train.py

import sys
from collections import OrderedDict
from rich.progress import Progress, TextColumn, BarColumn, TaskProgressColumn, TimeRemainingColumn

import data
from options.train_options import TrainOptions
from util.iter_counter import IterationCounter
from util.visualizer import Visualizer
from trainers.pix2pix_trainer import Pix2PixTrainer
from util.logging_wandb import init_project,stop
import torch
from torch.distributions import Beta

# parse options
opt = TrainOptions().parse()
# print options to help debugging
print(' '.join(sys.argv))

# load the dataset
dataloader = data.create_dataloader(opt)

# create trainer for our model
trainer = Pix2PixTrainer(opt)
print(trainer)

# create tool for counting iterations
iter_counter = IterationCounter(opt, len(dataloader))

# create tool for visualization
visualizer = Visualizer(opt)
T = len(dataloader)
b = torch.tensor([1], dtype=torch.float32)

# Setup rich progress bar
progress = Progress(
    TextColumn("[bold blue]{task.description}"),
    BarColumn(),
    TaskProgressColumn(),
    TimeRemainingColumn(),
)

## INIT WANDB
init_project("results")

epoch = 0
with progress:
    epoch_task = progress.add_task(f"[deep_pink4]Total Epochs[{epoch + 1}|{max(iter_counter.training_epochs())}]",
                                   total=len(iter_counter.training_epochs()))
    for epoch in iter_counter.training_epochs():
        iter_counter.record_epoch_start(epoch)
        progress.update(epoch_task, advance=1)
        img_index = 0
        iter_task = progress.add_task(f"[cyan]Images Index [{img_index}]", total=len(dataloader))

        iteration = 0
        while iteration < len(dataloader):
            for i,data_i in enumerate(dataloader,start=iter_counter.epoch_iter):  # This will loop indefinitely if using InfiniteSamplerWrapper
                iter_counter.record_one_iteration()
                progress.update(iter_task, advance=1)

                t = iteration + 1
                a = torch.tensor(((t - (0.5 * T)) / (0.25 * T)), dtype=torch.float32).exp()
                m = Beta(a, b)
                opt.alpha = m.sample().cuda()

                # Training
                # train generator
                if iteration % opt.D_steps_per_G == 0:
                    trainer.run_generator_one_step(data_i, iteration, progress=progress, epoch=epoch,
                                                   images_iter=img_index)
                    img_index += 1
                    progress.update(iter_task,
                                    description=f"[cyan]Images Index [{img_index}|{len(dataloader)}]")

                # train discriminator
                trainer.run_discriminator_one_step(data_i, iteration)

                # Visualizations
                if iter_counter.needs_printing():
                    losses = trainer.get_latest_losses()
                    visualizer.print_current_errors(epoch, iter_counter.epoch_iter,
                                                    losses, iter_counter.time_per_iter)
                    visualizer.plot_current_errors(losses, iter_counter.total_steps_so_far)

                if iter_counter.needs_displaying():
                    if opt.task == 'SIS':
                        visuals = OrderedDict([('input_label', data_i['label'][0]),
                                               ('synthesized_image', trainer.get_latest_generated()[0]),
                                               ('real_image', data_i['image'][0])])
                    else:
                        visuals = OrderedDict([('content', data_i['label'][0]),
                                               ('synthesized_image', trainer.get_latest_generated()[0]),
                                               ('style', data_i['image'][0])])
                    visualizer.display_current_results(visuals, epoch, iter_counter.total_steps_so_far)

                if iter_counter.needs_saving():
                    print('saving the latest model (epoch %d, total_steps %d)' %
                          (epoch, iter_counter.total_steps_so_far))
                    trainer.save('latest')
                    iter_counter.record_current_iter()

                # Visualizations and model saving as in the original script

                iteration += 1
                if iteration >= opt.max_iterations_per_epoch:
                    break

        trainer.update_learning_rate(epoch)
        iter_counter.record_epoch_end()
        progress.update(epoch_task,
                        description=f"[deep_pink4]Total Epochs[{epoch + 1}|{max(iter_counter.training_epochs())}]")

        if epoch % opt.save_epoch_freq == 0 or epoch == iter_counter.total_epochs:
            print('saving the model at the end of epoch %d, iters %d' %
                  (epoch, iter_counter.total_steps_so_far))
            trainer.save('latest')
            trainer.save(epoch)
    stop()

print('Training was successfully finished.')