🚪 Hatch — 8hr 自主研究報告 (v20260505)

import torch, timm, torch.nn as nn

class GenericClassifier(nn.Module):
    def __init__(self, backbone, n_attr, feat_dim):
        super().__init__()
        self.backbone = timm.create_model(backbone, pretrained=False, num_classes=0, global_pool="avg")
        self.dropout = nn.Dropout(0.3)
        self.cls = nn.Linear(feat_dim, n_attr)
    def forward(self, x): return self.cls(self.dropout(self.backbone(x)))

ckpt = torch.load("hatch_swa_v20260505.pt", weights_only=False)
# backbone_name = "convnext_tiny.fb_in1k", attrs = ["has_open", "has_close"]
# feat_dim = 768, img_size = 384
# thresholds = {"has_open": 0.50, "has_close": 0.76}（best.pt） 或 0.52/0.76（best_tta.pt）
model = GenericClassifier(ckpt["backbone_name"], len(ckpt["attrs"]), ckpt["feat_dim"]).eval()
model.load_state_dict(ckpt["model_state"])

# 推論：384×384 → ImageNet normalize → sigmoid 2 outputs
# TTA (best_tta.pt 適用): 加 hflip 一次平均