Bienqq
/
car-detection-bayes
2
1
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Merge remote-tracking branch 'origin/master'

This commit is contained in:
Glenn Jocher 2019-07-04 14:03:21 +02:00
parent 7a353a9c70 109991198c
commit a5592093ef
9 changed files with 169 additions and 143 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
.gitignore vendored
View File

@ -30,6 +30,7 @@ data/*
!data/trainvalno5k.shapes !data/trainvalno5k.shapes
!data/5k.shapes !data/5k.shapes
!data/5k.txt !data/5k.txt
!data/*.sh
pycocotools/* pycocotools/*
results*.txt results*.txt

1
README.md
View File

@ -211,7 +211,6 @@ Computing mAP: 100%|████████████████████
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.331 Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.331
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.517 Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.517
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.618 Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.618
``` ```
# Citation # Citation

11
data/get_coco_dataset_gdrive.sh Executable file
View File

@ -0,0 +1,11 @@
#!/bin/bash
# https://stackoverflow.com/questions/48133080/how-to-download-a-google-drive-url-via-curl-or-wget/48133859
# Download COCO dataset
fileid="1HaXkef9z6y5l4vUnCYgdmEAj61c6bfWO"
filename="coco_gdrive.zip"
curl -c ./cookie -s -L "https://drive.google.com/uc?export=download&id=${fileid}" > /dev/null
curl -Lb ./cookie "https://drive.google.com/uc?export=download&confirm=`awk '/download/ {print $NF}' ./cookie`&id=${fileid}" -o ${filename}
# Unzip
unzip -q coco_gdrive.zip

24
detect.py
View File

@ -127,20 +127,18 @@ if __name__ == '__main__':
parser.add_argument('--img-size', type=int, default=416, help='inference size (pixels)') parser.add_argument('--img-size', type=int, default=416, help='inference size (pixels)')
parser.add_argument('--conf-thres', type=float, default=0.5, help='object confidence threshold') parser.add_argument('--conf-thres', type=float, default=0.5, help='object confidence threshold')
parser.add_argument('--nms-thres', type=float, default=0.5, help='iou threshold for non-maximum suppression') parser.add_argument('--nms-thres', type=float, default=0.5, help='iou threshold for non-maximum suppression')
parser.add_argument('--fourcc', type=str, default='mp4v', help='specifies the fourcc code for output video encoding (make sure ffmpeg supports specified fourcc codec)') parser.add_argument('--fourcc', type=str, default='mp4v', help='fourcc output video codec (verify ffmpeg support)')
parser.add_argument('--output', type=str, default='output',help='specifies the output path for images and videos') parser.add_argument('--output', type=str, default='output', help='specifies the output path for images and videos')
opt = parser.parse_args() opt = parser.parse_args()
print(opt) print(opt)
with torch.no_grad(): with torch.no_grad():
detect( detect(opt.cfg,
opt.cfg, opt.data_cfg,
opt.data_cfg, opt.weights,
opt.weights, images=opt.images,
images=opt.images, img_size=opt.img_size,
img_size=opt.img_size, conf_thres=opt.conf_thres,
conf_thres=opt.conf_thres, nms_thres=opt.nms_thres,
nms_thres=opt.nms_thres, fourcc=opt.fourcc,
fourcc=opt.fourcc, output=opt.output)
output=opt.output
)

24
models.py
View File

@ -15,6 +15,7 @@ def create_modules(module_defs):
hyperparams = module_defs.pop(0) hyperparams = module_defs.pop(0)
output_filters = [int(hyperparams['channels'])] output_filters = [int(hyperparams['channels'])]
module_list = nn.ModuleList() module_list = nn.ModuleList()
yolo_index = -1
for i, module_def in enumerate(module_defs): for i, module_def in enumerate(module_defs):
modules = nn.Sequential() modules = nn.Sequential()
@ -44,8 +45,7 @@ def create_modules(module_defs):
modules.add_module('maxpool_%d' % i, maxpool) modules.add_module('maxpool_%d' % i, maxpool)
elif module_def['type'] == 'upsample': elif module_def['type'] == 'upsample':
# upsample = nn.Upsample(scale_factor=int(module_def['stride']), mode='nearest') # WARNING: deprecated upsample = nn.Upsample(scale_factor=int(module_def['stride']), mode='nearest')
upsample = Upsample(scale_factor=int(module_def['stride']))
modules.add_module('upsample_%d' % i, upsample) modules.add_module('upsample_%d' % i, upsample)
elif module_def['type'] == 'route': elif module_def['type'] == 'route':
@ -58,6 +58,7 @@ def create_modules(module_defs):
modules.add_module('shortcut_%d' % i, EmptyLayer()) modules.add_module('shortcut_%d' % i, EmptyLayer())
elif module_def['type'] == 'yolo': elif module_def['type'] == 'yolo':
yolo_index += 1
anchor_idxs = [int(x) for x in module_def['mask'].split(',')] anchor_idxs = [int(x) for x in module_def['mask'].split(',')]
# Extract anchors # Extract anchors
anchors = [float(x) for x in module_def['anchors'].split(',')] anchors = [float(x) for x in module_def['anchors'].split(',')]
@ -66,8 +67,7 @@ def create_modules(module_defs):
nc = int(module_def['classes']) # number of classes nc = int(module_def['classes']) # number of classes
img_size = hyperparams['height'] img_size = hyperparams['height']
# Define detection layer # Define detection layer
yolo_layer = YOLOLayer(anchors, nc, img_size, cfg=hyperparams['cfg']) modules.add_module('yolo_%d' % i, YOLOLayer(anchors, nc, img_size, yolo_index))
modules.add_module('yolo_%d' % i, yolo_layer)
# Register module list and number of output filters # Register module list and number of output filters
module_list.append(modules) module_list.append(modules)
@ -86,20 +86,8 @@ class EmptyLayer(nn.Module):
return x return x
class Upsample(nn.Module):
# Custom Upsample layer (nn.Upsample gives deprecated warning message)
def __init__(self, scale_factor=1, mode='nearest'):
super(Upsample, self).__init__()
self.scale_factor = scale_factor
self.mode = mode
def forward(self, x):
return F.interpolate(x, scale_factor=self.scale_factor, mode=self.mode)
class YOLOLayer(nn.Module): class YOLOLayer(nn.Module):
def __init__(self, anchors, nc, img_size, cfg): def __init__(self, anchors, nc, img_size, yolo_index):
super(YOLOLayer, self).__init__() super(YOLOLayer, self).__init__()
self.anchors = torch.Tensor(anchors) self.anchors = torch.Tensor(anchors)
@ -109,7 +97,7 @@ class YOLOLayer(nn.Module):
self.ny = 0 # initialize number of y gridpoints self.ny = 0 # initialize number of y gridpoints
if ONNX_EXPORT: # grids must be computed in __init__ if ONNX_EXPORT: # grids must be computed in __init__
stride = [32, 16, 8][yolo_layer] # stride of this layer stride = [32, 16, 8][yolo_index] # stride of this layer
nx = int(img_size[1] / stride) # number x grid points nx = int(img_size[1] / stride) # number x grid points
ny = int(img_size[0] / stride) # number y grid points ny = int(img_size[0] / stride) # number y grid points
create_grids(self, max(img_size), (nx, ny)) create_grids(self, max(img_size), (nx, ny))

20
test.py
View File

@ -201,14 +201,12 @@ if __name__ == '__main__':
print(opt) print(opt)
with torch.no_grad(): with torch.no_grad():
mAP = test( mAP = test(opt.cfg,
opt.cfg, opt.data_cfg,
opt.data_cfg, opt.weights,
opt.weights, opt.batch_size,
opt.batch_size, opt.img_size,
opt.img_size, opt.iou_thres,
opt.iou_thres, opt.conf_thres,
opt.conf_thres, opt.nms_thres,
opt.nms_thres, opt.save_json)
opt.save_json
)

186
train.py
View File

@ -11,27 +11,31 @@ from models import *
from utils.datasets import * from utils.datasets import *
from utils.utils import * from utils.utils import *
# Hyperparameters: train.py --data data/coco.data --img-size 320 --single-scale --batch-size 64 --accumulate 1 --epochs 1 --evolve 0.087 0.281 0.109 0.121 # 0.149 0.241 0.126 0.156 6.85 1.008 1.421 0.07989 16.94 6.215 10.61 4.272 0.251 0.001 -4 0.9 0.0005 320 64-1 giou
hyp = {'giou': .035, # giou loss gain # 0.111 0.27 0.132 0.131 3.96 1.276 0.3156 0.1425 21.21 6.224 11.59 8.83 0.376 0.001 -4 0.9 0.0005
'xy': 0.20, # xy loss gain hyp = {'giou': 1.008, # giou loss gain
'wh': 0.10, # wh loss gain 'xy': 1.421, # xy loss gain
'cls': 0.035, # cls loss gain 'wh': 0.07989, # wh loss gain
'cls_pw': 79.0, # cls BCELoss positive_weight 'cls': 16.94, # cls loss gain
'conf': 1.61, # conf loss gain 'cls_pw': 6.215, # cls BCELoss positive_weight
'conf_pw': 3.53, # conf BCELoss positive_weight 'conf': 10.61, # conf loss gain
'iou_t': 0.29, # iou target-anchor training threshold 'conf_pw': 4.272, # conf BCELoss positive_weight
'iou_t': 0.251, # iou target-anchor training threshold
'lr0': 0.001, # initial learning rate 'lr0': 0.001, # initial learning rate
'lrf': -4., # final learning rate = lr0 * (10 ** lrf) 'lrf': -4., # final learning rate = lr0 * (10 ** lrf)
'momentum': 0.90, # SGD momentum 'momentum': 0.90, # SGD momentum
'weight_decay': 0.0005} # optimizer weight decay 'weight_decay': 0.0005} # optimizer weight decay
# hyp = {'giou': 1.0, # giou loss gain
# 'xy': 1.0, # xy loss gain # 0.0945 0.279 0.114 0.131 25 0.035 0.2 0.1 0.035 79 1.61 3.53 0.29 0.001 -4 0.9 0.0005 320 64-1
# 'wh': 1.0, # wh loss gain # 0.112 0.265 0.111 0.144 12.6 0.035 0.2 0.1 0.035 79 1.61 3.53 0.29 0.001 -4 0.9 0.0005 320 32-2
# 'cls': 1.0, # cls loss gain # hyp = {'giou': .035, # giou loss gain
# 'xy': 0.20, # xy loss gain
# 'wh': 0.10, # wh loss gain
# 'cls': 0.035, # cls loss gain
# 'cls_pw': 79.0, # cls BCELoss positive_weight # 'cls_pw': 79.0, # cls BCELoss positive_weight
# 'conf': 1.0, # conf loss gain # 'conf': 1.61, # conf loss gain
# 'conf_pw': 6.0, # conf BCELoss positive_weight # 'conf_pw': 3.53, # conf BCELoss positive_weight
# 'iou_t': 0.29, # iou target-anchor training threshold # 'iou_t': 0.29, # iou target-anchor training threshold
# 'lr0': 0.001, # initial learning rate # 'lr0': 0.001, # initial learning rate
# 'lrf': -4., # final learning rate = lr0 * (10 ** lrf) # 'lrf': -4., # final learning rate = lr0 * (10 ** lrf)
@ -43,12 +47,10 @@ def train(
cfg, cfg,
data_cfg, data_cfg,
img_size=416, img_size=416,
resume=False, epochs=100, # 500200 batches at bs 16, 117263 images = 273 epochs
epochs=100, # 500200 batches at bs 4, 117263 images = 68 epochs batch_size=8,
batch_size=16, accumulate=8, # effective bs = batch_size * accumulate = 8 * 8 = 64
accumulate=4, # effective bs = 64 = batch_size * accumulate
freeze_backbone=False, freeze_backbone=False,
transfer=False # Transfer learning (train only YOLO layers)
): ):
init_seeds() init_seeds()
weights = 'weights' + os.sep weights = 'weights' + os.sep
@ -76,10 +78,10 @@ def train(
cutoff = -1 # backbone reaches to cutoff layer cutoff = -1 # backbone reaches to cutoff layer
start_epoch = 0 start_epoch = 0
best_loss = float('inf') best_fitness = 0.0
nf = int(model.module_defs[model.yolo_layers[0] - 1]['filters']) # yolo layer size (i.e. 255) nf = int(model.module_defs[model.yolo_layers[0] - 1]['filters']) # yolo layer size (i.e. 255)
if resume: # Load previously saved model if opt.resume or opt.transfer: # Load previously saved model
if transfer: # Transfer learning if opt.transfer: # Transfer learning
chkpt = torch.load(weights + 'yolov3-spp.pt', map_location=device) 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}, model.load_state_dict({k: v for k, v in chkpt['model'].items() if v.numel() > 1 and v.shape[0] != 255},
strict=False) strict=False)
@ -93,7 +95,7 @@ def train(
start_epoch = chkpt['epoch'] + 1 start_epoch = chkpt['epoch'] + 1
if chkpt['optimizer'] is not None: if chkpt['optimizer'] is not None:
optimizer.load_state_dict(chkpt['optimizer']) optimizer.load_state_dict(chkpt['optimizer'])
best_loss = chkpt['best_loss'] best_fitness = chkpt['best_fitness']
del chkpt del chkpt
else: # Initialize model with backbone (optional) else: # Initialize model with backbone (optional)
@ -135,7 +137,11 @@ def train(
# Initialize distributed training # Initialize distributed training
if torch.cuda.device_count() > 1: 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) 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) model = torch.nn.parallel.DistributedDataParallel(model)
# sampler = torch.utils.data.distributed.DistributedSampler(dataset) # sampler = torch.utils.data.distributed.DistributedSampler(dataset)
@ -148,12 +154,13 @@ def train(
collate_fn=dataset.collate_fn) collate_fn=dataset.collate_fn)
# Mixed precision training https://github.com/NVIDIA/apex # Mixed precision training https://github.com/NVIDIA/apex
try: mixed_precision = True
from apex import amp if mixed_precision:
model, optimizer = amp.initialize(model, optimizer, opt_level='O1') try:
mixed_precision = True from apex import amp
except: # not installed: install help: https://github.com/NVIDIA/apex/issues/259 model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
mixed_precision = False except: # not installed: install help: https://github.com/NVIDIA/apex/issues/259
mixed_precision = False
# Start training # Start training
model.hyp = hyp # attach hyperparameters to model model.hyp = hyp # attach hyperparameters to model
@ -166,7 +173,8 @@ def train(
t, t0 = time.time(), time.time() t, t0 = time.time(), time.time()
for epoch in range(start_epoch, epochs): for epoch in range(start_epoch, epochs):
model.train() model.train()
print(('\n%8s%12s' + '%10s' * 7) % ('Epoch', 'Batch', 'xy', 'wh', 'conf', 'cls', 'total', 'targets', 'time')) print(('\n%8s%12s' + '%10s' * 7) %
('Epoch', 'Batch', 'xy', 'wh', 'conf', 'cls', 'total', 'targets', 'img_size'))
# Update scheduler # Update scheduler
scheduler.step() scheduler.step()
@ -183,15 +191,16 @@ def train(
# dataset.indices = random.choices(range(dataset.n), weights=image_weights, k=dataset.n) # random weighted index # dataset.indices = random.choices(range(dataset.n), weights=image_weights, k=dataset.n) # random weighted index
mloss = torch.zeros(5).to(device) # mean losses mloss = torch.zeros(5).to(device) # mean losses
for i, (imgs, targets, _, _) in enumerate(dataloader): pbar = tqdm(enumerate(dataloader), total=nb) # progress bar
for i, (imgs, targets, _, _) in pbar:
imgs = imgs.to(device) imgs = imgs.to(device)
targets = targets.to(device) targets = targets.to(device)
# Multi-Scale training # Multi-Scale training TODO: short-side to 32-multiple https://github.com/ultralytics/yolov3/issues/358
if multi_scale: if multi_scale:
if (i + 1 + nb * epoch) / accumulate % 10 == 0: #  adjust (67% - 150%) every 10 batches if (i + nb * epoch) / accumulate % 10 == 0: #  adjust (67% - 150%) every 10 batches
img_size = random.choice(range(img_size_min, img_size_max + 1)) * 32 img_size = random.choice(range(img_size_min, img_size_max + 1)) * 32
print('img_size = %g' % img_size) # print('img_size = %g' % img_size)
scale_factor = img_size / max(imgs.shape[-2:]) scale_factor = img_size / max(imgs.shape[-2:])
imgs = F.interpolate(imgs, scale_factor=scale_factor, mode='bilinear', align_corners=False) imgs = F.interpolate(imgs, scale_factor=scale_factor, mode='bilinear', align_corners=False)
@ -228,11 +237,11 @@ def train(
# Print batch results # Print batch results
mloss = (mloss * i + loss_items) / (i + 1) # update mean losses 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)
s = ('%8s%12s' + '%10.3g' * 7) % ( s = ('%8s%12s' + '%10.3g' * 7) % (
'%g/%g' % (epoch, epochs - 1), '%g/%g' % (epoch, epochs - 1), '%g/%g' % (i, nb - 1), *mloss, len(targets), img_size)
'%g/%g' % (i, nb - 1), *mloss, len(targets), time.time() - t)
t = time.time() t = time.time()
print(s) pbar.set_description(s) # print(s)
# Report time # Report time
dt = (time.time() - t0) / 3600 dt = (time.time() - t0) / 3600
@ -248,17 +257,17 @@ def train(
with open('results.txt', 'a') as file: with open('results.txt', 'a') as file:
file.write(s + '%11.3g' * 5 % results + '\n') # P, R, mAP, F1, test_loss file.write(s + '%11.3g' * 5 % results + '\n') # P, R, mAP, F1, test_loss
# Update best loss # Update best map
test_loss = results[4] fitness = results[2]
if test_loss < best_loss: if fitness > best_fitness:
best_loss = test_loss best_fitness = fitness
# Save training results # Save training results
save = (not opt.nosave) or (epoch == epochs - 1) save = (not opt.nosave) or (epoch == epochs - 1)
if save: if save:
# Create checkpoint # Create checkpoint
chkpt = {'epoch': epoch, chkpt = {'epoch': epoch,
'best_loss': best_loss, 'best_fitness': best_fitness,
'model': model.module.state_dict() if type( 'model': model.module.state_dict() if type(
model) is nn.parallel.DistributedDataParallel else model.state_dict(), model) is nn.parallel.DistributedDataParallel else model.state_dict(),
'optimizer': optimizer.state_dict()} 'optimizer': optimizer.state_dict()}
@ -267,7 +276,7 @@ def train(
torch.save(chkpt, latest) torch.save(chkpt, latest)
# Save best checkpoint # Save best checkpoint
if best_loss == test_loss: if best_fitness == fitness:
torch.save(chkpt, best) torch.save(chkpt, best)
# Save backup every 10 epochs (optional) # Save backup every 10 epochs (optional)
@ -286,8 +295,15 @@ def print_mutation(hyp, results):
b = '%11.4g' * len(hyp) % tuple(hyp.values()) # hyperparam values b = '%11.4g' * len(hyp) % tuple(hyp.values()) # hyperparam values
c = '%11.3g' * len(results) % results # results (P, R, mAP, F1, test_loss) 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)) 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 opt.cloud_evolve:
os.system('gsutil cp gs://yolov4/evolve.txt .') # download evolve.txt
with open('evolve.txt', 'a') as f: # append result to evolve.txt
f.write(c + b + '\n')
os.system('gsutil cp evolve.txt gs://yolov4') # upload evolve.txt
else:
with open('evolve.txt', 'a') as f:
f.write(c + b + '\n')
if __name__ == '__main__': if __name__ == '__main__':
@ -302,51 +318,46 @@ if __name__ == '__main__':
parser.add_argument('--resume', action='store_true', help='resume training flag') 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('--transfer', action='store_true', help='transfer learning flag')
parser.add_argument('--num-workers', type=int, default=4, help='number of Pytorch DataLoader workers') parser.add_argument('--num-workers', type=int, default=4, 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='only save final checkpoint') parser.add_argument('--nosave', action='store_true', help='only save final checkpoint')
parser.add_argument('--notest', action='store_true', help='only test final epoch') 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('--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('--evolve', action='store_true', help='evolve hyperparameters')
parser.add_argument('--cloud-evolve', action='store_true', help='evolve hyperparameters from a cloud source')
parser.add_argument('--var', default=0, type=int, help='debug variable') parser.add_argument('--var', default=0, type=int, help='debug variable')
opt = parser.parse_args() opt = parser.parse_args()
print(opt) print(opt)
opt.evolve = opt.cloud_evolve or opt.evolve
if opt.evolve: if opt.evolve:
opt.notest = True # only test final epoch opt.notest = True # only test final epoch
opt.nosave = True # only save final checkpoint opt.nosave = True # only save final checkpoint
# Train # Train
results = train( results = train(opt.cfg,
opt.cfg, opt.data_cfg,
opt.data_cfg, img_size=opt.img_size,
img_size=opt.img_size, epochs=opt.epochs,
resume=opt.resume or opt.transfer, batch_size=opt.batch_size,
transfer=opt.transfer, accumulate=opt.accumulate)
epochs=opt.epochs,
batch_size=opt.batch_size,
accumulate=opt.accumulate,
)
# Evolve hyperparameters (optional) # Evolve hyperparameters (optional)
if opt.evolve: if opt.evolve:
best_fitness = results[2] # use mAP for fitness
# Write mutation results
print_mutation(hyp, results)
gen = 1000 # generations to evolve gen = 1000 # generations to evolve
for _ in range(gen): print_mutation(hyp, results) # Write mutation results
# Mutate hyperparameters for _ in range(gen):
old_hyp = hyp.copy() # Get best hyperparamters
init_seeds(seed=int(time.time())) x = np.loadtxt('evolve.txt', ndmin=2)
s = [.4, .4, .4, .4, .4, .4, .4, .4, .4 * 0, .4 * 0, .04 * 0, .4 * 0] # fractional sigmas x = x[x[:, 2].argmax()] # select best mAP as genetic fitness (col 2)
for i, k in enumerate(hyp.keys()): for i, k in enumerate(hyp.keys()):
x = (np.random.randn(1) * s[i] + 1) ** 1.1 # plt.hist(x.ravel(), 100) hyp[k] = x[i + 5]
hyp[k] = hyp[k] * float(x) # vary by about 30% 1sigma
# Mutate
init_seeds(seed=int(time.time()))
s = [.2, .2, .2, .2, .2, .2, .2, .2, .2 * 0, .2 * 0, .05 * 0, .2 * 0] # fractional sigmas
for i, k in enumerate(hyp.keys()):
x = (np.random.randn(1) * s[i] + 1) ** 2.0 # plt.hist(x.ravel(), 300)
hyp[k] *= float(x) # vary by 20% 1sigma
# Clip to limits # Clip to limits
keys = ['lr0', 'iou_t', 'momentum', 'weight_decay'] keys = ['lr0', 'iou_t', 'momentum', 'weight_decay']
@ -354,30 +365,17 @@ if __name__ == '__main__':
for k, v in zip(keys, limits): for k, v in zip(keys, limits):
hyp[k] = np.clip(hyp[k], v[0], v[1]) hyp[k] = np.clip(hyp[k], v[0], v[1])
# Determine mutation fitness # Train mutation
results = train( results = train(opt.cfg,
opt.cfg, opt.data_cfg,
opt.data_cfg, img_size=opt.img_size,
img_size=opt.img_size, epochs=opt.epochs,
resume=opt.resume or opt.transfer, batch_size=opt.batch_size,
transfer=opt.transfer, accumulate=opt.accumulate)
epochs=opt.epochs,
batch_size=opt.batch_size,
accumulate=opt.accumulate,
)
mutation_fitness = results[2]
# Write mutation results # Write mutation results
print_mutation(hyp, 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 # # Plot results
# import numpy as np # import numpy as np
# import matplotlib.pyplot as plt # import matplotlib.pyplot as plt

32
utils/google_utils.py Normal file
View File

@ -0,0 +1,32 @@
# This file contains google utils: https://cloud.google.com/storage/docs/reference/libraries
# pip install --upgrade google-cloud-storage
from google.cloud import storage
def upload_blob(bucket_name, source_file_name, destination_blob_name):
# Uploads a file to a bucket
# https://cloud.google.com/storage/docs/uploading-objects#storage-upload-object-python
storage_client = storage.Client()
bucket = storage_client.get_bucket(bucket_name)
blob = bucket.blob(destination_blob_name)
blob.upload_from_filename(source_file_name)
print('File {} uploaded to {}.'.format(
source_file_name,
destination_blob_name))
def download_blob(bucket_name, source_blob_name, destination_file_name):
# Uploads a blob from a bucket
storage_client = storage.Client()
bucket = storage_client.get_bucket(bucket_name)
blob = bucket.blob(source_blob_name)
blob.download_to_filename(destination_file_name)
print('Blob {} downloaded to {}.'.format(
source_blob_name,
destination_file_name))

13
utils/utils.py
View File

@ -11,6 +11,7 @@ from PIL import Image
from tqdm import tqdm from tqdm import tqdm
from . import torch_utils from . import torch_utils
from . import google_utils
matplotlib.rc('font', **{'size': 11}) matplotlib.rc('font', **{'size': 11})
@ -284,7 +285,7 @@ def compute_loss(p, targets, model, giou_loss=False): # predictions, targets, m
# Compute losses # Compute losses
bs = p[0].shape[0] # batch size bs = p[0].shape[0] # batch size
k = bs # loss gain k = bs / 64 # loss gain
for i, pi0 in enumerate(p): # layer i predictions, i for i, pi0 in enumerate(p): # layer i predictions, i
b, a, gj, gi = indices[i] # image, anchor, gridy, gridx b, a, gj, gi = indices[i] # image, anchor, gridy, gridx
tconf = torch.zeros_like(pi0[..., 0]) # conf tconf = torch.zeros_like(pi0[..., 0]) # conf
@ -303,12 +304,12 @@ def compute_loss(p, targets, model, giou_loss=False): # predictions, targets, m
lxy += (k * h['xy']) * MSE(torch.sigmoid(pi[..., 0:2]), txy[i]) # xy loss lxy += (k * h['xy']) * MSE(torch.sigmoid(pi[..., 0:2]), txy[i]) # xy loss
lwh += (k * h['wh']) * MSE(pi[..., 2:4], twh[i]) # wh yolo loss lwh += (k * h['wh']) * MSE(pi[..., 2:4], twh[i]) # wh yolo loss
# tclsm = torch.zeros_like(pi[..., 5:]) tclsm = torch.zeros_like(pi[..., 5:])
# tclsm[range(len(b)), tcls[i]] = 1.0 tclsm[range(len(b)), tcls[i]] = 1.0
# lcls += (k * h['cls']) * BCEcls(pi[..., 5:], tclsm) # class_conf loss lcls += (k * h['cls']) * BCEcls(pi[..., 5:], tclsm) # cls loss (BCE)
lcls += (k * h['cls']) * CE(pi[..., 5:], tcls[i]) # class_conf loss # lcls += (k * h['cls']) * CE(pi[..., 5:], tcls[i]) # cls loss (CE)
# # Append to text file # Append targets to text file
# with open('targets.txt', 'a') as file: # with open('targets.txt', 'a') as file:
# [file.write('%11.5g ' * 4 % tuple(x) + '\n') for x in torch.cat((txy[i], twh[i]), 1)] # [file.write('%11.5g ' * 4 % tuple(x) + '\n') for x in torch.cat((txy[i], twh[i]), 1)]