LeukoAI / leukoai /models.py
LeukoAI deploy
Deploy a23fa57 from synav/LeukoAI-dev
6e502b4
Raw History Blame Contribute Delete
2.96 kB
"""Model definitions, class labels and weight loading for LeukoAI v1.
The three models are DenseNet-121 classifiers recovered from the original
desktop application (toolv9, 2023). Weights in ``weights/`` are the original
checkpoints with training state removed; predictions are unchanged.
"""
from __future__ import annotations
import hashlib
import os
from dataclasses import dataclass
import torch
from torchvision.models import densenet121
# Class order exactly as in the original application.
DX1 = ["NO LEUKEMIA", "LEUKEMIA"]
DX2 = ["ACUTE LEUKEMIA", "CHRONIC LEUKEMIA", "NO LEUKEMIA", "REACTIVE CELLS"]
DX3 = ["ALL", "AML", "CLL", "CML", "NO LEUKEMIA"]
@dataclass(frozen=True)
class ModelSpec:
key: str
file: str
original_file: str
labels: tuple
sha256: str
SPECS = (
ModelSpec("m1", "m1_leukemia_binary.pt", "del_new_bin_flip_jit_scale.ckpt", tuple(DX1),
"e5b65eb8b31a187908faf77867949cfa8f70a08b29c956458a0f59d42f14c873"),
ModelSpec("m2", "m2_acute_chronic_normal_reactive.pt", "del_new_acnr_flip_scale_jit.ckpt", tuple(DX2),
"304f436254a084d3921daaca41bbeb0ed9c8a712c6b7389ded779c49075a3a05"),
ModelSpec("m3", "m3_subtype_5class.pt", "new_five.ckpt", tuple(DX3),
"78b520597a19d03647815b0ae332630f4fdc790b0c0475afacce5d804cdd21a9"),
)
DEFAULT_WEIGHTS_DIR = os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))), "weights")
def _sha256(path: str) -> str:
h = hashlib.sha256()
with open(path, "rb") as f:
for chunk in iter(lambda: f.read(1 << 20), b""):
h.update(chunk)
return h.hexdigest()
def _load_state_dict(spec: ModelSpec, weights_dir: str, verify: bool) -> dict:
slim = os.path.join(weights_dir, spec.file)
if os.path.exists(slim):
if verify and _sha256(slim) != spec.sha256:
raise ValueError(f"Checksum mismatch for {slim}; the weight file is not the released LeukoAI model.")
return torch.load(slim, map_location="cpu", weights_only=True)
original = os.path.join(weights_dir, spec.original_file)
if os.path.exists(original): # original toolv9 Lightning checkpoint
sd = torch.load(original, map_location="cpu", weights_only=True)["state_dict"]
return {k.replace("net.", "", 1): v for k, v in sd.items()}
raise FileNotFoundError(f"Neither {spec.file} nor {spec.original_file} found in {weights_dir}")
def load_models(weights_dir: str | None = None, device: str | None = None, verify: bool = True):
"""Return ``{"m1": model, "m2": model, "m3": model}`` in eval mode."""
weights_dir = weights_dir or DEFAULT_WEIGHTS_DIR
device = device or ("cuda" if torch.cuda.is_available() else "cpu")
models = {}
for spec in SPECS:
m = densenet121(num_classes=len(spec.labels))
m.load_state_dict(_load_state_dict(spec, weights_dir, verify), strict=True)
models[spec.key] = m.eval().to(device)
return models