deepellipse-circleid / code /infer_helpers.py
qinxunhui's picture
Add small artifacts (code, outputs, configs, LICENSE)
519ab98 verified
Raw History Blame Contribute Delete
28.3 kB
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""Inference-time model assembly helpers.
Extracted verbatim from the training-side scripts (``train_test_deepelipse.py``
and ``train_ay_bce.py``) so that the released inference path does not depend on
any training entry point. Nothing in this file trains: these are the
checkpoint-architecture resolution, checkpoint loading and frozen-backbone
feature-extraction utilities that ``reproduce_best_ay.py`` needs in order to
recompute ``Y`` and the style-affinity adjacency ``A`` from released weights.
"""
from __future__ import annotations
import argparse
from pathlib import Path
import os
from collections import OrderedDict, defaultdict
from typing import List, Optional, Tuple
import torch
import torch.nn as nn
from tqdm import tqdm
from ay_bce_head import PairwiseBCEHead
from convnext_point_feature import default_convnext_checkpoint
from deep_elipse_model import DeepEllipseConfig, DeepEllipseModel
from resnet_point_feature import default_resnet_checkpoint
class TrainConfig:
"""Inference-time stand-in for the training-side config.
The extracted helpers read exactly one attribute off the original training
config object -- the target device -- so a same-named placeholder keeps the
copied function bodies identical to their originals.
"""
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
def unpack_dataloader_batch(input_all):
"""exp6_data_loader: (fea_tensor, ske_tensor, img_tensor[, cwt_scalo])。"""
fea_tensor, ske_tensor, img_tensor = input_all[:3]
cwt_scalo = input_all[3] if len(input_all) > 3 else None
return fea_tensor, ske_tensor, img_tensor, cwt_scalo
def normalize_state_dict_keys(state_dict: dict) -> OrderedDict:
"""去掉 DataParallel 的 module. 前缀。"""
out = OrderedDict()
for k, v in state_dict.items():
nk = k[7:] if k.startswith("module.") else k
out[nk] = v
return out
def detect_enc43_backbone(state_dict: dict) -> str:
keys = list(state_dict.keys())
if any("enc43.pyramid." in k for k in keys):
return "convnext1d"
if any("enc43.ms_stem." in k for k in keys):
return "multiscale_stem"
return "unknown"
def peek_reload_checkpoint(pretrained_path: str):
raw = torch.load(pretrained_path, map_location="cpu", weights_only=False)
if isinstance(raw, dict) and "model_state_dict" in raw:
sd = normalize_state_dict_keys(raw["model_state_dict"])
meta_cfg = raw.get("config", {}) or {}
elif isinstance(raw, dict):
sd = normalize_state_dict_keys(raw)
meta_cfg = {}
else:
sd = normalize_state_dict_keys(raw)
meta_cfg = {}
return sd, meta_cfg, detect_enc43_backbone(sd)
def detect_architecture_from_state_dict(state_dict: dict) -> dict:
"""从权重键名推断训练时的 §4.3 / CWT 架构。"""
keys = list(state_dict.keys())
arch = {
"use_scheme_v3_1": False,
"use_scheme_v3": False,
"v3_1_stage_pool": "dbsa",
"cwt_scale_affinity": False,
"cwt_scale_index_mode": "uniform",
"angular_norm": "sum_relu",
"e3_cross_query": "cwt_cnn",
"e3_cross_query_path2": "maxpool",
"e3_cwt_branch": True,
"e2_cwt_branch": False,
"cwt_phi_valid_emb": 0,
"e2_cross_path2": "orig_cnn",
"phi_tx_layers": 0,
"gcc_enabled": False,
}
if any("enc43.segment_pool.gcc." in k for k in keys):
arch["gcc_enabled"] = True
if any("enc43.pyramid." in k for k in keys):
arch["use_scheme_v3_1"] = True
elif any("enc43.ms_stem." in k for k in keys):
return arch
if any("enc43.segment_pool.cwt_pyramid." in k for k in keys):
arch["v3_1_stage_pool"] = "cwt_dual"
elif any(k.startswith("cwt_encoder.") for k in keys):
arch["v3_1_stage_pool"] = "cwt_angular"
elif any("enc43.segment_pool.pools." in k for k in keys):
arch["v3_1_stage_pool"] = "cwt_angular"
elif any("enc43.segment_pool.fine_pool." in k for k in keys):
arch["v3_1_stage_pool"] = "dbsa"
elif any("enc43.segment_pool.gates." in k for k in keys):
arch["v3_1_stage_pool"] = "seg_sigmoid"
elif arch["use_scheme_v3_1"] and not any("enc43.segment_pool." in k for k in keys):
arch["v3_1_stage_pool"] = "global"
if any("enc43.segment_pool.stage_phi_enc." in k for k in keys):
arch["cwt_scale_affinity"] = True
if any("enc43.segment_pool.stage_phi_enc.valid_emb." in k for k in keys):
arch["cwt_phi_valid_emb"] = 8
if any("enc43.segment_pool.phi_transformers." in k for k in keys):
arch["phi_tx_layers"] = 2
if arch["v3_1_stage_pool"] == "cwt_angular" and any(
"enc43.segment_pool.cwt_pyramid." in k for k in keys
):
arch["e2_cwt_branch"] = True
if any("enc43.segment_pool.path2_pools." in k for k in keys):
arch["e2_cross_path2"] = "orig_cnn"
if arch["v3_1_stage_pool"] == "cwt_dual":
arch["e3_cwt_branch"] = any(
"enc43.segment_pool.cwt_pyramid." in k for k in keys
)
if any("enc43.segment_pool.path2_pools." in k for k in keys):
arch["e3_cross_query_path2"] = "orig_cnn"
q_w_key = "enc43.segment_pool.pools.0.q_proj.weight"
if q_w_key in state_dict:
in_dim = state_dict[q_w_key].shape[1]
arch["e3_cross_query"] = "phi" if in_dim == 64 else "cwt_cnn"
return arch
def _merge_checkpoint_arch(meta_cfg: dict, state_dict: dict) -> dict:
inferred = detect_architecture_from_state_dict(state_dict)
merged = dict(inferred)
if meta_cfg.get("use_scheme_v3_1"):
merged["use_scheme_v3_1"] = True
if meta_cfg.get("use_scheme_v3"):
merged["use_scheme_v3"] = True
if meta_cfg.get("v3_1_stage_pool"):
merged["v3_1_stage_pool"] = meta_cfg["v3_1_stage_pool"]
if meta_cfg.get("cwt_scale_affinity"):
merged["cwt_scale_affinity"] = bool(meta_cfg["cwt_scale_affinity"])
if meta_cfg.get("cwt_scale_index_mode"):
merged["cwt_scale_index_mode"] = meta_cfg["cwt_scale_index_mode"]
if meta_cfg.get("angular_norm"):
merged["angular_norm"] = meta_cfg["angular_norm"]
if meta_cfg.get("e3_cross_query"):
merged["e3_cross_query"] = meta_cfg["e3_cross_query"]
if meta_cfg.get("e3_cross_query_path2"):
merged["e3_cross_query_path2"] = meta_cfg["e3_cross_query_path2"]
if "e3_cwt_branch" in meta_cfg:
merged["e3_cwt_branch"] = bool(meta_cfg["e3_cwt_branch"])
if "e2_cwt_branch" in meta_cfg:
merged["e2_cwt_branch"] = bool(meta_cfg["e2_cwt_branch"])
if meta_cfg.get("cwt_phi_valid_emb"):
merged["cwt_phi_valid_emb"] = int(meta_cfg["cwt_phi_valid_emb"])
if meta_cfg.get("e2_cross_path2"):
merged["e2_cross_path2"] = meta_cfg["e2_cross_path2"]
if meta_cfg.get("phi_tx_layers") is not None:
merged["phi_tx_layers"] = int(meta_cfg["phi_tx_layers"])
if meta_cfg.get("gcc_enabled") is not None:
merged["gcc_enabled"] = bool(meta_cfg["gcc_enabled"])
return merged
def align_args_with_checkpoint(args, parser, ckpt_path: str, *, purpose: str) -> None:
"""在 build_elipse_model 之前,按 checkpoint 对齐 scheme / stage_pool。"""
if not ckpt_path or not os.path.isfile(ckpt_path):
return
sd, ckpt_cfg, ckpt_backbone = peek_reload_checkpoint(ckpt_path)
arch = _merge_checkpoint_arch(ckpt_cfg, sd)
user_scheme = args.scheme_v2 or args.scheme_v3 or args.scheme_v3_1
if purpose in ("test", "resume"):
if arch["use_scheme_v3_1"]:
args.scheme_v3_1 = True
args.scheme_v2 = False
args.scheme_v3 = False
elif arch["use_scheme_v3"]:
args.scheme_v3 = True
args.scheme_v2 = False
args.scheme_v3_1 = False
args.v3_1_stage_pool = arch["v3_1_stage_pool"]
args.cwt_scale_affinity = arch["cwt_scale_affinity"]
args.gcc_enabled = arch.get("gcc_enabled", False)
if hasattr(args, "cwt_scale_index_mode"):
args.cwt_scale_index_mode = arch.get("cwt_scale_index_mode", "uniform")
if hasattr(args, "angular_norm"):
args.angular_norm = arch.get("angular_norm", "sum_relu")
if hasattr(args, "e3_cross_query"):
args.e3_cross_query = arch.get("e3_cross_query", "cwt_cnn")
if hasattr(args, "e3_cross_query_path2"):
args.e3_cross_query_path2 = arch.get("e3_cross_query_path2", "maxpool")
if hasattr(args, "e3_no_cwt_branch"):
args.e3_no_cwt_branch = not arch.get("e3_cwt_branch", True)
if hasattr(args, "e2_cwt_branch"):
args.e2_cwt_branch = arch.get("e2_cwt_branch", False)
if hasattr(args, "e2_valid_phi"):
args.e2_valid_phi = arch.get("cwt_phi_valid_emb", 0) > 0
if hasattr(args, "e2_phi_tx_layers"):
args.e2_phi_tx_layers = arch.get("phi_tx_layers", 0)
if hasattr(args, "e2_cross_path2"):
args.e2_cross_path2 = arch.get("e2_cross_path2", "orig_cnn")
print(
f"📦 checkpoint 对齐架构 ({purpose}): "
f"scheme_v3_1={args.scheme_v3_1}, pool={args.v3_1_stage_pool}, "
f"cwt_scale_affinity={args.cwt_scale_affinity}, "
f"cwt_scale_index_mode={arch.get('cwt_scale_index_mode', 'uniform')}, "
f"angular_norm={arch.get('angular_norm', 'sum_relu')}, "
f"e3_cross_query={arch.get('e3_cross_query', 'cwt_cnn')}, "
f"e3_cross_query_path2={arch.get('e3_cross_query_path2', 'maxpool')}, "
f"e3_cwt_branch={arch.get('e3_cwt_branch', True)}, "
f"e2_cwt_branch={arch.get('e2_cwt_branch', False)}, "
f"cwt_phi_valid_emb={arch.get('cwt_phi_valid_emb', 0)}, "
f"phi_tx_layers={arch.get('phi_tx_layers', 0)}"
)
return
if ckpt_backbone == "convnext1d" and not user_scheme:
print("⚠️ checkpoint 含 enc43.pyramid → 自动启用 --scheme-v3-1")
args.scheme_v3_1 = True
if ckpt_backbone == "convnext1d" and not args.scheme_v3_1 and not args.scheme_v3:
parser.error("checkpoint 为 ConvNeXt1D,须加 --scheme-v3-1 或 --scheme-v3")
if ckpt_backbone == "multiscale_stem" and (args.scheme_v3_1 or args.scheme_v3):
parser.error("checkpoint 为 v1 multiscale_stem,不应加 --scheme-v3-1")
def load_partial_weights(model, pretrained_path, verbose_ignore: bool = False):
"""
从预训练模型加载权重,自动跳过形状不匹配的层。
"""
print(f"load_partial_weights: {pretrained_path}")
raw = torch.load(pretrained_path, map_location=TrainConfig.DEVICE, weights_only=False)
if isinstance(raw, dict) and "model_state_dict" in raw:
pretrained_dict = normalize_state_dict_keys(raw["model_state_dict"])
elif isinstance(raw, dict):
pretrained_dict = normalize_state_dict_keys(raw)
else:
pretrained_dict = normalize_state_dict_keys(raw)
model_dict = model.state_dict()
ckpt_backbone = detect_enc43_backbone(pretrained_dict)
model_backbone = detect_enc43_backbone(model_dict)
print(f"checkpoint §4.3 骨干: {ckpt_backbone} | 当前模型: {model_backbone}")
filtered_dict = OrderedDict()
matched_count = 0
skipped_count = 0
ignored_count = 0
ignored_groups: defaultdict = defaultdict(int)
print("--- 开始权重过滤与加载 ---")
for k, v in pretrained_dict.items():
if k in model_dict:
if v.shape == model_dict[k].shape:
filtered_dict[k] = v
matched_count += 1
else:
skipped_count += 1
print(f"[SKIP] 形状不匹配: {k}")
print(f" 旧权重形状: {v.shape}")
print(f" 新模型形状: {model_dict[k].shape}")
if k == "enc41.proj_deep.weight":
print(" (切换 deep_backend 时 proj_deep 维变化属预期)")
else:
ignored_count += 1
prefix = ".".join(k.split(".")[:3])
ignored_groups[prefix] += 1
if verbose_ignore:
print(f"[IGNORE] 新模型中不存在该层: {k}")
model_dict.update(filtered_dict)
model.load_state_dict(model_dict, strict=False)
def _count_loaded(tag: str) -> int:
return sum(1 for k in filtered_dict if tag in k)
pyramid_ckpt = sum(1 for k in pretrained_dict if "enc43.pyramid." in k)
pyramid_loaded = _count_loaded("enc43.pyramid.")
stage_proj_ckpt = sum(1 for k in pretrained_dict if "enc43.stage_proj." in k)
stage_proj_loaded = _count_loaded("enc43.stage_proj.")
cwt_ckpt = sum(1 for k in pretrained_dict if k.startswith("cwt_encoder."))
cwt_loaded = _count_loaded("cwt_encoder.")
print("-----------------------------")
print(f"成功加载层数: {matched_count}")
print(f"跳过层数 (形状不符): {skipped_count}")
print(f"忽略层数 (新模型无对应键): {ignored_count}")
if ignored_groups and not verbose_ignore:
top = sorted(ignored_groups.items(), key=lambda x: -x[1])[:8]
print("忽略键前缀 Top:", ", ".join(f"{p}({n})" for p, n in top))
if pyramid_ckpt:
print(f"enc43.pyramid: {pyramid_loaded}/{pyramid_ckpt} 层已加载")
if stage_proj_ckpt:
print(f"enc43.stage_proj: {stage_proj_loaded}/{stage_proj_ckpt} 层已加载")
if cwt_ckpt or cwt_loaded:
print(f"cwt_encoder: {cwt_loaded}/{cwt_ckpt} 层已加载(CWT 头通常重新初始化)")
print("权重加载完成。")
if ckpt_backbone == "convnext1d" and model_backbone != "convnext1d":
raise ValueError(
"checkpoint 含 enc43.pyramid,但当前模型不是 v3/v3.1 ConvNeXt1D。\n"
"请在训练命令中加 --scheme-v3-1(CWT 微调示例:\n"
" --scheme-v3-1 --v3-1-stage-pool cwt_angular --reload_model <ckpt>)"
)
if pyramid_ckpt > 0 and pyramid_loaded == 0:
raise ValueError(
"enc43.pyramid 权重一层都未加载。请确认已加 --scheme-v3-1 且 deep_elipse_model.py / "
"section43_v3_1.py 已同步到服务器。"
)
return model
def build_elipse_model(
num_classes: int,
use_single_model_weiht: bool = False,
use_scheme_v2: bool = False,
use_scheme_v3: bool = False,
use_scheme_v3_1: bool = False,
v2_pre_smooth: bool = True,
v2_pool_mode: str = "sp",
v3_pool_mode: str = "avg",
v3_1_stage_pool: str = "global",
cwt_scale_affinity: bool = False,
cwt_scale_index_mode: str = "uniform",
angular_norm: str = "sum_relu",
e3_cross_query: str = "cwt_cnn",
e3_cross_query_path2: str = "maxpool",
e3_cwt_branch: bool = True,
e2_cwt_branch: bool = False,
cwt_phi_valid_emb: int = 0,
e2_cross_path2: str = "orig_cnn",
phi_tx_layers: int = 0,
phi_tx_heads: int = 4,
phi_tx_ffn: int = 256,
gcc_enabled: bool = False,
deep_backend: str = "convnext",
resnet_checkpoint: Optional[str] = None,
convnext_checkpoint: Optional[str] = None,
seq_backend: Optional[str] = None,
):
"""构建 DeepEllipseModel(函数名保留以兼容原脚本调用)。"""
cfg = DeepEllipseConfig(
num_classes=num_classes,
deep_backend=deep_backend,
resnet_checkpoint=resnet_checkpoint or default_resnet_checkpoint(),
convnext_checkpoint=convnext_checkpoint or default_convnext_checkpoint(),
use_scheme_v2=use_scheme_v2,
use_scheme_v3=use_scheme_v3,
use_scheme_v3_1=use_scheme_v3_1,
seq_backend=seq_backend,
v2_pre_smooth=v2_pre_smooth,
v2_pool_mode=v2_pool_mode,
v3_pool_mode=v3_pool_mode,
v3_1_stage_pool=v3_1_stage_pool,
cwt_scale_affinity=cwt_scale_affinity,
cwt_scale_index_mode=cwt_scale_index_mode,
angular_norm=angular_norm,
e3_cross_query=e3_cross_query,
e3_cross_query_path2=e3_cross_query_path2,
e3_cwt_branch=e3_cwt_branch,
e2_cwt_branch=e2_cwt_branch,
cwt_phi_valid_emb=cwt_phi_valid_emb,
e2_cross_path2=e2_cross_path2,
phi_tx_layers=phi_tx_layers,
phi_tx_heads=phi_tx_heads,
phi_tx_ffn=phi_tx_ffn,
gcc_enabled=gcc_enabled,
)
model = DeepEllipseModel(cfg)
if use_single_model_weiht:
checkpoint_path = "./checkpoints/deepelipse0526/best_model_epoch0_best.pth"
if os.path.isfile(checkpoint_path):
print(f"load deep ellipse: {checkpoint_path}")
checkpoint = torch.load(checkpoint_path, map_location=TrainConfig.DEVICE, weights_only=False)
model.load_state_dict(checkpoint["model_state_dict"], strict=False)
print("DeepEllipseModel 创建成功")
if use_scheme_v3_1:
scheme = model.enc43.config.scheme_label()
elif use_scheme_v3:
scheme = model.enc43.config.scheme_label()
elif use_scheme_v2:
scheme = model.enc43.config.scheme_label()
else:
scheme = "v1(MaxPool)"
print(f" §4.3 方案: {scheme}")
print(f" f_ellipse dim: {model.d_ellipse}")
print(f" §4.1.3 deep_backend: {model.enc41.deep_branch.backend}")
if model.enc41.deep_branch.backend == "convnext":
ckpt = getattr(model.enc41.config, "convnext_checkpoint", None)
print(f" §4.1.3 convnext_ckpt: {ckpt}")
elif model.enc41.deep_branch.backend == "resnet":
ckpt = getattr(model.enc41.config, "resnet_checkpoint", None)
print(f" §4.1.3 resnet_ckpt: {ckpt}")
print(f" E_pen dim (d_pen): {model.d_pen}")
total_params = sum(p.numel() for p in model.parameters())
trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
print(f" 总参数量:{total_params:,}")
print(f" 可训练参数:{trainable_params:,} ({trainable_params / total_params * 100:.2f}%)")
return model
def load_checkpoint(filepath: str, model, optimizer=None, scheduler=None, device=None):
"""加载检查点;模型权重走 partial 匹配,避免 strict 因架构不一致报错。"""
if device is None:
device = TrainConfig.DEVICE
print(f"📦 加载检查点: {filepath}")
checkpoint = torch.load(filepath, map_location=device, weights_only=False)
load_partial_weights(model, filepath)
if optimizer and "optimizer_state_dict" in checkpoint:
try:
optimizer.load_state_dict(checkpoint["optimizer_state_dict"])
except Exception as exc:
print(f"⚠️ 优化器状态未加载: {exc}")
if scheduler and checkpoint.get("scheduler_state_dict"):
try:
scheduler.load_state_dict(checkpoint["scheduler_state_dict"])
except Exception as exc:
print(f"⚠️ scheduler 状态未加载: {exc}")
print("✅ 检查点加载完成")
return {
"start_epoch": checkpoint.get("epoch", 0) + 1,
"best_val_acc": checkpoint.get("val_acc", 0.0),
"label_encoder_classes": checkpoint.get("label_encoder_classes"),
}
def unpack_batch(input_all):
return unpack_dataloader_batch(input_all)
def freeze_backbone(model: nn.Module) -> None:
model.eval()
for p in model.parameters():
p.requires_grad = False
def unfreeze_backbone(model: nn.Module, *, unfreeze_deep: bool = True) -> int:
"""Enable grads on backbone. Returns #trainable params."""
model.train()
n = 0
for name, p in model.named_parameters():
if (not unfreeze_deep) and (
"enc41.deep_branch" in name or name.startswith("enc41.deep_branch")
):
p.requires_grad = False
continue
p.requires_grad = True
n += p.numel()
return n
def extract_f_ellipse(
model,
fea,
img,
pen_ids,
cwt_scalo,
device,
*,
train_mode: bool = False,
):
if model is None:
# head-only mode: fea already carries precomputed f_ellipse
f = fea.to(device)
return f if train_mode else f.detach()
if hasattr(fea, "to"):
fea = fea.to(device)
img = img.to(device)
pen_ids = pen_ids.to(device)
if cwt_scalo is not None:
cwt_scalo = cwt_scalo.to(device)
_, details = model(
fea,
img,
pen_ids,
training=train_mode,
return_details=True,
cwt_scalo=cwt_scalo,
)
f = details["f_ellipse"]
return f if train_mode else f.detach()
def extract_logits(
model,
fea,
img,
pen_ids,
cwt_scalo,
device,
*,
train_mode: bool = False,
):
"""CE-head writer logits from a DeepEllipse forward (same as training export)."""
if hasattr(fea, "to"):
fea = fea.to(device)
img = img.to(device)
pen_ids = pen_ids.to(device)
if cwt_scalo is not None:
cwt_scalo = cwt_scalo.to(device)
logits = model(
fea,
img,
pen_ids,
training=train_mode,
return_details=False,
cwt_scalo=cwt_scalo,
)
return logits if train_mode else logits.detach()
def collect_test_y(
backbone,
loader,
device,
) -> Tuple[List[str], torch.Tensor]:
"""Run the frozen CE head over the test loader → (image_ids, Y[N,C])."""
backbone.eval()
ids: List[str] = []
ys: List[torch.Tensor] = []
for input_all, labels, pen_ids, writer_ids, image_ids in tqdm(
loader, desc="Extract Y (CE head)"
):
fea, ske, img, cwt = unpack_batch(input_all)
logits = extract_logits(
backbone, fea, img, pen_ids, cwt, device, train_mode=False
)
ids.extend([str(x) for x in image_ids])
ys.append(logits.cpu())
return ids, torch.cat(ys, dim=0)
def collect_test_h_and_y(
backbone,
head: PairwiseBCEHead,
loader,
device,
) -> Tuple[List[str], torch.Tensor, torch.Tensor]:
"""One backbone pass: affinity embedding h and CE logits Y."""
head.eval()
backbone.eval()
ids: List[str] = []
hs: List[torch.Tensor] = []
ys: List[torch.Tensor] = []
for input_all, labels, pen_ids, writer_ids, image_ids in tqdm(
loader, desc="Extract h+Y (test)"
):
fea, ske, img, cwt = unpack_batch(input_all)
if hasattr(fea, "to"):
fea = fea.to(device)
img = img.to(device)
pen_ids = pen_ids.to(device)
if cwt is not None:
cwt = cwt.to(device)
logits, details = backbone(
fea, img, pen_ids, training=False, return_details=True, cwt_scalo=cwt
)
f = details["f_ellipse"].detach()
h = head(f)
ids.extend([str(x) for x in image_ids])
hs.append(h.cpu())
ys.append(logits.detach().cpu())
return ids, torch.cat(hs, dim=0), torch.cat(ys, dim=0)
def collect_test_h(
backbone,
head: PairwiseBCEHead,
loader,
device,
) -> Tuple[List[str], torch.Tensor]:
head.eval()
if backbone is not None:
backbone.eval()
ids: List[str] = []
hs: List[torch.Tensor] = []
for input_all, labels, pen_ids, writer_ids, image_ids in tqdm(
loader, desc="Extract h (test)"
):
fea, ske, img, cwt = unpack_batch(input_all)
f = extract_f_ellipse(
backbone, fea, img, pen_ids, cwt, device, train_mode=False
)
h = head(f)
ids.extend([str(x) for x in image_ids])
hs.append(h.cpu())
return ids, torch.cat(hs, dim=0)
def _checkpoint_has_paper679_phi(ckpt_path) -> bool:
"""True if the backbone checkpoint carries the recovered standalone phi weights."""
try:
raw = torch.load(str(ckpt_path), map_location="cpu", weights_only=False)
except Exception:
return False
state = raw.get("model_state_dict", raw) if isinstance(raw, dict) else raw
return any(
k.endswith("segment_pool.phi_transformers.0.norm.weight") for k in state
)
def build_backbone(
args, num_classes: int, device, *, freeze: bool = True
) -> DeepEllipseModel:
ns = argparse.Namespace(
scheme_v2=False,
scheme_v3=False,
scheme_v3_1=True,
v3_1_stage_pool="cwt_angular",
cwt_scale_affinity=True,
cwt_scale_index_mode="uniform",
angular_norm="sum_relu",
e3_cross_query="cwt_cnn",
e3_cross_query_path2="maxpool",
e3_no_cwt_branch=False,
e2_cwt_branch=True,
e2_valid_phi=True,
e2_phi_tx_layers=2,
e2_cross_path2="orig_cnn",
gcc_enabled=False,
e2b=True,
)
parser = argparse.ArgumentParser()
align_args_with_checkpoint(ns, parser, args.backbone_ckpt, purpose="test")
# Paper-679 path: keep GCC module in the codebase/ckpt, but run the
# standalone phi_transformers branch with recovered weights so A/Y from
# the model match Private 0.67963 (do not load segment_pool.gcc.*).
paper_679 = bool(getattr(args, "paper_679", False))
phi_sidecar = getattr(args, "phi_sidecar", None)
if paper_679:
ns.gcc_enabled = False
if getattr(ns, "e2_phi_tx_layers", 0) <= 0:
ns.e2_phi_tx_layers = 2
print("📄 paper-679: gcc_enabled=False (standalone phi_transformers)")
model = build_elipse_model(
num_classes,
use_scheme_v3_1=True,
v3_1_stage_pool=ns.v3_1_stage_pool,
cwt_scale_affinity=ns.cwt_scale_affinity,
cwt_scale_index_mode=getattr(ns, "cwt_scale_index_mode", "uniform"),
angular_norm=getattr(ns, "angular_norm", "sum_relu"),
e3_cross_query=getattr(ns, "e3_cross_query", "cwt_cnn"),
e3_cross_query_path2=getattr(ns, "e3_cross_query_path2", "maxpool"),
e3_cwt_branch=not getattr(ns, "e3_no_cwt_branch", False),
e2_cwt_branch=getattr(ns, "e2_cwt_branch", True),
cwt_phi_valid_emb=8 if getattr(ns, "e2_valid_phi", True) else 0,
e2_cross_path2=getattr(ns, "e2_cross_path2", "orig_cnn"),
phi_tx_layers=getattr(ns, "e2_phi_tx_layers", 2),
gcc_enabled=getattr(ns, "gcc_enabled", False),
deep_backend=args.deep,
resnet_checkpoint=args.resnet_ckpt,
convnext_checkpoint=args.convnext_ckpt,
)
model.to(device)
load_checkpoint(args.backbone_ckpt, model, device=device)
if paper_679 or phi_sidecar:
pool = model.enc43.segment_pool
if getattr(pool, "phi_transformers", None) is None:
raise RuntimeError("paper-679 model has no phi_transformers (gcc still on?)")
# In-ckpt weights are the primary source: the released backbone carries the
# recovered standalone phi under enc43.segment_pool.phi_transformers.*, and
# load_partial_weights above already matched them.
in_ckpt = _checkpoint_has_paper679_phi(args.backbone_ckpt)
side = phi_sidecar
if side is None:
cand = [
Path("paper679_phi_transformers.pt"),
Path(args.backbone_ckpt).resolve().parent / "paper679_phi_transformers.pt",
]
side = next((c for c in cand if c.is_file()), None)
if in_ckpt:
print("📄 paper-679 phi loaded from backbone checkpoint "
"(enc43.segment_pool.phi_transformers.*)")
if side is not None and Path(side).is_file():
blob = torch.load(str(side), map_location="cpu", weights_only=False)
phi_sd = blob.get("phi_transformers", blob)
pool.phi_transformers.load_state_dict(phi_sd, strict=True)
print(f"📄 paper-679 phi sidecar override: {side} ({len(phi_sd)} tensors)")
elif not in_ckpt:
raise FileNotFoundError(
"paper-679 needs recovered phi weights: neither "
"enc43.segment_pool.phi_transformers.* in the checkpoint nor a "
"paper679_phi_transformers.pt sidecar was found"
)
if freeze:
freeze_backbone(model)
else:
n = unfreeze_backbone(
model, unfreeze_deep=bool(getattr(args, "unfreeze_deep", True))
)
print(f"Backbone unfrozen: {n:,} trainable params (unfreeze_deep={args.unfreeze_deep})")
return model