predict.py

import os
import random
from argparse import ArgumentParser
from importlib import import_module

import SimpleITK as sitk
import numpy as np
import pandas as pd
import torch
import torch.nn as nn
from tqdm import tqdm

from data.dataset import LungSegmentDataset
from data import transforms as aug
from models.implicit_autoencoder import ImplicitAutoEncoder, ImplicitDecoder
from models.resnet18 import ResNet3d18Backbone
from utils.logger import logger


def _parse_cmd_args():
    arg_parser = ArgumentParser()
    arg_parser.add_argument("--gpu", default="0", help="GPU ID.")
    arg_parser.add_argument("--cfg", required=True,
        help="Python config module.")
    arg_parser.add_argument("--data_dir", required=True,
        help="Data directory")
    arg_parser.add_argument("--df_path", required=True,
        help="Data info csv path.")
    arg_parser.add_argument("--weight_path", required=True,
        help="Train model weight path.")
    arg_parser.add_argument("--output_dir", required=True,
        help="Prediction output directory.")
    args = arg_parser.parse_args()

    return args


def _set_rng_seed(seed):
    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)


def _init_dataloaders(args):
    data_dir = args.data_dir
    df = pd.read_csv(args.df_path)

    transforms = [
        aug.CropLung(),
        aug.GetLobe(),
        aug.Resample(cfg.resample_configs),
        aug.OnehotEncode("lobe", 6),
        aug.MinMaxNormalize(cfg.win_min, cfg.win_max),
        aug.SampleGrid(mode="regular"),
        aug.SampleTarget("lungsegment"),
        aug.ConcatInputs(cfg.input_keys),
        aug.ToTensor()
    ]

    ds_val = LungSegmentDataset(df, data_dir, transforms, "val")
    dl_val = LungSegmentDataset.get_dataloader(ds_val, cfg.infer_batch_size,
        False, cfg.num_workers, "infer")
    ds_test = LungSegmentDataset(df, data_dir, transforms, "test")
    dl_test = LungSegmentDataset.get_dataloader(ds_test, cfg.infer_batch_size,
        False, cfg.num_workers, "infer")

    return dl_val, dl_test


def _init_model(args):
    encoder = ResNet3d18Backbone(**cfg.enc_cfgs)
    decoder = ImplicitDecoder(**cfg.dec_cfgs)
    model = ImplicitAutoEncoder(encoder, decoder)

    devices = [torch.device(f"cuda:{i}") for i in args.gpu.split(",")]
    if len(devices) > 1:
        model = nn.DataParallel(model.cuda(), devices)
    else:
        model = model.cuda()

    return model


def _calculate_bboxes(image_shape, crop_size):
    steps = [np.arange(0, image_shape[i], crop_size[i]).tolist()
        + [image_shape[i]] for i in range(3)]
    begs = [steps[i][:-1] for i in range(3)]
    ends = [steps[i][1:] for i in range(3)]
    bboxes = []
    for i in range(len(begs[0])):
        for j in range(len(begs[1])):
            for k in range(len(begs[2])):
                bboxes.append(np.array([
                    [begs[0][i], ends[0][i]],
                    [begs[1][j], ends[1][j]],
                    [begs[2][k], ends[2][k]],
                ]))
    bboxes = np.stack(bboxes)

    return bboxes


def _sliding_window_predict(model, inputs, grids, window_size):
    resolution = grids.size()[-2:0:-1]
    bboxes = _calculate_bboxes(resolution, window_size)
    output = np.zeros(resolution, dtype=np.uint8)
    features = model.encoder(inputs)

    for i in range(bboxes.shape[0]):
        grid_patch = grids[
            :,
            bboxes[i, 2, 0]:bboxes[i, 2, 1],
            bboxes[i, 1, 0]:bboxes[i, 1, 1],
            bboxes[i, 0, 0]:bboxes[i, 0, 1],
            :].cuda()
        point_encodings = model.make_point_encoding(features, grid_patch)
        output_patch = model.decoder(point_encodings)
        output_patch = output_patch.cpu().numpy().squeeze(axis=0)
        output_patch = output_patch.argmax(axis=0).astype(np.uint8)
        output[
            bboxes[i, 0, 0]:bboxes[i, 0, 1],
            bboxes[i, 1, 0]:bboxes[i, 1, 1],
            bboxes[i, 2, 0]:bboxes[i, 2, 1]
        ] = output_patch.transpose((2, 1, 0))

    return output


@logger
@torch.no_grad()
def _predict(model, dataloader, output_dir, subset):
    model.eval()
    progress = tqdm(total=len(dataloader))
    os.makedirs(os.path.join(output_dir, subset), exist_ok=True)

    for i, sample in enumerate(dataloader):
        inputs, grids, pids, bboxes, shapes = sample
        pid = pids[0]
        bbox = bboxes[0]
        grids = grids[0][None]
        original_shape = shapes[0]
        inputs = inputs.cuda()

        y_pred_lung = _sliding_window_predict(model, inputs, grids,
            cfg.window_size)
        y_pred = np.zeros(original_shape, dtype=np.uint8)
        y_pred[
            bbox[0, 0]:bbox[0, 1] + 1,
            bbox[1, 0]:bbox[1, 1] + 1,
            bbox[2, 0]:bbox[2, 1] + 1
        ] = y_pred_lung

        y_pred_img = sitk.GetImageFromArray(y_pred.astype(np.uint8))
        sitk.WriteImage(y_pred_img, os.path.join(output_dir, subset,
            f"{pid}_pred.nii.gz"))

        progress.update()

    progress.close()


def _load_weights(weight_path):
    model_weights = torch.load(weight_path)
    new_model_weights = {}
    for k in model_weights.keys():
        if k.startswith("module."):
            new_k = k[7:]
            new_model_weights[new_k] = model_weights[k]

    return new_model_weights


def main():
    _set_rng_seed(42)

    args = _parse_cmd_args()
    os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu
    global cfg
    cfg = import_module(f"configs.{args.cfg}_config")

    dl_val, dl_test = _init_dataloaders(args)

    model = _init_model(args)
    model_weights = _load_weights(args.weight_path)
    model.load_state_dict(model_weights)

    _predict(model, dl_val, args.output_dir, "val")
    _predict(model, dl_test, args.output_dir, "test")


if __name__ == "__main__":
    main()