Merge remote-tracking branch 'origin/master'
This commit is contained in:
Glenn Jocher
commit
3a7711856e
10
README.md
10
README.md
|
|
@ -5,7 +5,7 @@
|
|||
</td>
|
||||
<td align="center">
|
||||
<a href="https://www.ultralytics.com" target="_blank">
|
||||
<img src="https://storage.googleapis.com/ultralytics/logo/logoname1000.png" width="200"></a>
|
||||
<img src="https://storage.googleapis.com/ultralytics/logo/logoname1000.png" width="160"></a>
|
||||
<img src="https://user-images.githubusercontent.com/26833433/61591093-2b4d4480-abc2-11e9-8b46-d88eb1dabba1.jpg">
|
||||
<a href="https://itunes.apple.com/app/id1452689527" target="_blank">
|
||||
<img src="https://user-images.githubusercontent.com/26833433/50044365-9b22ac00-0082-11e9-862f-e77aee7aa7b0.png" width="180"></a>
|
||||
|
|
@ -43,8 +43,7 @@ Python 3.7 or later with the following `pip3 install -U -r requirements.txt` pac
|
|||
|
||||
# Jupyter Notebook
|
||||
|
||||
A jupyter notebook with training, inference and testing examples is available at:
|
||||
https://colab.research.google.com/drive/1G8T-VFxQkjDe4idzN8F-hbIBqkkkQnxw
|
||||
Our Jupyter [notebook](https://colab.research.google.com/github/ultralytics/yolov3/blob/master/examples.ipynb) provides quick training, inference and testing examples.
|
||||
|
||||
# Training
|
||||
|
||||
|
|
@ -87,10 +86,11 @@ https://cloud.google.com/deep-learning-vm/
|
|||
GPUs | `batch_size` | batch time | epoch time | epoch cost
|
||||
--- |---| --- | --- | ---
|
||||
1 K80 | 64 (32x2) | 2.9s | 175min | $0.58
|
||||
1 T4 | 64 (32x2) | 0.8s | 49min | $0.29
|
||||
1 T4 | 64 (32x2) | 0.80s | 49min | $0.29
|
||||
2 T4 | 64 (64x1) | 0.52s | 32min | $0.36
|
||||
1 2080ti | 64 (32x2) | - | - | -
|
||||
1 V100 | 64 (32x2) | 0.38s | 23min | $0.31
|
||||
2 V100 | 64 (64x1) | 0.38s | 23min | $0.62
|
||||
2 V100 | 64 (64x1) | 0.30s | 18min | $0.46
|
||||
|
||||
# Inference
|
||||
|
||||
|
|
|
|||
File diff suppressed because one or more lines are too long
|
|
@ -1,7 +1,3 @@
|
|||
import os
|
||||
|
||||
import torch.nn.functional as F
|
||||
|
||||
from utils.parse_config import *
|
||||
from utils.utils import *
|
||||
|
||||
|
|
@ -145,6 +141,7 @@ class YOLOLayer(nn.Module):
|
|||
return torch.cat((xy / ngu, wh, p_conf, p_cls), 2).squeeze().t()
|
||||
|
||||
else: # inference
|
||||
# s = 1.5 # scale_xy (pxy = pxy * s - (s - 1) / 2)
|
||||
io = p.clone() # inference output
|
||||
io[..., 0:2] = torch.sigmoid(io[..., 0:2]) + self.grid_xy # xy
|
||||
io[..., 2:4] = torch.exp(io[..., 2:4]) * self.anchor_wh # wh yolo method
|
||||
|
|
|
|||
|
|
@ -1,7 +1,7 @@
|
|||
# pip3 install -U -r requirements.txt
|
||||
# conda install numpy opencv matplotlib tqdm pillow
|
||||
# conda install pytorch torchvision -c pytorch
|
||||
# conda install -c conda-forge scikit-image
|
||||
# conda install scikit-image -c conda-forge
|
||||
numpy
|
||||
opencv-python
|
||||
torch >= 1.1.0
|
||||
|
|
|
|||
206
train.py
206
train.py
|
|
@ -4,19 +4,29 @@ 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.adabound import *
|
||||
from utils.datasets import *
|
||||
from utils.utils import *
|
||||
|
||||
mixed_precision = True
|
||||
try: # Mixed precision training https://github.com/NVIDIA/apex
|
||||
from apex import amp
|
||||
except: # not installed: install help: https://github.com/NVIDIA/apex/issues/259
|
||||
mixed_precision = False
|
||||
|
||||
# 320 --epochs 1
|
||||
# 0.109 0.297 0.15 0.126 7.04 1.666 4.062 0.1845 42.6 3.34 12.61 8.338 0.2705 0.001 -4 0.9 0.0005 a 320 giou + best_anchor False
|
||||
# 0.109 0.297 0.150 0.126 7.04 1.666 4.062 0.1845 42.6 3.34 12.61 8.338 0.2705 0.001 -4 0.9 0.0005 a 320 giou + best_anchor False
|
||||
# 0.223 0.218 0.138 0.189 9.28 1.153 4.376 0.08263 24.28 3.05 20.93 2.842 0.2759 0.001357 -5.036 0.9158 0.0005722 b mAP/F1 - 50/50 weighting
|
||||
# 0.231 0.215 0.135 0.191 9.51 1.432 3.007 0.06082 24.87 3.477 24.13 2.802 0.3436 0.001127 -5.036 0.9232 0.0005874 c
|
||||
# 0.246 0.194 0.128 0.192 8.12 1.101 3.954 0.0817 22.83 3.967 19.83 1.779 0.3352 0.000895 -5.036 0.9238 0.0007973 d
|
||||
# 0.187 0.237 0.144 0.186 14.6 1.607 4.202 0.09439 39.27 3.726 31.26 2.634 0.273 0.001542 -5.036 0.8364 0.0008393 e
|
||||
# 0.250 0.217 0.136 0.195 3.3 1.2 2 0.604 15.7 3.67 20 1.36 0.194 0.00128 -4 0.95 0.000201 0.8 0.388 1.2 0.119 0.0589 0.401 f
|
||||
# 0.269 0.225 0.149 0.218 6.71 1.13 5.25 0.246 22.4 3.64 17.8 1.31 0.256 0.00146 -4 0.936 0.00042 0.123 0.18 1.81 0.0987 0.0788 0.441 g
|
||||
# 0.179 0.274 0.165 0.187 7.95 1.22 7.62 0.224 17 5.71 17.7 3.28 0.295 0.00136 -4 0.875 0.000319 0.131 0.208 2.14 0.14 0.0773 0.228 h
|
||||
# 0.296 0.228 0.152 0.220 5.18 1.43 4.27 0.265 11.7 4.81 11.5 1.56 0.281 0.0013 -4 0.944 0.000427 0.0599 0.142 1.03 0.0552 0.0555 0.434 i
|
||||
|
||||
# 320 --epochs 2
|
||||
# 0.242 0.296 0.196 0.231 5.67 0.8541 4.286 0.1539 21.61 1.957 22.9 2.894 0.3689 0.001844 -4 0.913 0.000467 # ha 0.417 mAP @ epoch 100
|
||||
|
|
@ -25,40 +35,45 @@ from utils.utils import *
|
|||
# 0.161 0.327 0.190 0.193 7.82 1.153 4.062 0.1845 24.28 3.05 20.93 2.842 0.2759 0.001357 -4 0.916 0.000572 # hd 0.438 mAP @ epoch 100
|
||||
|
||||
|
||||
# Training hyperparameters d
|
||||
hyp = {'giou': 1.153, # giou loss gain
|
||||
'xy': 4.062, # xy loss gain
|
||||
'wh': 0.1845, # wh loss gain
|
||||
'cls': 24.28, # cls loss gain
|
||||
'cls_pw': 3.05, # cls BCELoss positive_weight
|
||||
'obj': 20.93, # obj loss gain
|
||||
'obj_pw': 2.842, # obj BCELoss positive_weight
|
||||
'iou_t': 0.2759, # iou training threshold
|
||||
'lr0': 0.001357, # initial learning rate
|
||||
'lrf': -4., # final LambdaLR learning rate = lr0 * (10 ** lrf)
|
||||
'momentum': 0.916, # SGD momentum
|
||||
'weight_decay': 0.0000572, # optimizer weight decay
|
||||
'hsv_s': 0.5, # image HSV-Saturation augmentation (fraction)
|
||||
'hsv_v': 0.5, # image HSV-Value augmentation (fraction)
|
||||
'degrees': 5, # image rotation (+/- deg)
|
||||
'translate': 0.1, # image translation (+/- fraction)
|
||||
'scale': 0.1, # image scale (+/- gain)
|
||||
'shear': 2} # image shear (+/- deg)
|
||||
|
||||
|
||||
# # Training hyperparameters e
|
||||
# hyp = {'giou': 1.607, # giou loss gain
|
||||
# 'xy': 4.062, # xy loss gain
|
||||
# 'wh': 0.1845, # wh loss gain
|
||||
# 'cls': 39.27, # cls loss gain
|
||||
# 'cls_pw': 3.726, # cls BCELoss positive_weight
|
||||
# 'obj': 31.26, # obj loss gain
|
||||
# 'obj_pw': 2.634, # obj BCELoss positive_weight
|
||||
# 'iou_t': 0.273, # iou target-anchor training threshold
|
||||
# 'lr0': 0.001542, # initial learning rate
|
||||
# Training hyperparameters g
|
||||
# hyp = {'giou': 1.13, # giou loss gain
|
||||
# 'xy': 5.25, # xy loss gain
|
||||
# 'wh': 0.246, # wh loss gain
|
||||
# 'cls': 22.4, # cls loss gain
|
||||
# 'cls_pw': 3.64, # cls BCELoss positive_weight
|
||||
# 'obj': 17.8, # obj loss gain
|
||||
# 'obj_pw': 1.31, # obj BCELoss positive_weight
|
||||
# 'iou_t': 0.256, # iou training threshold
|
||||
# 'lr0': 0.00146, # initial learning rate
|
||||
# 'lrf': -4., # final LambdaLR learning rate = lr0 * (10 ** lrf)
|
||||
# 'momentum': 0.8364, # SGD momentum
|
||||
# 'weight_decay': 0.0008393} # optimizer weight decay
|
||||
# 'momentum': 0.936, # SGD momentum
|
||||
# 'weight_decay': 0.00042, # optimizer weight decay
|
||||
# 'hsv_s': 0.123, # image HSV-Saturation augmentation (fraction)
|
||||
# 'hsv_v': 0.18, # image HSV-Value augmentation (fraction)
|
||||
# 'degrees': 1.81, # image rotation (+/- deg)
|
||||
# 'translate': 0.0987, # image translation (+/- fraction)
|
||||
# 'scale': 0.0788, # image scale (+/- gain)
|
||||
# 'shear': 0.441} # image shear (+/- deg)
|
||||
|
||||
# Training hyperparameters i
|
||||
hyp = {'giou': 1.43, # giou loss gain
|
||||
'xy': 4.27, # xy loss gain
|
||||
'wh': 0.265, # wh loss gain
|
||||
'cls': 11.7, # cls loss gain
|
||||
'cls_pw': 4.81, # cls BCELoss positive_weight
|
||||
'obj': 11.5, # obj loss gain
|
||||
'obj_pw': 1.56, # obj BCELoss positive_weight
|
||||
'iou_t': 0.281, # iou training threshold
|
||||
'lr0': 0.0013, # initial learning rate
|
||||
'lrf': -4., # final LambdaLR learning rate = lr0 * (10 ** lrf)
|
||||
'momentum': 0.944, # SGD momentum
|
||||
'weight_decay': 0.000427, # optimizer weight decay
|
||||
'hsv_s': 0.0599, # image HSV-Saturation augmentation (fraction)
|
||||
'hsv_v': 0.142, # image HSV-Value augmentation (fraction)
|
||||
'degrees': 1.03, # image rotation (+/- deg)
|
||||
'translate': 0.0552, # image translation (+/- fraction)
|
||||
'scale': 0.0555, # image scale (+/- gain)
|
||||
'shear': 0.434} # image shear (+/- deg)
|
||||
|
||||
|
||||
def train(cfg,
|
||||
|
|
@ -66,13 +81,13 @@ def train(cfg,
|
|||
img_size=416,
|
||||
epochs=100, # 500200 batches at bs 16, 117263 images = 273 epochs
|
||||
batch_size=16,
|
||||
accumulate=4): # effective bs = batch_size * accumulate = 8 * 8 = 64
|
||||
accumulate=4): # effective bs = batch_size * accumulate = 16 * 4 = 64
|
||||
# Initialize
|
||||
init_seeds()
|
||||
weights = 'weights' + os.sep
|
||||
last = weights + 'last.pt'
|
||||
best = weights + 'best.pt'
|
||||
device = torch_utils.select_device()
|
||||
device = torch_utils.select_device(apex=mixed_precision)
|
||||
multi_scale = opt.multi_scale
|
||||
|
||||
if multi_scale:
|
||||
|
|
@ -89,11 +104,13 @@ def train(cfg,
|
|||
model = Darknet(cfg).to(device)
|
||||
|
||||
# Optimizer
|
||||
optimizer = optim.SGD(model.parameters(), lr=hyp['lr0'], momentum=hyp['momentum'], weight_decay=hyp['weight_decay'])
|
||||
optimizer = optim.SGD(model.parameters(), lr=hyp['lr0'], momentum=hyp['momentum'], weight_decay=hyp['weight_decay'],
|
||||
nesterov=True)
|
||||
# optimizer = AdaBound(model.parameters(), lr=hyp['lr0'], final_lr=0.1)
|
||||
|
||||
cutoff = -1 # backbone reaches to cutoff layer
|
||||
start_epoch = 0
|
||||
best_fitness = 0.0
|
||||
best_fitness = 0.
|
||||
if opt.resume or opt.transfer: # Load previously saved model
|
||||
if opt.transfer: # Transfer learning
|
||||
nf = int(model.module_defs[model.yolo_layers[0] - 1]['filters']) # yolo layer size (i.e. 255)
|
||||
|
|
@ -136,7 +153,7 @@ def train(cfg,
|
|||
# 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 = lr_scheduler.MultiStepLR(optimizer, milestones=[round(opt.epochs * x) for x in [0.8]], gamma=0.1)
|
||||
scheduler.last_epoch = start_epoch - 1
|
||||
|
||||
# # Plot lr schedule
|
||||
|
|
@ -150,6 +167,18 @@ def train(cfg,
|
|||
# plt.tight_layout()
|
||||
# plt.savefig('LR.png', dpi=300)
|
||||
|
||||
# Mixed precision training https://github.com/NVIDIA/apex
|
||||
if mixed_precision:
|
||||
model, optimizer = amp.initialize(model, optimizer, opt_level='O1', verbosity=0)
|
||||
|
||||
# Initialize distributed training
|
||||
if torch.cuda.device_count() > 1:
|
||||
dist.init_process_group(backend='nccl', # 'distributed backend'
|
||||
init_method='tcp://127.0.0.1:9999', # distributed training init method
|
||||
world_size=1, # number of nodes for distributed training
|
||||
rank=0) # distributed training node rank
|
||||
model = torch.nn.parallel.DistributedDataParallel(model)
|
||||
|
||||
# Dataset
|
||||
dataset = LoadImagesAndLabels(train_path,
|
||||
img_size,
|
||||
|
|
@ -158,33 +187,14 @@ def train(cfg,
|
|||
hyp=hyp, # augmentation hyperparameters
|
||||
rect=opt.rect) # rectangular training
|
||||
|
||||
# Initialize distributed training
|
||||
if torch.cuda.device_count() > 1:
|
||||
dist.init_process_group(backend='nccl', # 'distributed backend'
|
||||
init_method='tcp://127.0.0.1:9999', # distributed training init method
|
||||
world_size=1, # number of nodes for distributed training
|
||||
rank=0) # distributed training node rank
|
||||
|
||||
model = torch.nn.parallel.DistributedDataParallel(model)
|
||||
# sampler = torch.utils.data.distributed.DistributedSampler(dataset)
|
||||
|
||||
# Dataloader
|
||||
dataloader = DataLoader(dataset,
|
||||
dataloader = torch.utils.data.DataLoader(dataset,
|
||||
batch_size=batch_size,
|
||||
num_workers=opt.num_workers,
|
||||
shuffle=not opt.rect, # Shuffle=True unless rectangular training is used
|
||||
pin_memory=True,
|
||||
collate_fn=dataset.collate_fn)
|
||||
|
||||
# Mixed precision training https://github.com/NVIDIA/apex
|
||||
mixed_precision = True
|
||||
if mixed_precision:
|
||||
try:
|
||||
from apex import amp
|
||||
model, optimizer = amp.initialize(model, optimizer, opt_level='O1', verbosity=0)
|
||||
except: # not installed: install help: https://github.com/NVIDIA/apex/issues/259
|
||||
mixed_precision = False
|
||||
|
||||
# Start training
|
||||
model.hyp = hyp # attach hyperparameters to model
|
||||
# model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) # attach class weights
|
||||
|
|
@ -192,7 +202,7 @@ def train(cfg,
|
|||
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
|
||||
# n_burnin = min(round(nb / 5 + 1), 1000) # burn-in batches
|
||||
t0 = time.time()
|
||||
for epoch in range(start_epoch, epochs):
|
||||
model.train()
|
||||
|
|
@ -234,11 +244,11 @@ def train(cfg,
|
|||
plot_images(imgs=imgs, targets=targets, paths=paths, fname='train_batch%g.jpg' % i)
|
||||
|
||||
# SGD burn-in
|
||||
if epoch == 0 and i <= n_burnin:
|
||||
g = (i / n_burnin) ** 4 # gain
|
||||
for x in optimizer.param_groups:
|
||||
x['lr'] = hyp['lr0'] * g
|
||||
x['weight_decay'] = hyp['weight_decay'] * g
|
||||
# if epoch == 0 and i <= n_burnin:
|
||||
# g = (i / n_burnin) ** 4 # gain
|
||||
# for x in optimizer.param_groups:
|
||||
# x['lr'] = hyp['lr0'] * g
|
||||
# x['weight_decay'] = hyp['weight_decay'] * g
|
||||
|
||||
# Run model
|
||||
pred = model(imgs)
|
||||
|
|
@ -313,33 +323,11 @@ def train(cfg,
|
|||
|
||||
# Report time
|
||||
print('%g epochs completed in %.3f hours.' % (epoch - start_epoch + 1, (time.time() - t0) / 3600))
|
||||
del model, optimizer
|
||||
dist.destroy_process_group() if torch.cuda.device_count() > 1 else None
|
||||
torch.cuda.empty_cache()
|
||||
return results
|
||||
|
||||
|
||||
def print_mutation(hyp, results):
|
||||
# Write mutation results
|
||||
a = '%11s' * len(hyp) % tuple(hyp.keys()) # hyperparam keys
|
||||
b = '%11.3g' * 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))
|
||||
|
||||
if opt.bucket:
|
||||
os.system('gsutil cp gs://%s/evolve.txt .' % opt.bucket) # download evolve.txt
|
||||
with open('evolve.txt', 'a') as f: # append result
|
||||
f.write(c + b + '\n')
|
||||
x = np.unique(np.loadtxt('evolve.txt', ndmin=2), axis=0) # load unique rows
|
||||
np.savetxt('evolve.txt', x[np.argsort(-fitness(x))], '%11.3g') # save sort by fitness
|
||||
os.system('gsutil cp evolve.txt gs://%s' % opt.bucket) # upload evolve.txt
|
||||
else:
|
||||
with open('evolve.txt', 'a') as f:
|
||||
f.write(c + b + '\n')
|
||||
|
||||
|
||||
def fitness(x): # returns fitness of hyp evolution vectors
|
||||
return x[:, 2] * 0.5 + x[:, 3] * 0.5 # fitness = weighted combination of mAP and F1
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument('--epochs', type=int, default=100, help='number of epochs')
|
||||
|
|
@ -362,11 +350,7 @@ if __name__ == '__main__':
|
|||
opt = parser.parse_args()
|
||||
print(opt)
|
||||
|
||||
if opt.evolve:
|
||||
opt.notest = True # only test final epoch
|
||||
opt.nosave = True # only save final checkpoint
|
||||
|
||||
# Train
|
||||
if not opt.evolve: # Train normally
|
||||
results = train(opt.cfg,
|
||||
opt.data,
|
||||
img_size=opt.img_size,
|
||||
|
|
@ -374,10 +358,14 @@ if __name__ == '__main__':
|
|||
batch_size=opt.batch_size,
|
||||
accumulate=opt.accumulate)
|
||||
|
||||
# Evolve hyperparameters (optional)
|
||||
if opt.evolve:
|
||||
print_mutation(hyp, results) # Write mutation results
|
||||
for _ in range(1000): # generations to evolve
|
||||
else: # Evolve hyperparameters (optional)
|
||||
opt.notest = True # only test final epoch
|
||||
opt.nosave = True # only save final checkpoint
|
||||
if opt.bucket:
|
||||
os.system('gsutil cp gs://%s/evolve.txt .' % opt.bucket) # download evolve.txt if exists
|
||||
|
||||
for _ in range(1): # generations to evolve
|
||||
if os.path.exists('evolve.txt'): # if evolve.txt exists: select best hyps and mutate
|
||||
# Get best hyperparameters
|
||||
x = np.loadtxt('evolve.txt', ndmin=2)
|
||||
x = x[fitness(x).argmax()] # select best fitness hyps
|
||||
|
|
@ -393,7 +381,7 @@ if __name__ == '__main__':
|
|||
|
||||
# Clip to limits
|
||||
keys = ['lr0', 'iou_t', 'momentum', 'weight_decay', 'hsv_s', 'hsv_v', 'translate', 'scale']
|
||||
limits = [(1e-4, 1e-2), (0.00, 0.70), (0.60, 0.95), (0, 0.001), (0, .8), (0, .8), (0, .8), (0, .8)]
|
||||
limits = [(1e-4, 1e-2), (0.00, 0.70), (0.60, 0.97), (0, 0.001), (0, .9), (0, .9), (0, .9), (0, .9)]
|
||||
for k, v in zip(keys, limits):
|
||||
hyp[k] = np.clip(hyp[k], v[0], v[1])
|
||||
|
||||
|
|
@ -406,19 +394,7 @@ if __name__ == '__main__':
|
|||
accumulate=opt.accumulate)
|
||||
|
||||
# Write mutation results
|
||||
print_mutation(hyp, results)
|
||||
print_mutation(hyp, results, opt.bucket)
|
||||
|
||||
# # Plot results
|
||||
# import numpy as np
|
||||
# import matplotlib.pyplot as plt
|
||||
# a = np.loadtxt('evolve.txt')
|
||||
# x = fitness(a)
|
||||
# weights = (x - x.min()) ** 2
|
||||
# fig = plt.figure(figsize=(10, 10))
|
||||
# for i in range(len(hyp)):
|
||||
# y = a[:, i + 5]
|
||||
# mu = (y * weights).sum() / weights.sum()
|
||||
# plt.subplot(4, 5, i + 1)
|
||||
# plt.plot(x.max(), mu, 'o')
|
||||
# plt.plot(x, y, '.')
|
||||
# print(list(hyp.keys())[i], '%.4g' % mu)
|
||||
# Plot results
|
||||
# plot_evolution_results(hyp)
|
||||
|
|
|
|||
|
|
@ -0,0 +1,236 @@
|
|||
import math
|
||||
|
||||
import torch
|
||||
from torch.optim import Optimizer
|
||||
|
||||
|
||||
class AdaBound(Optimizer):
|
||||
"""Implements AdaBound algorithm.
|
||||
It has been proposed in `Adaptive Gradient Methods with Dynamic Bound of Learning Rate`_.
|
||||
Arguments:
|
||||
params (iterable): iterable of parameters to optimize or dicts defining
|
||||
parameter groups
|
||||
lr (float, optional): Adam learning rate (default: 1e-3)
|
||||
betas (Tuple[float, float], optional): coefficients used for computing
|
||||
running averages of gradient and its square (default: (0.9, 0.999))
|
||||
final_lr (float, optional): final (SGD) learning rate (default: 0.1)
|
||||
gamma (float, optional): convergence speed of the bound functions (default: 1e-3)
|
||||
eps (float, optional): term added to the denominator to improve
|
||||
numerical stability (default: 1e-8)
|
||||
weight_decay (float, optional): weight decay (L2 penalty) (default: 0)
|
||||
amsbound (boolean, optional): whether to use the AMSBound variant of this algorithm
|
||||
.. Adaptive Gradient Methods with Dynamic Bound of Learning Rate:
|
||||
https://openreview.net/forum?id=Bkg3g2R9FX
|
||||
"""
|
||||
|
||||
def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), final_lr=0.1, gamma=1e-3,
|
||||
eps=1e-8, weight_decay=0, amsbound=False):
|
||||
if not 0.0 <= lr:
|
||||
raise ValueError("Invalid learning rate: {}".format(lr))
|
||||
if not 0.0 <= eps:
|
||||
raise ValueError("Invalid epsilon value: {}".format(eps))
|
||||
if not 0.0 <= betas[0] < 1.0:
|
||||
raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
|
||||
if not 0.0 <= betas[1] < 1.0:
|
||||
raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
|
||||
if not 0.0 <= final_lr:
|
||||
raise ValueError("Invalid final learning rate: {}".format(final_lr))
|
||||
if not 0.0 <= gamma < 1.0:
|
||||
raise ValueError("Invalid gamma parameter: {}".format(gamma))
|
||||
defaults = dict(lr=lr, betas=betas, final_lr=final_lr, gamma=gamma, eps=eps,
|
||||
weight_decay=weight_decay, amsbound=amsbound)
|
||||
super(AdaBound, self).__init__(params, defaults)
|
||||
|
||||
self.base_lrs = list(map(lambda group: group['lr'], self.param_groups))
|
||||
|
||||
def __setstate__(self, state):
|
||||
super(AdaBound, self).__setstate__(state)
|
||||
for group in self.param_groups:
|
||||
group.setdefault('amsbound', False)
|
||||
|
||||
def step(self, closure=None):
|
||||
"""Performs a single optimization step.
|
||||
Arguments:
|
||||
closure (callable, optional): A closure that reevaluates the model
|
||||
and returns the loss.
|
||||
"""
|
||||
loss = None
|
||||
if closure is not None:
|
||||
loss = closure()
|
||||
|
||||
for group, base_lr in zip(self.param_groups, self.base_lrs):
|
||||
for p in group['params']:
|
||||
if p.grad is None:
|
||||
continue
|
||||
grad = p.grad.data
|
||||
if grad.is_sparse:
|
||||
raise RuntimeError(
|
||||
'Adam does not support sparse gradients, please consider SparseAdam instead')
|
||||
amsbound = group['amsbound']
|
||||
|
||||
state = self.state[p]
|
||||
|
||||
# State initialization
|
||||
if len(state) == 0:
|
||||
state['step'] = 0
|
||||
# Exponential moving average of gradient values
|
||||
state['exp_avg'] = torch.zeros_like(p.data)
|
||||
# Exponential moving average of squared gradient values
|
||||
state['exp_avg_sq'] = torch.zeros_like(p.data)
|
||||
if amsbound:
|
||||
# Maintains max of all exp. moving avg. of sq. grad. values
|
||||
state['max_exp_avg_sq'] = torch.zeros_like(p.data)
|
||||
|
||||
exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
|
||||
if amsbound:
|
||||
max_exp_avg_sq = state['max_exp_avg_sq']
|
||||
beta1, beta2 = group['betas']
|
||||
|
||||
state['step'] += 1
|
||||
|
||||
if group['weight_decay'] != 0:
|
||||
grad = grad.add(group['weight_decay'], p.data)
|
||||
|
||||
# Decay the first and second moment running average coefficient
|
||||
exp_avg.mul_(beta1).add_(1 - beta1, grad)
|
||||
exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)
|
||||
if amsbound:
|
||||
# Maintains the maximum of all 2nd moment running avg. till now
|
||||
torch.max(max_exp_avg_sq, exp_avg_sq, out=max_exp_avg_sq)
|
||||
# Use the max. for normalizing running avg. of gradient
|
||||
denom = max_exp_avg_sq.sqrt().add_(group['eps'])
|
||||
else:
|
||||
denom = exp_avg_sq.sqrt().add_(group['eps'])
|
||||
|
||||
bias_correction1 = 1 - beta1 ** state['step']
|
||||
bias_correction2 = 1 - beta2 ** state['step']
|
||||
step_size = group['lr'] * math.sqrt(bias_correction2) / bias_correction1
|
||||
|
||||
# Applies bounds on actual learning rate
|
||||
# lr_scheduler cannot affect final_lr, this is a workaround to apply lr decay
|
||||
final_lr = group['final_lr'] * group['lr'] / base_lr
|
||||
lower_bound = final_lr * (1 - 1 / (group['gamma'] * state['step'] + 1))
|
||||
upper_bound = final_lr * (1 + 1 / (group['gamma'] * state['step']))
|
||||
step_size = torch.full_like(denom, step_size)
|
||||
step_size.div_(denom).clamp_(lower_bound, upper_bound).mul_(exp_avg)
|
||||
|
||||
p.data.add_(-step_size)
|
||||
|
||||
return loss
|
||||
|
||||
|
||||
class AdaBoundW(Optimizer):
|
||||
"""Implements AdaBound algorithm with Decoupled Weight Decay (arxiv.org/abs/1711.05101)
|
||||
It has been proposed in `Adaptive Gradient Methods with Dynamic Bound of Learning Rate`_.
|
||||
Arguments:
|
||||
params (iterable): iterable of parameters to optimize or dicts defining
|
||||
parameter groups
|
||||
lr (float, optional): Adam learning rate (default: 1e-3)
|
||||
betas (Tuple[float, float], optional): coefficients used for computing
|
||||
running averages of gradient and its square (default: (0.9, 0.999))
|
||||
final_lr (float, optional): final (SGD) learning rate (default: 0.1)
|
||||
gamma (float, optional): convergence speed of the bound functions (default: 1e-3)
|
||||
eps (float, optional): term added to the denominator to improve
|
||||
numerical stability (default: 1e-8)
|
||||
weight_decay (float, optional): weight decay (L2 penalty) (default: 0)
|
||||
amsbound (boolean, optional): whether to use the AMSBound variant of this algorithm
|
||||
.. Adaptive Gradient Methods with Dynamic Bound of Learning Rate:
|
||||
https://openreview.net/forum?id=Bkg3g2R9FX
|
||||
"""
|
||||
|
||||
def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), final_lr=0.1, gamma=1e-3,
|
||||
eps=1e-8, weight_decay=0, amsbound=False):
|
||||
if not 0.0 <= lr:
|
||||
raise ValueError("Invalid learning rate: {}".format(lr))
|
||||
if not 0.0 <= eps:
|
||||
raise ValueError("Invalid epsilon value: {}".format(eps))
|
||||
if not 0.0 <= betas[0] < 1.0:
|
||||
raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
|
||||
if not 0.0 <= betas[1] < 1.0:
|
||||
raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
|
||||
if not 0.0 <= final_lr:
|
||||
raise ValueError("Invalid final learning rate: {}".format(final_lr))
|
||||
if not 0.0 <= gamma < 1.0:
|
||||
raise ValueError("Invalid gamma parameter: {}".format(gamma))
|
||||
defaults = dict(lr=lr, betas=betas, final_lr=final_lr, gamma=gamma, eps=eps,
|
||||
weight_decay=weight_decay, amsbound=amsbound)
|
||||
super(AdaBoundW, self).__init__(params, defaults)
|
||||
|
||||
self.base_lrs = list(map(lambda group: group['lr'], self.param_groups))
|
||||
|
||||
def __setstate__(self, state):
|
||||
super(AdaBoundW, self).__setstate__(state)
|
||||
for group in self.param_groups:
|
||||
group.setdefault('amsbound', False)
|
||||
|
||||
def step(self, closure=None):
|
||||
"""Performs a single optimization step.
|
||||
Arguments:
|
||||
closure (callable, optional): A closure that reevaluates the model
|
||||
and returns the loss.
|
||||
"""
|
||||
loss = None
|
||||
if closure is not None:
|
||||
loss = closure()
|
||||
|
||||
for group, base_lr in zip(self.param_groups, self.base_lrs):
|
||||
for p in group['params']:
|
||||
if p.grad is None:
|
||||
continue
|
||||
grad = p.grad.data
|
||||
if grad.is_sparse:
|
||||
raise RuntimeError(
|
||||
'Adam does not support sparse gradients, please consider SparseAdam instead')
|
||||
amsbound = group['amsbound']
|
||||
|
||||
state = self.state[p]
|
||||
|
||||
# State initialization
|
||||
if len(state) == 0:
|
||||
state['step'] = 0
|
||||
# Exponential moving average of gradient values
|
||||
state['exp_avg'] = torch.zeros_like(p.data)
|
||||
# Exponential moving average of squared gradient values
|
||||
state['exp_avg_sq'] = torch.zeros_like(p.data)
|
||||
if amsbound:
|
||||
# Maintains max of all exp. moving avg. of sq. grad. values
|
||||
state['max_exp_avg_sq'] = torch.zeros_like(p.data)
|
||||
|
||||
exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
|
||||
if amsbound:
|
||||
max_exp_avg_sq = state['max_exp_avg_sq']
|
||||
beta1, beta2 = group['betas']
|
||||
|
||||
state['step'] += 1
|
||||
|
||||
# Decay the first and second moment running average coefficient
|
||||
exp_avg.mul_(beta1).add_(1 - beta1, grad)
|
||||
exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)
|
||||
if amsbound:
|
||||
# Maintains the maximum of all 2nd moment running avg. till now
|
||||
torch.max(max_exp_avg_sq, exp_avg_sq, out=max_exp_avg_sq)
|
||||
# Use the max. for normalizing running avg. of gradient
|
||||
denom = max_exp_avg_sq.sqrt().add_(group['eps'])
|
||||
else:
|
||||
denom = exp_avg_sq.sqrt().add_(group['eps'])
|
||||
|
||||
bias_correction1 = 1 - beta1 ** state['step']
|
||||
bias_correction2 = 1 - beta2 ** state['step']
|
||||
step_size = group['lr'] * math.sqrt(bias_correction2) / bias_correction1
|
||||
|
||||
# Applies bounds on actual learning rate
|
||||
# lr_scheduler cannot affect final_lr, this is a workaround to apply lr decay
|
||||
final_lr = group['final_lr'] * group['lr'] / base_lr
|
||||
lower_bound = final_lr * (1 - 1 / (group['gamma'] * state['step'] + 1))
|
||||
upper_bound = final_lr * (1 + 1 / (group['gamma'] * state['step']))
|
||||
step_size = torch.full_like(denom, step_size)
|
||||
step_size.div_(denom).clamp_(lower_bound, upper_bound).mul_(exp_avg)
|
||||
|
||||
if group['weight_decay'] != 0:
|
||||
decayed_weights = torch.mul(p.data, group['weight_decay'])
|
||||
p.data.add_(-step_size)
|
||||
p.data.sub_(decayed_weights)
|
||||
else:
|
||||
p.data.add_(-step_size)
|
||||
|
||||
return loss
|
||||
|
|
@ -8,9 +8,9 @@ from pathlib import Path
|
|||
import cv2
|
||||
import numpy as np
|
||||
import torch
|
||||
from PIL import Image, ExifTags
|
||||
from torch.utils.data import Dataset
|
||||
from tqdm import tqdm
|
||||
from PIL import Image, ExifTags
|
||||
|
||||
from utils.utils import xyxy2xywh, xywh2xyxy
|
||||
|
||||
|
|
@ -40,8 +40,6 @@ def exif_size(img):
|
|||
|
||||
class LoadImages: # for inference
|
||||
def __init__(self, path, img_size=416):
|
||||
self.height = img_size
|
||||
|
||||
files = []
|
||||
if os.path.isdir(path):
|
||||
files = sorted(glob.glob('%s/*.*' % path))
|
||||
|
|
@ -52,6 +50,7 @@ class LoadImages: # for inference
|
|||
videos = [x for x in files if os.path.splitext(x)[-1].lower() in vid_formats]
|
||||
nI, nV = len(images), len(videos)
|
||||
|
||||
self.img_size = img_size
|
||||
self.files = images + videos
|
||||
self.nF = nI + nV # number of files
|
||||
self.video_flag = [False] * nI + [True] * nV
|
||||
|
|
@ -96,7 +95,7 @@ class LoadImages: # for inference
|
|||
print('image %g/%g %s: ' % (self.count, self.nF, path), end='')
|
||||
|
||||
# Padded resize
|
||||
img, *_ = letterbox(img0, new_shape=self.height)
|
||||
img, *_ = letterbox(img0, new_shape=self.img_size)
|
||||
|
||||
# Normalize RGB
|
||||
img = img[:, :, ::-1].transpose(2, 0, 1) # BGR to RGB
|
||||
|
|
@ -117,8 +116,10 @@ class LoadImages: # for inference
|
|||
|
||||
class LoadWebcam: # for inference
|
||||
def __init__(self, img_size=416):
|
||||
self.cam = cv2.VideoCapture(0)
|
||||
self.height = img_size
|
||||
self.img_size = img_size
|
||||
self.cam = cv2.VideoCapture(0) # local camera
|
||||
# self.cam = cv2.VideoCapture('rtsp://192.168.1.64/1') # IP camera
|
||||
# self.cam = cv2.VideoCapture('rtsp://username:password@192.168.1.64/1') # IP camera with login
|
||||
|
||||
def __iter__(self):
|
||||
self.count = -1
|
||||
|
|
@ -138,7 +139,7 @@ class LoadWebcam: # for inference
|
|||
print('webcam %g: ' % self.count, end='')
|
||||
|
||||
# Padded resize
|
||||
img, *_ = letterbox(img0, new_shape=self.height)
|
||||
img, *_ = letterbox(img0, new_shape=self.img_size)
|
||||
|
||||
# Normalize RGB
|
||||
img = img[:, :, ::-1].transpose(2, 0, 1) # BGR to RGB
|
||||
|
|
@ -154,8 +155,7 @@ class LoadWebcam: # for inference
|
|||
class LoadImagesAndLabels(Dataset): # for training/testing
|
||||
def __init__(self, path, img_size=416, batch_size=16, augment=False, hyp=None, rect=False, image_weights=False):
|
||||
with open(path, 'r') as f:
|
||||
img_files = f.read().splitlines()
|
||||
self.img_files = [x for x in img_files if os.path.splitext(x)[-1].lower() in img_formats]
|
||||
self.img_files = [x for x in f.read().splitlines() if os.path.splitext(x)[-1].lower() in img_formats]
|
||||
|
||||
n = len(self.img_files)
|
||||
bi = np.floor(np.arange(n) / batch_size).astype(np.int) # batch index
|
||||
|
|
@ -405,10 +405,11 @@ def letterbox(img, new_shape=416, color=(128, 128, 128), mode='auto'):
|
|||
new_unpad = (new_shape, new_shape)
|
||||
ratiow, ratioh = new_shape / shape[1], new_shape / shape[0]
|
||||
|
||||
if shape[::-1] != new_unpad:
|
||||
img = cv2.resize(img, new_unpad, interpolation=cv2.INTER_AREA) # resize
|
||||
top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))
|
||||
left, right = int(round(dw - 0.1)), int(round(dw + 0.1))
|
||||
img = cv2.resize(img, new_unpad, interpolation=cv2.INTER_AREA) # resized, no border
|
||||
img = cv2.copyMakeBorder(img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color) # padded square
|
||||
img = cv2.copyMakeBorder(img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color) # add border
|
||||
return img, ratiow, ratioh, dw, dh
|
||||
|
||||
|
||||
|
|
|
|||
|
|
@ -28,6 +28,12 @@ python3 detect.py
|
|||
# Test
|
||||
python3 test.py --save-json
|
||||
|
||||
# Evolve
|
||||
for i in {0..500}
|
||||
do
|
||||
python3 train.py --data data/coco.data --img-size 320 --epochs 1 --batch-size 64 --accumulate 1 --evolve --bucket yolov4
|
||||
done
|
||||
|
||||
# Git pull
|
||||
git pull https://github.com/ultralytics/yolov3 # master
|
||||
git pull https://github.com/ultralytics/yolov3 test # branch
|
||||
|
|
|
|||
|
|
@ -9,24 +9,19 @@ def init_seeds(seed=0):
|
|||
# torch.backends.cudnn.deterministic = True # https://pytorch.org/docs/stable/notes/randomness.html
|
||||
|
||||
|
||||
def select_device(force_cpu=False):
|
||||
def select_device(force_cpu=False, apex=False):
|
||||
# apex if mixed precision training https://github.com/NVIDIA/apex
|
||||
cuda = False if force_cpu else torch.cuda.is_available()
|
||||
device = torch.device('cuda:0' if cuda else 'cpu')
|
||||
|
||||
if not cuda:
|
||||
print('Using CPU')
|
||||
if cuda:
|
||||
try: # Mixed precision training https://github.com/NVIDIA/apex
|
||||
from apex import amp
|
||||
apex_str = 'with Apex '
|
||||
except:
|
||||
apex_str = ''
|
||||
|
||||
torch.backends.cudnn.benchmark = True # set False for reproducible results
|
||||
c = 1024 ** 2 # bytes to MB
|
||||
ng = torch.cuda.device_count()
|
||||
x = [torch.cuda.get_device_properties(i) for i in range(ng)]
|
||||
cuda_str = 'Using CUDA ' + apex_str
|
||||
cuda_str = 'Using CUDA ' + ('Apex ' if apex else '')
|
||||
for i in range(0, ng):
|
||||
if i == 1:
|
||||
# torch.cuda.set_device(0) # OPTIONAL: Set GPU ID
|
||||
|
|
@ -42,14 +37,12 @@ def fuse_conv_and_bn(conv, bn):
|
|||
# https://tehnokv.com/posts/fusing-batchnorm-and-conv/
|
||||
with torch.no_grad():
|
||||
# init
|
||||
fusedconv = torch.nn.Conv2d(
|
||||
conv.in_channels,
|
||||
fusedconv = torch.nn.Conv2d(conv.in_channels,
|
||||
conv.out_channels,
|
||||
kernel_size=conv.kernel_size,
|
||||
stride=conv.stride,
|
||||
padding=conv.padding,
|
||||
bias=True
|
||||
)
|
||||
bias=True)
|
||||
|
||||
# prepare filters
|
||||
w_conv = conv.weight.clone().view(conv.out_channels, -1)
|
||||
|
|
|
|||
|
|
@ -1,5 +1,7 @@
|
|||
import glob
|
||||
import os
|
||||
import random
|
||||
from pathlib import Path
|
||||
|
||||
import cv2
|
||||
import matplotlib
|
||||
|
|
@ -9,7 +11,6 @@ import torch
|
|||
import torch.nn as nn
|
||||
from PIL import Image
|
||||
from tqdm import tqdm
|
||||
from pathlib import Path
|
||||
|
||||
from . import torch_utils # , google_utils
|
||||
|
||||
|
|
@ -303,12 +304,14 @@ def compute_loss(p, targets, model, giou_loss=True): # predictions, targets, mo
|
|||
tobj[b, a, gj, gi] = 1.0 # obj
|
||||
# pi[..., 2:4] = torch.sigmoid(pi[..., 2:4]) # wh power loss (uncomment)
|
||||
|
||||
# s = 1.5 # scale_xy
|
||||
pxy = torch.sigmoid(pi[..., 0:2]) # * s - (s - 1) / 2
|
||||
if giou_loss:
|
||||
pbox = torch.cat((torch.sigmoid(pi[..., 0:2]), torch.exp(pi[..., 2:4]) * anchor_vec[i]), 1) # predicted
|
||||
pbox = torch.cat((pxy, torch.exp(pi[..., 2:4]) * anchor_vec[i]), 1) # predicted
|
||||
giou = bbox_iou(pbox.t(), tbox[i], x1y1x2y2=False, GIoU=True) # giou computation
|
||||
lxy += (k * h['giou']) * (1.0 - giou).mean() # giou loss
|
||||
else:
|
||||
lxy += (k * h['xy']) * MSE(torch.sigmoid(pi[..., 0:2]), txy[i]) # xy loss
|
||||
lxy += (k * h['xy']) * MSE(pxy, txy[i]) # xy loss
|
||||
lwh += (k * h['wh']) * MSE(pi[..., 2:4], twh[i]) # wh yolo loss
|
||||
|
||||
tclsm = torch.zeros_like(pi[..., 5:])
|
||||
|
|
@ -542,23 +545,20 @@ def select_best_evolve(path='evolve*.txt'): # from utils.utils import *; select
|
|||
print(file, x[fitness.argmax()])
|
||||
|
||||
|
||||
def kmeans_targets(path='./data/coco_64img.txt'): # from utils.utils import *; kmeans_targets()
|
||||
def kmeans_targets(path='./data/coco_64img.txt', n=9, img_size=320): # from utils.utils import *; kmeans_targets()
|
||||
# Produces a list of target kmeans suitable for use in *.cfg files
|
||||
img_formats = ['.bmp', '.jpg', '.jpeg', '.png', '.tif']
|
||||
with open(path, 'r') as f:
|
||||
img_files = f.read().splitlines()
|
||||
img_files = list(filter(lambda x: len(x) > 0, img_files))
|
||||
img_files = [x for x in f.read().splitlines() if os.path.splitext(x)[-1].lower() in img_formats]
|
||||
|
||||
# Read shapes
|
||||
n = len(img_files)
|
||||
assert n > 0, 'No images found in %s' % path
|
||||
label_files = [x.replace('images', 'labels').
|
||||
replace('.jpeg', '.txt').
|
||||
replace('.jpg', '.txt').
|
||||
replace('.bmp', '.txt').
|
||||
replace('.png', '.txt') for x in img_files]
|
||||
nf = len(img_files)
|
||||
assert nf > 0, 'No images found in %s' % path
|
||||
label_files = [x.replace('images', 'labels').replace(os.path.splitext(x)[-1], '.txt') for x in img_files]
|
||||
s = np.array([Image.open(f).size for f in tqdm(img_files, desc='Reading image shapes')]) # (width, height)
|
||||
|
||||
# Read targets
|
||||
labels = [np.zeros((0, 5))] * n
|
||||
labels = [np.zeros((0, 5))] * nf
|
||||
iter = tqdm(label_files, desc='Reading labels')
|
||||
for i, file in enumerate(iter):
|
||||
try:
|
||||
|
|
@ -570,19 +570,43 @@ def kmeans_targets(path='./data/coco_64img.txt'): # from utils.utils import *;
|
|||
assert (l[:, 1:] <= 1).all(), 'non-normalized or out of bounds coordinate labels: %s' % file
|
||||
l[:, [1, 3]] *= s[i][0]
|
||||
l[:, [2, 4]] *= s[i][1]
|
||||
l[:, 1:] *= 320 / max(s[i])
|
||||
l[:, 1:] *= img_size / max(s[i]) # nominal img_size for training here
|
||||
labels[i] = l
|
||||
except:
|
||||
pass # print('Warning: missing labels for %s' % self.img_files[i]) # missing label file
|
||||
assert len(np.concatenate(labels, 0)) > 0, 'No labels found. Incorrect label paths provided.'
|
||||
|
||||
# kmeans
|
||||
# kmeans calculation
|
||||
from scipy import cluster
|
||||
wh = np.concatenate(labels, 0)[:, 3:5]
|
||||
k = cluster.vq.kmeans(wh, 9)[0]
|
||||
k = cluster.vq.kmeans(wh, n)[0]
|
||||
k = k[np.argsort(k.prod(1))]
|
||||
for x in k.ravel():
|
||||
print('%.1f, ' % x, end='')
|
||||
print('%.1f, ' % x, end='') # drop-in replacement for *.cfg anchors
|
||||
|
||||
|
||||
def print_mutation(hyp, results, bucket=''):
|
||||
# Print mutation results to evolve.txt (for use with train.py --evolve)
|
||||
a = '%11s' * len(hyp) % tuple(hyp.keys()) # hyperparam keys
|
||||
b = '%11.3g' * 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))
|
||||
|
||||
if bucket:
|
||||
os.system('gsutil cp gs://%s/evolve.txt .' % bucket) # download evolve.txt
|
||||
with open('evolve.txt', 'a') as f: # append result
|
||||
f.write(c + b + '\n')
|
||||
x = np.unique(np.loadtxt('evolve.txt', ndmin=2), axis=0) # load unique rows
|
||||
np.savetxt('evolve.txt', x[np.argsort(-fitness(x))], '%11.3g') # save sort by fitness
|
||||
os.system('gsutil cp evolve.txt gs://%s' % bucket) # upload evolve.txt
|
||||
else:
|
||||
with open('evolve.txt', 'a') as f:
|
||||
f.write(c + b + '\n')
|
||||
|
||||
|
||||
def fitness(x):
|
||||
# Returns fitness (for use with results.txt or evolve.txt)
|
||||
return 0.50 * x[:, 2] + 0.50 * x[:, 3] # fitness = 0.9 * mAP + 0.1 * F1
|
||||
|
||||
|
||||
# Plotting functions ---------------------------------------------------------------------------------------------------
|
||||
|
|
@ -617,7 +641,7 @@ def plot_wh_methods(): # from utils.utils import *; plot_wh_methods()
|
|||
plt.ylabel('output')
|
||||
plt.legend()
|
||||
fig.tight_layout()
|
||||
fig.savefig('comparison.png', dpi=300)
|
||||
fig.savefig('comparison.png', dpi=200)
|
||||
|
||||
|
||||
def plot_images(imgs, targets, paths=None, fname='images.jpg'):
|
||||
|
|
@ -642,7 +666,7 @@ def plot_images(imgs, targets, paths=None, fname='images.jpg'):
|
|||
s = Path(paths[i]).name
|
||||
plt.title(s[:min(len(s), 40)], fontdict={'size': 8}) # limit to 40 characters
|
||||
fig.tight_layout()
|
||||
fig.savefig(fname, dpi=300)
|
||||
fig.savefig(fname, dpi=200)
|
||||
plt.close()
|
||||
|
||||
|
||||
|
|
@ -662,7 +686,7 @@ def plot_test_txt(): # from utils.utils import *; plot_test()
|
|||
ax[0].hist(cx, bins=600)
|
||||
ax[1].hist(cy, bins=600)
|
||||
fig.tight_layout()
|
||||
plt.savefig('hist1d.jpg', dpi=300)
|
||||
plt.savefig('hist1d.jpg', dpi=200)
|
||||
|
||||
|
||||
def plot_targets_txt(): # from utils.utils import *; plot_targets_txt()
|
||||
|
|
@ -678,7 +702,27 @@ def plot_targets_txt(): # from utils.utils import *; plot_targets_txt()
|
|||
ax[i].legend()
|
||||
ax[i].set_title(s[i])
|
||||
fig.tight_layout()
|
||||
plt.savefig('targets.jpg', dpi=300)
|
||||
plt.savefig('targets.jpg', dpi=200)
|
||||
|
||||
|
||||
def plot_evolution_results(hyp): # from utils.utils import *; plot_evolution_results(hyp)
|
||||
# Plot hyperparameter evolution results in evolve.txt
|
||||
x = np.loadtxt('evolve.txt')
|
||||
f = fitness(x)
|
||||
weights = (f - f.min()) ** 2 # for weighted results
|
||||
fig = plt.figure(figsize=(12, 10))
|
||||
matplotlib.rc('font', **{'size': 8})
|
||||
for i, (k, v) in enumerate(hyp.items()):
|
||||
y = x[:, i + 5]
|
||||
# mu = (y * weights).sum() / weights.sum() # best weighted result
|
||||
mu = y[f.argmax()] # best single result
|
||||
plt.subplot(4, 5, i + 1)
|
||||
plt.plot(mu, f.max(), 'o', markersize=10)
|
||||
plt.plot(y, f, '.')
|
||||
plt.title('%s = %.3g' % (k, mu), fontdict={'size': 9}) # limit to 40 characters
|
||||
print('%15s: %.3g' % (k, mu))
|
||||
fig.tight_layout()
|
||||
plt.savefig('evolve.png', dpi=200)
|
||||
|
||||
|
||||
def plot_results(start=0, stop=0): # from utils.utils import *; plot_results()
|
||||
|
|
@ -698,4 +742,4 @@ def plot_results(start=0, stop=0): # from utils.utils import *; plot_results()
|
|||
ax[i].set_title(s[i])
|
||||
fig.tight_layout()
|
||||
ax[4].legend()
|
||||
fig.savefig('results.png', dpi=300)
|
||||
fig.savefig('results.png', dpi=200)
|
||||
|
|
|
|||
Loading…
Reference in New Issue