main.py

import argparse, os
import pdb
import torch
import random
import torch.backends.cudnn as cudnn
import torch.nn as nn
import torch.optim as optim
from torch.autograd import Variable
from torch.utils.data import DataLoader
from srdensenet import Net, L1_Charbonnier_loss
from dataset import DatasetFromHdf5

# Training settings
parser = argparse.ArgumentParser(description="PyTorch DenseNet")
parser.add_argument("--batchSize", type=int, default=32, help="training batch size")
parser.add_argument("--nEpochs", type=int, default=60, help="number of epochs to train for")
parser.add_argument("--lr", type=float, default=1e-4, help="Learning Rate. Default=1e-4")
parser.add_argument("--step", type=int, default=30, help="Sets the learning rate to the initial LR decayed by momentum every n epochs, Default: n=10")
parser.add_argument("--cuda", action="store_true", help="Use cuda?")
parser.add_argument("--resume", default="", type=str, help="Path to checkpoint (default: none)")
parser.add_argument("--start-epoch", default=1, type=int, help="Manual epoch number (useful on restarts)")
parser.add_argument("--threads", type=int, default=1, help="Number of threads for data loader to use, Default: 1")
parser.add_argument("--momentum", default=0.9, type=float, help="Momentum, Default: 0.9")
parser.add_argument("--weight-decay", "--wd", default=1e-4, type=float, help="weight decay, Default: 1e-4")
parser.add_argument("--pretrained", default="", type=str, help="path to pretrained model (default: none)")

def main():

    global opt, model 
    opt = parser.parse_args()
    print opt

    cuda = opt.cuda
    if cuda and not torch.cuda.is_available():
        raise Exception("No GPU found, please run without --cuda")

    opt.seed = random.randint(1, 10000)
    print("Random Seed: ", opt.seed)
    torch.manual_seed(opt.seed)
    if cuda:
        torch.cuda.manual_seed(opt.seed)

    cudnn.benchmark = True
        
    print("===> Loading datasets")
    train_set = DatasetFromHdf5("/path/to/your/dataset/like/imagenet_50K.h5")
    training_data_loader = DataLoader(dataset=train_set, num_workers=opt.threads, batch_size=opt.batchSize, shuffle=True)

    print("===> Building model")
    model = Net()
    criterion = L1_Charbonnier_loss()

    print("===> Setting GPU")
    if cuda:
        model = model.cuda()
        criterion = criterion.cuda()

    # optionally resume from a checkpoint
    if opt.resume:
        if os.path.isfile(opt.resume):
            print("=> loading checkpoint '{}'".format(opt.resume))
            checkpoint = torch.load(opt.resume)
            opt.start_epoch = checkpoint["epoch"] + 1
            model.load_state_dict(checkpoint["model"].state_dict())
        else:
            print("=> no checkpoint found at '{}'".format(opt.resume))
            
    # optionally copy weights from a checkpoint
    if opt.pretrained:
        if os.path.isfile(opt.pretrained):
            print("=> loading model '{}'".format(opt.pretrained))
            weights = torch.load(opt.pretrained)
            model.load_state_dict(weights['model'].state_dict())
        else:
            print("=> no model found at '{}'".format(opt.pretrained)) 
            
    print("===> Setting Optimizer")
    optimizer = optim.Adam(model.parameters(), lr=opt.lr)

    print("===> Training")
    for epoch in range(opt.start_epoch, opt.nEpochs + 1): 
        train(training_data_loader, optimizer, model, criterion, epoch)
        save_checkpoint(model, epoch)

def total_gradient(parameters):
    """Computes a gradient clipping coefficient based on gradient norm."""
    parameters = list(filter(lambda p: p.grad is not None, parameters))
    totalnorm = 0
    for p in parameters: 
        modulenorm = p.grad.data.norm()
        totalnorm += modulenorm ** 2
    totalnorm = totalnorm ** (1./2)
    return totalnorm
    
def adjust_learning_rate(optimizer, epoch):
    """Sets the learning rate to the initial LR decayed by 10 every 10 epochs"""
    lr = opt.lr * (0.1 ** (epoch // opt.step))
    return lr

def train(training_data_loader, optimizer, model, criterion, epoch):

    lr = adjust_learning_rate(optimizer, epoch-1)
    
    for param_group in optimizer.param_groups:
        param_group["lr"] = lr  
    print "epoch =", epoch,"lr =",optimizer.param_groups[0]["lr"]
    model.train()    

    for iteration, batch in enumerate(training_data_loader, 1):

        input, label = Variable(batch[0]), Variable(batch[1], requires_grad=False)

        if opt.cuda:
            input = input.cuda()
            label = label.cuda()
        
        out = model(input)
        
        loss = criterion(out, label)
                
        optimizer.zero_grad()

        loss.backward()

        optimizer.step()
        
        if iteration%100 == 0:
            print("===> Epoch[{}]({}/{}): Loss: {:.10f}".format(epoch, iteration, len(training_data_loader), loss.data[0]))
            print "total gradient", total_gradient(model.parameters())
    
def save_checkpoint(model, epoch):
    model_folder = "model_adam/"
    model_out_path = model_folder + "model_epoch_{}.pth".format(epoch)
    state = {"epoch": epoch ,"model": model}
    if not os.path.exists(model_folder):
        os.makedirs(model_folder)

    torch.save(state, model_out_path)
        
    print("Checkpoint saved to {}".format(model_out_path))

if __name__ == "__main__":
    main()