import torch import inspect import json import yaml import math import os import sys from general_utils import log import numpy as np from functools import partial from os.path import expanduser, join, isfile, basename from torch.cuda.amp import autocast, GradScaler from torch.optim.lr_scheduler import LambdaLR from contextlib import nullcontext from torch.utils.data import DataLoader from general_utils import TrainingLogger, get_attribute, filter_args, log, training_config_from_cli_args def cosine_warmup_lr(i, warmup=10, max_iter=90): """ Cosine LR with Warmup """ if i < warmup: return (i+1)/(warmup+1) else: return 0.5 + 0.5*math.cos(math.pi*(((i-warmup)/(max_iter- warmup)))) def validate(model, dataset, config): data_loader = torch.utils.data.DataLoader(dataset, batch_size=4, shuffle=False) metric_class, use_metric = config.val_metric_class, config.use_val_metric loss_fn = get_attribute(config.loss) model.eval() model.cuda() if metric_class is not None: metric = get_attribute(metric_class)() with torch.no_grad(): i, losses = 0, [] for data_x, data_y in data_loader: data_x = [x.cuda() if isinstance(x, torch.Tensor) else x for x in data_x] data_y = [x.cuda() if isinstance(x, torch.Tensor) else x for x in data_y] prompts = model.sample_prompts(data_x[1], prompt_list=('a photo of a {}',)) pred, visual_q, _, _ = model(data_x[0], prompts, return_features=True) if metric_class is not None: metric.add([pred], data_y) # pred = model(data_x[0], prompts) # loss = loss_fn(pred[0], data_y[0]) loss = loss_fn(pred, data_y[0]) losses += [float(loss)] i += 1 if config.val_max_iterations is not None and i > config.val_max_iterations: break if use_metric is None: return np.mean(losses), {}, False else: metric_scores = {m: s for m, s in zip(metric.names(), metric.value())} if metric is not None else {} return np.mean(losses), metric_scores, True def main(): config = training_config_from_cli_args() val_interval, best_val_loss, best_val_score = config.val_interval, float('inf'), float('-inf') model_cls = get_attribute(config.model) _, model_args, _ = filter_args(config, inspect.signature(model_cls).parameters) model = model_cls(**model_args).cuda() dataset_cls = get_attribute(config.dataset) _, dataset_args, _ = filter_args(config, inspect.signature(dataset_cls).parameters) dataset = dataset_cls(**dataset_args) log.info(f'Train dataset {dataset.__class__.__name__} (length: {len(dataset)})') if val_interval is not None: dataset_val_args = {k[4:]: v for k,v in config.items() if k.startswith('val_') and k != 'val_interval'} _, dataset_val_args, _ = filter_args(dataset_val_args, inspect.signature(dataset_cls).parameters) print('val args', {**dataset_args, **{'split': 'val', 'aug': 0}, **dataset_val_args}) dataset_val = dataset_cls(**{**dataset_args, **{'split': 'val', 'aug': 0}, **dataset_val_args}) # optimizer opt_cls = get_attribute(config.optimizer) if config.optimize == 'torch.optim.SGD': opt_args = {'momentum': config.momentum if 'momentum' in config else 0} else: opt_args = {} opt = opt_cls(model.parameters(), lr=config.lr, **opt_args) if config.lr_scheduler == 'cosine': assert config.T_max is not None and config.eta_min is not None lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(opt, config.T_max, config.eta_min) elif config.lr_scheduler == 'warmup_cosine': lr_scheduler = LambdaLR(opt, partial(cosine_warmup_lr, max_iter=(config.max_iterations), warmup=config.warmup)) else: lr_scheduler = None batch_size, max_iterations = config.batch_size, config.max_iterations loss_fn = get_attribute(config.loss) if config.amp: log.info('Using AMP') autocast_fn = autocast scaler = GradScaler() else: autocast_fn, scaler = nullcontext, None save_only_trainable = True data_loader = DataLoader(dataset, batch_size=batch_size, num_workers=4) # disable config when hyperparam. opt. to avoid writing logs. tracker_config = config if not config.hyperparameter_optimization else None with TrainingLogger(log_dir=config.name, model=model, config=tracker_config) as logger: i = 0 while True: for data_x, data_y in data_loader: # between caption and output feature. # 1. Sample random captions # 2. Check alignment with CLIP # randomly mix text and visual support conditionals if config.mix: assert config.mask.startswith('text_and') with autocast_fn(): # data_x[1] = text label prompts = model.sample_prompts(data_x[1]) # model.clip_model() text_cond = model.compute_conditional(prompts) if model.__class__.__name__ == 'CLIPDensePredTMasked': # when mask=='separate' visual_s_cond, _, _ = model.visual_forward_masked(data_x[2].cuda(), data_x[3].cuda()) else: # data_x[2] = visual prompt visual_s_cond, _, _ = model.visual_forward(data_x[2].cuda()) max_txt = config.mix_text_max if config.mix_text_max is not None else 1 batch_size = text_cond.shape[0] # sample weights for each element in batch text_weights = torch.distributions.Uniform(config.mix_text_min, max_txt).sample((batch_size,))[:, None] text_weights = text_weights.cuda() if dataset.__class__.__name__ == 'PhraseCut': # give full weight to text where support_image is invalid visual_is_valid = data_x[4] if model.__class__.__name__ == 'CLIPDensePredTMasked' else data_x[3] text_weights = torch.max(text_weights[:,0], 1 - visual_is_valid.float().cuda()).unsqueeze(1) cond = text_cond * text_weights + visual_s_cond * (1 - text_weights) else: # no mix if model.__class__.__name__ == 'CLIPDensePredTMasked': # compute conditional vector using CLIP masking with autocast_fn(): assert config.mask == 'separate' cond, _, _ = model.visual_forward_masked(data_x[1].cuda(), data_x[2].cuda()) else: cond = data_x[1] if isinstance(cond, torch.Tensor): cond = cond.cuda() with autocast_fn(): visual_q = None pred, visual_q, _, _ = model(data_x[0].cuda(), cond, return_features=True) loss = loss_fn(pred, data_y[0].cuda()) if torch.isnan(loss) or torch.isinf(loss): # skip if loss is nan log.warning('Training stopped due to inf/nan loss.') sys.exit(-1) extra_loss = 0 loss += extra_loss opt.zero_grad() if scaler is None: loss.backward() opt.step() else: scaler.scale(loss).backward() scaler.step(opt) scaler.update() if lr_scheduler is not None: lr_scheduler.step() if i % 2000 == 0: current_lr = [g['lr'] for g in opt.param_groups][0] log.info(f'current lr: {current_lr:.5f} ({len(opt.param_groups)} parameter groups)') logger.iter(i=i, loss=loss) i += 1 if i >= max_iterations: if not isfile(join(logger.base_path, 'weights.pth')): # only write if no weights were already written logger.save_weights(only_trainable=save_only_trainable) sys.exit(0) if config.checkpoint_iterations is not None and i in config.checkpoint_iterations: logger.save_weights(only_trainable=save_only_trainable, weight_file=f'weights_{i}.pth') if val_interval is not None and i % val_interval == val_interval - 1: val_loss, val_scores, maximize = validate(model, dataset_val, config) if len(val_scores) > 0: score_str = f', scores: ' + ', '.join(f'{k}: {v}' for k, v in val_scores.items()) if maximize and val_scores[config.use_val_metric] > best_val_score: logger.save_weights(only_trainable=save_only_trainable) best_val_score = val_scores[config.use_val_metric] elif not maximize and val_scores[config.use_val_metric] < best_val_score: logger.save_weights(only_trainable=save_only_trainable) best_val_score = val_scores[config.use_val_metric] else: score_str = '' # if no score is used, fall back to loss if val_loss < best_val_loss: logger.save_weights(only_trainable=save_only_trainable) best_val_loss = val_loss log.info(f'Validation loss: {val_loss}' + score_str) logger.iter(i=i, val_loss=val_loss, extra_loss=float(extra_loss), **val_scores) model.train() print('epoch complete') if __name__ == '__main__': main()