412 lines
17 KiB
Python
412 lines
17 KiB
Python
import argparse
|
||
import time
|
||
|
||
import torch.distributed as dist
|
||
import torch.optim as optim
|
||
import torch.optim.lr_scheduler as lr_scheduler
|
||
from torch.utils.data import DataLoader
|
||
|
||
import test # Import test.py to get mAP after each epoch
|
||
from models import *
|
||
from utils.datasets import *
|
||
from utils.utils import *
|
||
|
||
# Hyperparameters: train.py --evolve --epochs 2 --img-size 320, Metrics: 0.204 0.302 0.175 0.234 (square smart)
|
||
hyp = {'giou': .035, # giou loss gain
|
||
'xy': 0.20, # xy loss gain
|
||
'wh': 0.10, # wh loss gain
|
||
'cls': 0.035, # cls loss gain
|
||
'conf': 1.61, # conf loss gain
|
||
'conf_bpw': 3.53, # conf BCELoss positive_weight
|
||
'iou_t': 0.29, # iou target-anchor training threshold
|
||
'lr0': 0.001, # initial learning rate
|
||
'lrf': -4., # final learning rate = lr0 * (10 ** lrf)
|
||
'momentum': 0.90, # SGD momentum
|
||
'weight_decay': 0.0005} # optimizer weight decay
|
||
|
||
|
||
# Hyperparameters: Original, Metrics: 0.172 0.304 0.156 0.205 (square)
|
||
# hyp = {'xy': 0.5, # xy loss gain
|
||
# 'wh': 0.0625, # wh loss gain
|
||
# 'cls': 0.0625, # cls loss gain
|
||
# 'conf': 4, # conf loss gain
|
||
# 'iou_t': 0.1, # iou target-anchor training threshold
|
||
# 'lr0': 0.001, # initial learning rate
|
||
# 'lrf': -5., # final learning rate = lr0 * (10 ** lrf)
|
||
# 'momentum': 0.9, # SGD momentum
|
||
# 'weight_decay': 0.0005} # optimizer weight decay
|
||
|
||
# Hyperparameters: train.py --evolve --epochs 2 --img-size 320, Metrics: 0.225 0.251 0.145 0.218 (rect)
|
||
# hyp = {'xy': 0.4499, # xy loss gain
|
||
# 'wh': 0.05121, # wh loss gain
|
||
# 'cls': 0.04207, # cls loss gain
|
||
# 'conf': 2.853, # conf loss gain
|
||
# 'iou_t': 0.2487, # iou target-anchor training threshold
|
||
# 'lr0': 0.0005301, # initial learning rate
|
||
# 'lrf': -5., # final learning rate = lr0 * (10 ** lrf)
|
||
# 'momentum': 0.8823, # SGD momentum
|
||
# 'weight_decay': 0.0004149} # optimizer weight decay
|
||
|
||
# Hyperparameters: train.py --evolve --epochs 2 --img-size 320, Metrics: 0.178 0.313 0.167 0.212 (square)
|
||
# hyp = {'xy': 0.4664, # xy loss gain
|
||
# 'wh': 0.08437, # wh loss gain
|
||
# 'cls': 0.05145, # cls loss gain
|
||
# 'conf': 4.244, # conf loss gain
|
||
# 'iou_t': 0.09121, # iou target-anchor training threshold
|
||
# 'lr0': 0.0004938, # initial learning rate
|
||
# 'lrf': -5., # final learning rate = lr0 * (10 ** lrf)
|
||
# 'momentum': 0.9025, # SGD momentum
|
||
# 'weight_decay': 0.0005417} # optimizer weight decay
|
||
|
||
def train(
|
||
cfg,
|
||
data_cfg,
|
||
img_size=416,
|
||
resume=False,
|
||
epochs=100, # 500200 batches at bs 4, 117263 images = 68 epochs
|
||
batch_size=16,
|
||
accumulate=4, # effective bs = 64 = batch_size * accumulate
|
||
freeze_backbone=False,
|
||
transfer=False # Transfer learning (train only YOLO layers)
|
||
):
|
||
init_seeds()
|
||
weights = 'weights' + os.sep
|
||
latest = weights + 'latest.pt'
|
||
best = weights + 'best.pt'
|
||
device = torch_utils.select_device()
|
||
torch.backends.cudnn.benchmark = True # possibly unsuitable for multiscale
|
||
img_size_test = img_size # image size for testing
|
||
multi_scale = not opt.single_scale
|
||
|
||
if multi_scale:
|
||
img_size_min = round(img_size / 32 / 1.5)
|
||
img_size_max = round(img_size / 32 * 1.5)
|
||
img_size = img_size_max * 32 # initiate with maximum multi_scale size
|
||
|
||
# Configure run
|
||
data_dict = parse_data_cfg(data_cfg)
|
||
train_path = data_dict['train']
|
||
nc = int(data_dict['classes']) # number of classes
|
||
|
||
# Initialize model
|
||
model = Darknet(cfg).to(device)
|
||
|
||
# Optimizer
|
||
optimizer = optim.SGD(model.parameters(), lr=hyp['lr0'], momentum=hyp['momentum'], weight_decay=hyp['weight_decay'])
|
||
|
||
cutoff = -1 # backbone reaches to cutoff layer
|
||
start_epoch = 0
|
||
best_loss = float('inf')
|
||
nf = int(model.module_defs[model.yolo_layers[0] - 1]['filters']) # yolo layer size (i.e. 255)
|
||
if resume: # Load previously saved model
|
||
if transfer: # Transfer learning
|
||
chkpt = torch.load(weights + 'yolov3-spp.pt', map_location=device)
|
||
model.load_state_dict({k: v for k, v in chkpt['model'].items() if v.numel() > 1 and v.shape[0] != 255},
|
||
strict=False)
|
||
for p in model.parameters():
|
||
p.requires_grad = True if p.shape[0] == nf else False
|
||
|
||
else: # resume from latest.pt
|
||
chkpt = torch.load(latest, map_location=device) # load checkpoint
|
||
model.load_state_dict(chkpt['model'])
|
||
|
||
start_epoch = chkpt['epoch'] + 1
|
||
if chkpt['optimizer'] is not None:
|
||
optimizer.load_state_dict(chkpt['optimizer'])
|
||
best_loss = chkpt['best_loss']
|
||
del chkpt
|
||
|
||
else: # Initialize model with backbone (optional)
|
||
if '-tiny.cfg' in cfg:
|
||
cutoff = load_darknet_weights(model, weights + 'yolov3-tiny.conv.15')
|
||
else:
|
||
cutoff = load_darknet_weights(model, weights + 'darknet53.conv.74')
|
||
|
||
# Remove old results
|
||
for f in glob.glob('*_batch*.jpg') + glob.glob('results.txt'):
|
||
os.remove(f)
|
||
|
||
# Scheduler https://github.com/ultralytics/yolov3/issues/238
|
||
# lf = lambda x: 1 - x / epochs # linear ramp to zero
|
||
# lf = lambda x: 10 ** (hyp['lrf'] * x / epochs) # exp ramp
|
||
# lf = lambda x: 1 - 10 ** (hyp['lrf'] * (1 - x / epochs)) # inverse exp ramp
|
||
# scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
|
||
scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=[round(opt.epochs * x) for x in (0.8, 0.9)], gamma=0.1)
|
||
scheduler.last_epoch = start_epoch - 1
|
||
|
||
# # Plot lr schedule
|
||
# y = []
|
||
# for _ in range(epochs):
|
||
# scheduler.step()
|
||
# y.append(optimizer.param_groups[0]['lr'])
|
||
# plt.plot(y, label='LambdaLR')
|
||
# plt.xlabel('epoch')
|
||
# plt.xlabel('LR')
|
||
# plt.tight_layout()
|
||
# plt.savefig('LR.png', dpi=300)
|
||
|
||
# Dataset
|
||
rectangular_training = False
|
||
dataset = LoadImagesAndLabels(train_path,
|
||
img_size,
|
||
batch_size,
|
||
augment=True,
|
||
rect=rectangular_training)
|
||
|
||
# Initialize distributed training
|
||
if torch.cuda.device_count() > 1:
|
||
dist.init_process_group(backend=opt.backend, init_method=opt.dist_url, world_size=opt.world_size, rank=opt.rank)
|
||
model = torch.nn.parallel.DistributedDataParallel(model)
|
||
# sampler = torch.utils.data.distributed.DistributedSampler(dataset)
|
||
|
||
# Dataloader
|
||
dataloader = DataLoader(dataset,
|
||
batch_size=batch_size,
|
||
num_workers=opt.num_workers,
|
||
shuffle=not rectangular_training, # Shuffle=True unless rectangular training is used
|
||
pin_memory=True,
|
||
collate_fn=dataset.collate_fn)
|
||
|
||
# Mixed precision training https://github.com/NVIDIA/apex
|
||
# install help: https://github.com/NVIDIA/apex/issues/259
|
||
mixed_precision = False
|
||
if mixed_precision:
|
||
from apex import amp
|
||
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
|
||
|
||
# Start training
|
||
model.hyp = hyp # attach hyperparameters to model
|
||
model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) # attach class weights
|
||
model_info(model)
|
||
nb = len(dataloader)
|
||
maps = np.zeros(nc) # mAP per class
|
||
results = (0, 0, 0, 0, 0) # P, R, mAP, F1, test_loss
|
||
n_burnin = min(round(nb / 5 + 1), 1000) # burn-in batches
|
||
t, t0 = time.time(), time.time()
|
||
for epoch in range(start_epoch, epochs):
|
||
model.train()
|
||
print(('\n%8s%12s' + '%10s' * 7) % ('Epoch', 'Batch', 'xy', 'wh', 'conf', 'cls', 'total', 'targets', 'time'))
|
||
|
||
# Update scheduler
|
||
scheduler.step()
|
||
|
||
# Freeze backbone at epoch 0, unfreeze at epoch 1 (optional)
|
||
if freeze_backbone and epoch < 2:
|
||
for name, p in model.named_parameters():
|
||
if int(name.split('.')[1]) < cutoff: # if layer < 75
|
||
p.requires_grad = False if epoch == 0 else True
|
||
|
||
# # Update image weights (optional)
|
||
# w = model.class_weights.cpu().numpy() * (1 - maps) # class weights
|
||
# image_weights = labels_to_image_weights(dataset.labels, nc=nc, class_weights=w)
|
||
# dataset.indices = random.choices(range(dataset.n), weights=image_weights, k=dataset.n) # random weighted index
|
||
|
||
mloss = torch.zeros(5).to(device) # mean losses
|
||
for i, (imgs, targets, _, _) in enumerate(dataloader):
|
||
imgs = imgs.to(device)
|
||
targets = targets.to(device)
|
||
|
||
# Multi-Scale training
|
||
if multi_scale:
|
||
if (i + 1 + nb * epoch) / accumulate % 10 == 0: # adjust (67% - 150%) every 10 batches
|
||
img_size = random.choice(range(img_size_min, img_size_max + 1)) * 32
|
||
print('img_size = %g' % img_size)
|
||
scale_factor = img_size / max(imgs.shape[-2:])
|
||
imgs = F.interpolate(imgs, scale_factor=scale_factor, mode='bilinear', align_corners=False)
|
||
|
||
# Plot images with bounding boxes
|
||
if epoch == 0 and i == 0:
|
||
plot_images(imgs=imgs, targets=targets, fname='train_batch%g.jpg' % i)
|
||
|
||
# SGD burn-in
|
||
if epoch == 0 and i <= n_burnin:
|
||
lr = hyp['lr0'] * (i / n_burnin) ** 4
|
||
for x in optimizer.param_groups:
|
||
x['lr'] = lr
|
||
|
||
# Run model
|
||
pred = model(imgs)
|
||
|
||
# Compute loss
|
||
loss, loss_items = compute_loss(pred, targets, model, giou_loss=opt.giou)
|
||
if torch.isnan(loss):
|
||
print('WARNING: nan loss detected, ending training')
|
||
return results
|
||
|
||
# Compute gradient
|
||
if mixed_precision:
|
||
with amp.scale_loss(loss, optimizer) as scaled_loss:
|
||
scaled_loss.backward()
|
||
else:
|
||
loss.backward()
|
||
|
||
# Accumulate gradient for x batches before optimizing
|
||
if (i + 1) % accumulate == 0 or (i + 1) == nb:
|
||
optimizer.step()
|
||
optimizer.zero_grad()
|
||
|
||
# Print batch results
|
||
mloss = (mloss * i + loss_items) / (i + 1) # update mean losses
|
||
s = ('%8s%12s' + '%10.3g' * 7) % (
|
||
'%g/%g' % (epoch, epochs - 1),
|
||
'%g/%g' % (i, nb - 1), *mloss, len(targets), time.time() - t)
|
||
t = time.time()
|
||
print(s)
|
||
|
||
# Calculate mAP (always test final epoch, skip first 5 if opt.nosave)
|
||
if not (opt.notest or (opt.nosave and epoch < 10)) or epoch == epochs - 1:
|
||
with torch.no_grad():
|
||
results, maps = test.test(cfg, data_cfg, batch_size=batch_size, img_size=img_size_test, model=model,
|
||
conf_thres=0.1)
|
||
|
||
# Write epoch results
|
||
with open('results.txt', 'a') as file:
|
||
file.write(s + '%11.3g' * 5 % results + '\n') # P, R, mAP, F1, test_loss
|
||
|
||
# Update best loss
|
||
test_loss = results[4]
|
||
if test_loss < best_loss:
|
||
best_loss = test_loss
|
||
|
||
# Save training results
|
||
save = (not opt.nosave) or (epoch == epochs - 1)
|
||
if save:
|
||
# Create checkpoint
|
||
chkpt = {'epoch': epoch,
|
||
'best_loss': best_loss,
|
||
'model': model.module.state_dict() if type(
|
||
model) is nn.parallel.DistributedDataParallel else model.state_dict(),
|
||
'optimizer': optimizer.state_dict()}
|
||
|
||
# Save latest checkpoint
|
||
torch.save(chkpt, latest)
|
||
|
||
# Save best checkpoint
|
||
if best_loss == test_loss:
|
||
torch.save(chkpt, best)
|
||
|
||
# Save backup every 10 epochs (optional)
|
||
if epoch > 0 and epoch % 10 == 0:
|
||
torch.save(chkpt, weights + 'backup%g.pt' % epoch)
|
||
|
||
# Delete checkpoint
|
||
del chkpt
|
||
|
||
dt = (time.time() - t0) / 3600
|
||
print('%g epochs completed in %.3f hours.' % (epoch - start_epoch + 1, dt))
|
||
return results
|
||
|
||
|
||
def print_mutation(hyp, results):
|
||
# Write mutation results
|
||
a = '%11s' * len(hyp) % tuple(hyp.keys()) # hyperparam keys
|
||
b = '%11.4g' * len(hyp) % tuple(hyp.values()) # hyperparam values
|
||
c = '%11.3g' * len(results) % results # results (P, R, mAP, F1, test_loss)
|
||
print('\n%s\n%s\nEvolved fitness: %s\n' % (a, b, c))
|
||
with open('evolve.txt', 'a') as f:
|
||
f.write(c + b + '\n')
|
||
|
||
|
||
if __name__ == '__main__':
|
||
parser = argparse.ArgumentParser()
|
||
parser.add_argument('--epochs', type=int, default=68, help='number of epochs')
|
||
parser.add_argument('--batch-size', type=int, default=8, help='batch size')
|
||
parser.add_argument('--accumulate', type=int, default=8, help='number of batches to accumulate before optimizing')
|
||
parser.add_argument('--cfg', type=str, default='cfg/yolov3-spp.cfg', help='cfg file path')
|
||
parser.add_argument('--data-cfg', type=str, default='data/coco_64img.data', help='coco.data file path')
|
||
parser.add_argument('--single-scale', action='store_true', help='train at fixed size (no multi-scale)')
|
||
parser.add_argument('--img-size', type=int, default=416, help='inference size (pixels)')
|
||
parser.add_argument('--resume', action='store_true', help='resume training flag')
|
||
parser.add_argument('--transfer', action='store_true', help='transfer learning flag')
|
||
parser.add_argument('--num-workers', type=int, default=0, help='number of Pytorch DataLoader workers')
|
||
parser.add_argument('--dist-url', default='tcp://127.0.0.1:9999', type=str, help='distributed training init method')
|
||
parser.add_argument('--rank', default=0, type=int, help='distributed training node rank')
|
||
parser.add_argument('--world-size', default=1, type=int, help='number of nodes for distributed training')
|
||
parser.add_argument('--backend', default='nccl', type=str, help='distributed backend')
|
||
parser.add_argument('--nosave', action='store_true', help='do not save training results')
|
||
parser.add_argument('--notest', action='store_true', help='only test final epoch')
|
||
parser.add_argument('--giou', action='store_true', help='use GIoU loss instead of xy, wh loss')
|
||
parser.add_argument('--evolve', action='store_true', help='run hyperparameter evolution')
|
||
parser.add_argument('--var', default=0, type=int, help='debug variable')
|
||
opt = parser.parse_args()
|
||
print(opt)
|
||
|
||
if opt.evolve:
|
||
opt.notest = True # save time by only testing final epoch
|
||
opt.nosave = True # do not save checkpoints
|
||
|
||
# Train
|
||
results = train(
|
||
opt.cfg,
|
||
opt.data_cfg,
|
||
img_size=opt.img_size,
|
||
resume=opt.resume or opt.transfer,
|
||
transfer=opt.transfer,
|
||
epochs=opt.epochs,
|
||
batch_size=opt.batch_size,
|
||
accumulate=opt.accumulate,
|
||
)
|
||
|
||
# Evolve hyperparameters (optional)
|
||
if opt.evolve:
|
||
best_fitness = results[2] # use mAP for fitness
|
||
|
||
# Write mutation results
|
||
print_mutation(hyp, results)
|
||
|
||
gen = 1000 # generations to evolve
|
||
for _ in range(gen):
|
||
|
||
# Mutate hyperparameters
|
||
old_hyp = hyp.copy()
|
||
init_seeds(seed=int(time.time()))
|
||
s = [.4, .4, .4, .4, .4, .4, .4, .4, .4, .04, .4] # fractional sigmas
|
||
for i, k in enumerate(hyp.keys()):
|
||
x = (np.random.randn(1) * s[i] + 1) ** 1.1 # plt.hist(x.ravel(), 100)
|
||
hyp[k] = hyp[k] * float(x) # vary by about 30% 1sigma
|
||
|
||
# Clip to limits
|
||
keys = ['lr0', 'iou_t', 'momentum', 'weight_decay']
|
||
limits = [(1e-4, 1e-2), (0, 0.00), (0.70, 0.99), (0, 0.01)]
|
||
for k, v in zip(keys, limits):
|
||
hyp[k] = np.clip(hyp[k], v[0], v[1])
|
||
|
||
# Determine mutation fitness
|
||
results = train(
|
||
opt.cfg,
|
||
opt.data_cfg,
|
||
img_size=opt.img_size,
|
||
resume=opt.resume or opt.transfer,
|
||
transfer=opt.transfer,
|
||
epochs=opt.epochs,
|
||
batch_size=opt.batch_size,
|
||
accumulate=opt.accumulate,
|
||
)
|
||
mutation_fitness = results[2]
|
||
|
||
# Write mutation results
|
||
print_mutation(hyp, results)
|
||
|
||
# Update hyperparameters if fitness improved
|
||
if mutation_fitness > best_fitness:
|
||
# Fitness improved!
|
||
print('Fitness improved!')
|
||
best_fitness = mutation_fitness
|
||
else:
|
||
hyp = old_hyp.copy() # reset hyp to
|
||
|
||
# # Plot results
|
||
# import numpy as np
|
||
# import matplotlib.pyplot as plt
|
||
# a = np.loadtxt('evolve_1000val.txt')
|
||
# x = a[:, 2] * a[:, 3] # metric = mAP * F1
|
||
# weights = (x - x.min()) ** 2
|
||
# fig = plt.figure(figsize=(14, 7))
|
||
# for i in range(len(hyp)):
|
||
# y = a[:, i + 5]
|
||
# mu = (y * weights).sum() / weights.sum()
|
||
# plt.subplot(2, 5, i+1)
|
||
# plt.plot(x.max(), mu, 'o')
|
||
# plt.plot(x, y, '.')
|
||
# print(list(hyp.keys())[i],'%.4g' % mu)
|