Spaces:
Sleeping
Sleeping
File size: 6,069 Bytes
c70812a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 | r""" training (validation) code """
import torch.optim as optim
import torch.nn as nn
import torch
from model.DCAMA import DCAMA
from common.logger import Logger, AverageMeter
from common.evaluation import Evaluator
from common.config import parse_opts
from common import utils
from data.dataset import FSSDataset
import torch.nn.functional as F
average_loss = torch.tensor(0.).float().cuda()
global_idx = 0
def train(epoch, model, dataloader, optimizer, training):
r""" Train """
# Force randomness during training / freeze randomness during testing
utils.fix_randseed(None) if training else utils.fix_randseed(0)
model.module.train_mode() if training else model.module.eval()
average_meter = AverageMeter(dataloader.dataset)
global average_loss, global_idx
average_loss = average_loss.to("cuda:{}".format(torch.cuda.current_device()))
stats = [[], []]
for idx, batch in enumerate(dataloader):
# 1. forward pass
batch = utils.to_cuda(batch)
logit_mask, score_preds = model(batch['query_img'], batch['support_imgs'], batch['support_masks'], nshot=batch['support_imgs'].size(1))
pred_mask = logit_mask.argmax(dim=1)
# 2. Compute loss & update model parameters
loss = model.module.compute_objective(logit_mask, batch['query_mask'])
# loss_obj = loss.detach()
area_inter, area_union = Evaluator.classify_prediction(pred_mask, batch)
iou = area_inter[1] / area_union[1]
loss_obj = iou.detach()
score_loss = F.l1_loss(score_preds, loss_obj)
stats[0].append(score_preds.detach().cpu().numpy())
stats[1].append(loss_obj.detach().cpu().numpy()[0])
if global_idx == 0:
average_loss = loss_obj.detach()
global_idx += 1
else:
average_loss = loss_obj.detach() * 0.05 + 0.95 * average_loss
print(loss_obj.item(), " ", score_preds.item(), " ", score_loss.item())
loss += score_loss
if training:
optimizer.zero_grad()
loss.backward()
optimizer.step()
# 3. Evaluate prediction
area_inter, area_union = Evaluator.classify_prediction(pred_mask, batch)
average_meter.update(area_inter, area_union, batch['class_id'], loss.detach().clone())
average_meter.write_process(idx, len(dataloader), epoch, write_batch_idx=50)
# Write evaluation results
average_meter.write_result('Training' if training else 'Validation', epoch)
avg_loss = utils.mean(average_meter.loss_buf)
miou, fb_iou = average_meter.compute_iou()
import matplotlib.pyplot as plt
idx = 0
plt.scatter(stats[0], stats[1], c="red", s=2, alpha=0.1)
plt.savefig('stat.png')
plt.close()
return avg_loss, miou, fb_iou
if __name__ == '__main__':
# Arguments parsing
args = parse_opts()
# ddp backend initialization
torch.distributed.init_process_group(backend='nccl')
torch.cuda.set_device(args.local_rank)
# Model initialization
model = DCAMA(args.backbone, args.feature_extractor_path, False)
device = torch.device("cuda", args.local_rank)
model.to(device)
model = nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank], output_device=args.local_rank,
find_unused_parameters=True)
params = model.state_dict()
state_dict = torch.load(args.load)
state_dict2 = {}
for k in state_dict.keys():
if "scorer" in k:
continue
state_dict2[k] = state_dict[k]
state_dict = state_dict2
for k1, k2 in zip(list(state_dict.keys()), params.keys()):
state_dict[k2] = state_dict.pop(k1)
model.load_state_dict(state_dict, strict=False)
## TODO:
for i in range(len(model.module.model.DCAMA_blocks)):
torch.nn.init.constant_(model.module.model.DCAMA_blocks[i].linears[1].weight, 0.)
torch.nn.init.constant_(model.module.model.DCAMA_blocks[i].linears[1].bias, 1.)
# Helper classes (for training) initialization
optimizer = optim.SGD([{"params": model.module.model.parameters(), "lr": args.lr,
"momentum": 0.9, "weight_decay": args.lr/10, "nesterov": True}])
Evaluator.initialize()
if args.local_rank == 0:
Logger.initialize(args, training=True)
Logger.info('# available GPUs: %d' % torch.cuda.device_count())
# Dataset initialization
FSSDataset.initialize(img_size=384, datapath=args.datapath, use_original_imgsize=False)
dataloader_trn = FSSDataset.build_dataloader(args.benchmark, args.bsz, args.nworker, args.fold, 'trn', args.nshot)
if args.local_rank == 0:
dataloader_val = FSSDataset.build_dataloader(args.benchmark, args.bsz, args.nworker, args.fold, 'val', args.nshot)
# Train
best_val_miou = float('-inf')
best_val_loss = float('inf')
for epoch in range(args.nepoch):
dataloader_trn.sampler.set_epoch(epoch)
trn_loss, trn_miou, trn_fb_iou = train(epoch, model, dataloader_trn, optimizer, training=True)
# evaluation
if args.local_rank == 0:
# with torch.no_grad():
# val_loss, val_miou, val_fb_iou = train(epoch, model, dataloader_val, optimizer, training=False)
# Save the best model
# if val_miou > best_val_miou:
# best_val_miou = val_miou
# Logger.save_model_miou(model, epoch, val_miou)
Logger.save_model_miou(model, epoch , 1.)
# Logger.tbd_writer.add_scalars('data/loss', {'trn_loss': trn_loss, 'val_loss': val_loss}, epoch)
# Logger.tbd_writer.add_scalars('data/miou', {'trn_miou': trn_miou, 'val_miou': val_miou}, epoch)
# Logger.tbd_writer.add_scalars('data/fb_iou', {'trn_fb_iou': trn_fb_iou, 'val_fb_iou': val_fb_iou}, epoch)
# Logger.tbd_writer.flush()
if args.local_rank == 0:
Logger.tbd_writer.close()
Logger.info('==================== Finished Training ====================')
|