third_party/r2d2/train.py

# Copyright 2019-present NAVER Corp.
# CC BY-NC-SA 3.0
# Available only for non-commercial use

import os, pdb
import torch
import torch.optim as optim

from tools import common, trainer
from tools.dataloader import *
from nets.patchnet import *
from nets.losses import *

default_net = "Quad_L2Net_ConfCFS()"

toy_db_debug = """SyntheticPairDataset(
    ImgFolder('imgs'), 
            'RandomScale(256,1024,can_upscale=True)', 
            'RandomTilting(0.5), PixelNoise(25)')"""

db_web_images = """SyntheticPairDataset(
    web_images, 
        'RandomScale(256,1024,can_upscale=True)',
        'RandomTilting(0.5), PixelNoise(25)')"""

db_aachen_images = """SyntheticPairDataset(
    aachen_db_images, 
        'RandomScale(256,1024,can_upscale=True)', 
        'RandomTilting(0.5), PixelNoise(25)')"""

db_aachen_style_transfer = """TransformedPairs(
    aachen_style_transfer_pairs,
            'RandomScale(256,1024,can_upscale=True), RandomTilting(0.5), PixelNoise(25)')"""

db_aachen_flow = "aachen_flow_pairs"

data_sources = dict(
    D=toy_db_debug,
    W=db_web_images,
    A=db_aachen_images,
    F=db_aachen_flow,
    S=db_aachen_style_transfer,
)

default_dataloader = """PairLoader(CatPairDataset(`data`),
    scale   = 'RandomScale(256,1024,can_upscale=True)',
    distort = 'ColorJitter(0.2,0.2,0.2,0.1)',
    crop    = 'RandomCrop(192)')"""

default_sampler = """NghSampler2(ngh=7, subq=-8, subd=1, pos_d=3, neg_d=5, border=16,
                            subd_neg=-8,maxpool_pos=True)"""

default_loss = """MultiLoss(
        1, ReliabilityLoss(`sampler`, base=0.5, nq=20),
        1, CosimLoss(N=`N`),
        1, PeakyLoss(N=`N`))"""


class MyTrainer(trainer.Trainer):
    """This class implements the network training.
    Below is the function I need to overload to explain how to do the backprop.
    """

    def forward_backward(self, inputs):
        output = self.net(imgs=[inputs.pop("img1"), inputs.pop("img2")])
        allvars = dict(inputs, **output)
        loss, details = self.loss_func(**allvars)
        if torch.is_grad_enabled():
            loss.backward()
        return loss, details


if __name__ == "__main__":
    import argparse

    parser = argparse.ArgumentParser("Train R2D2")

    parser.add_argument("--data-loader", type=str, default=default_dataloader)
    parser.add_argument(
        "--train-data",
        type=str,
        default=list("WASF"),
        nargs="+",
        choices=set(data_sources.keys()),
    )
    parser.add_argument(
        "--net", type=str, default=default_net, help="network architecture"
    )

    parser.add_argument(
        "--pretrained", type=str, default="", help="pretrained model path"
    )
    parser.add_argument(
        "--save-path", type=str, required=True, help="model save_path path"
    )

    parser.add_argument("--loss", type=str, default=default_loss, help="loss function")
    parser.add_argument(
        "--sampler", type=str, default=default_sampler, help="AP sampler"
    )
    parser.add_argument(
        "--N", type=int, default=16, help="patch size for repeatability"
    )

    parser.add_argument(
        "--epochs", type=int, default=25, help="number of training epochs"
    )
    parser.add_argument("--batch-size", "--bs", type=int, default=8, help="batch size")
    parser.add_argument("--learning-rate", "--lr", type=str, default=1e-4)
    parser.add_argument("--weight-decay", "--wd", type=float, default=5e-4)

    parser.add_argument(
        "--threads", type=int, default=8, help="number of worker threads"
    )
    parser.add_argument("--gpu", type=int, nargs="+", default=[0], help="-1 for CPU")

    args = parser.parse_args()

    iscuda = common.torch_set_gpu(args.gpu)
    common.mkdir_for(args.save_path)

    # Create data loader
    from datasets import *

    db = [data_sources[key] for key in args.train_data]
    db = eval(args.data_loader.replace("`data`", ",".join(db)).replace("\n", ""))
    print("Training image database =", db)
    loader = threaded_loader(db, iscuda, args.threads, args.batch_size, shuffle=True)

    # create network
    print("\n>> Creating net = " + args.net)
    net = eval(args.net)
    print(f" ( Model size: {common.model_size(net)/1000:.0f}K parameters )")

    # initialization
    if args.pretrained:
        checkpoint = torch.load(args.pretrained, lambda a, b: a)
        net.load_pretrained(checkpoint["state_dict"])

    # create losses
    loss = args.loss.replace("`sampler`", args.sampler).replace("`N`", str(args.N))
    print("\n>> Creating loss = " + loss)
    loss = eval(loss.replace("\n", ""))

    # create optimizer
    optimizer = optim.Adam(
        [p for p in net.parameters() if p.requires_grad],
        lr=args.learning_rate,
        weight_decay=args.weight_decay,
    )

    train = MyTrainer(net, loader, loss, optimizer)
    if iscuda:
        train = train.cuda()

    # Training loop #
    for epoch in range(args.epochs):
        print(f"\n>> Starting epoch {epoch}...")
        train()

    print(f"\n>> Saving model to {args.save_path}")
    torch.save({"net": args.net, "state_dict": net.state_dict()}, args.save_path)