base_model: canvit/canvitb16-add-vpe-pretrain-g128px-s512px-in21k-dv3b16-2026-02-02 --- # ADE20K Segmentation Probe — canvas 12×12 @ 512px scene Linear segmentation probe on the canvas features of [canvit/canvitb16-add-vpe-pretrain-g128px-s512px-in21k-dv3b16-2026-02-02](https://huggingface.co/canvit/canvitb16-add-vpe-pretrain-g128px-s512px-in21k-dv3b16-2026-02-02). - **Paper**: [arXiv:2603.22570](https://arxiv.org/abs/2603.22570) - **Training code**: [github.com/m2b3/CanViT-specialize](https://github.com/m2b3/CanViT-specialize) ## Usage ```bash uv add "canvit-pytorch @ git+https://github.com/m2b3/CanViT-PyTorch.git" ``` ```python import torch from canvit_pytorch.probes import SegmentationProbe probe = SegmentationProbe.from_pretrained("canvit/probe-ade20k-40k-s512-c12-in21k").eval() # [B, H, W, D] canvas features from a CanViT forward pass features = torch.randn(1, 12, 12, 1024) with torch.inference_mode(): logits = probe(features) # [B, num_classes, H, W] assert logits.shape == (1, 150, 12, 12) ``` ## Training Architecture: `LayerNorm → Dropout → BatchNorm → Conv1×1`. | Hyperparameter | Value | |---|---| | Scene size | 512 px | | Canvas grid | 12 × 12 | | Glimpse size | 128 px | | Timesteps (T) | 10 | | Training policy | R-IID | | Optimizer | AdamW | | Peak LR | \\( 3 \times 10^{-4} \\) | | Weight decay | \\( 10^{-3} \\) | | LR schedule | 1,500-step warmup → cosine decay | | Batch size | 16 | | Max steps | 40,000 | | Dropout | 0.1 | | Augmentation | RandomResizedCrop scale [0.5, 2] + HFlip | | Precision | bf16 (AMP) |