train.py

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

# Modified for RoBLo feature extraction

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

from tools import common, trainer
from tools.robotic_burst_dataloader import *
from nets.patchnet import *
from nets.losses import *
from torch.utils.tensorboard import SummaryWriter

default_net = "Fast_Quad_L2Net_ConfCFS(inchan=15)" # For a burst of 5 RGB images, 5 x 3 => 15

# For training using toy dataset
#toy_db_debug = """SyntheticPairDataset(ImgFolder('imgs'),
#        'RandomScale(256,1024,can_upscale=True)',
#        'RandomTilting(0.01), PixelNoise(0)')"""

# Aachen dataset images
db_aachen_images = """SyntheticPairDataset(
    aachen_db_images,
        'RandomScale(256,1024,can_upscale=True)',
        'RandomTilting(0.01), PixelNoise(0)')"""

# Random web images
db_web_images = """SyntheticPairDataset(
    web_images,
        'RandomScale(256,1024,can_upscale=True)',
        'RandomTilting(0.01), PixelNoise(0)')"""

# Calling all training datasets
data_sources = dict(
    #T = toy_db_debug,
    W = db_web_images,
    A = db_aachen_images,
    )

default_dataloader = """PairLoader(CatPairDataset(`data`),
    scale   = 'RandomScale(256,1024,can_upscale=True)',
    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.
    """
    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

    def log_run(writer, loss, epoch):
        for key in loss:
            writer.add_scalar(key, loss[key], epoch + 1)

if __name__ == '__main__':
    import argparse
    parser = argparse.ArgumentParser("Train RoBLo")

    parser.add_argument("--data-loader", type=str, default=default_dataloader)
    parser.add_argument("--train-data", type=str, default=list('WA'), 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=2, 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)

    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)

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

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

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

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

    tb_dir = os.path.join(args.save_path, "tensorboard")
    common.mkdir_for(tb_dir)
    writer = SummaryWriter(tb_dir)

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

    for epoch in range(args.epochs):
        print(f"\n>> Starting epoch {epoch}...")
        loss = train()
        writer.add_scalar('loss', loss, epoch + 1)
        model_name = f"{args.save_path[:-4]}_{epoch}.pt"
        print(f"\n>> Saving model to {model_name}")
        torch.save({'net': args.net, 'state_dict': net.state_dict()}, model_name)

    print(f"\n>> Saving model to {args.save_path}")
    torch.save({'net': args.net, 'state_dict': net.state_dict()}, os.path.join(args.save_path, 'RoBLo_N16_B5.pt'))