layer.set_weights(parent_layer_weights)
except:
# PARTIAL WEIGHT TRANSFER
partial_weight_transfer(layer, parent_layer_weights,
disp)
else:
if disp:
print('Did not transfer weights to {}'.format(
layer.name))
return child
if __name__ == '__main__':
from utils import get_flops
from dataset.coco import cn as cfg
cfg.merge_from_file('../configs/evopose_s4_656_XS.yaml')
model = EvoPose(cfg)
print('{:.2f}M / {:.2f}G'.format(model.count_params() / 1e6,
get_flops(model) / 1e9 / 2))
# search space size:
# default_genotype = np.array(genotype_from_blocks_args(DEFAULT_BLOCKS_ARGS))
# s = 2**default_genotype.shape[0] * 4 ** default_genotype.shape[0] # kernel and repeats
# for c in default_genotype[:, 2]: # channels
# s *= c
# s *= 2**3 # stride
# print('Search space size: 10^{:.0f}'.format(np.log10(s)))
# np.random.seed(0)
# cfg.MODEL.LOAD_WEIGHTS = False
# parent_genotype = genotype_from_blocks_args(DEFAULT_BLOCKS_ARGS)
parser.add_argument('-c',
'--cfg',
default='./configs/evopose2d_M_f32.yaml')
parser.add_argument(
'-p',
'--coco-path',
required=True,
default='./data/',
help='Path to folder containing COCO images and annotation directories.'
)
parser.add_argument('-i', '--img-id', type=int, default=785)
parser.add_argument('--alpha', type=float, default=0.8)
args = parser.parse_args()
# load the config .yaml file
cfg.merge_from_file('configs/' + args.cfg)
# load the trained model
model = tf.keras.models.load_model('models/{}.h5'.format(
args.cfg.split('.yaml')[0]))
cfg.DATASET.OUTPUT_SHAPE = model.output_shape[1:]
# load the dataset annotations
coco = COCO(
osp.join(args.coco_path, 'annotations',
'person_keypoints_val2017.json'))
img_data = coco.loadImgs([args.img_id])[0]
annotation = coco.loadAnns(coco.getAnnIds([args.img_id]))[0]
bbox = annotation['bbox']
kp = np.array(annotation['keypoints']).reshape(-1, 3) # not used
cfg.MODEL.HEAD_KERNEL,
strides=2,
padding='same',
use_bias=False,
kernel_regularizer=regularizer,
name='head_conv{}'.format(i + 1))(x)
x = layers.BatchNormalization(name='head_bn{}'.format(i + 1))(x)
x = layers.Activation(cfg.MODEL.HEAD_ACTIVATION,
name='head_act{}'.format(i + 1))(x)
x = layers.Conv2D(cfg.DATASET.OUTPUT_SHAPE[-1],
1,
padding='same',
use_bias=True,
kernel_regularizer=regularizer,
name='final_conv')(x)
return Model(backbone.input, x, name='sb_{}'.format(cfg.MODEL.BACKBONE))
if __name__ == '__main__':
from dataset.coco import cn as cfg
cfg.merge_from_file('../configs/sb_resnet50_256x192.yaml')
cfg.DATASET.INPUT_SHAPE = [384, 288, 3]
cfg.MODEL.BACKBONE = 'resnet152'
model = SimpleBaseline(cfg)
model.summary()
print('{:.2f}M parameters | {:.2f}G multiply-adds'.format(
model.count_params() / 1e6,
get_flops(model) / 1e9 / 2))
# stem
for i in range(2):
x = Conv2D(64,
3,
2,
'same',
use_bias=False,
name='conv{}'.format(i + 1))(x)
x = BatchNormalization(name='bn{}'.format(i + 1))(x)
x = ReLU(name='relu{}'.format(i + 1))(x)
xs = [x]
stage_cfgs = [cfg.MODEL['STAGE{}'.format(i + 1)] for i in range(4)]
for i, stage_cfg in enumerate(stage_cfgs):
xs = stage(xs, stage_cfg, name='s{}'.format(i + 1))
output = Conv2D(cfg.DATASET.OUTPUT_SHAPE[-1], 1, name='final_conv')(xs[0])
model = Model(inputs=input, outputs=output, name='hrnet')
add_regularization(model, l2(cfg.TRAIN.WD))
return model
if __name__ == '__main__':
from utils import get_flops
from dataset.coco import cn as cfg
cfg.merge_from_file('../configs/hrnet_w32_256x192.yaml')
# cfg.DATASET.INPUT_SHAPE = [256, 192, 3]
model = HRNet(cfg)
# model.summary()
print('{:.2f}M / {:.2f}G'.format(model.count_params() / 1e6,
get_flops(model) / 1e9 / 2))
parser = argparse.ArgumentParser()
parser.add_argument('-c', '--cfg', default='E3.yaml')
parser.add_argument('--accelerator-prefix', default='node-')
parser.add_argument('-a',
'--accelerator-ids',
type=int,
nargs='+',
default=[0])
parser.add_argument('-ar',
'--accelerator-range',
type=int,
nargs='+',
default=None)
args = parser.parse_args()
cfg.merge_from_file(osp.join('configs/', args.cfg))
if args.accelerator_range:
accelerators = [
'{}{}'.format(args.accelerator_prefix, a)
for a in range(*args.accelerator_range)
]
else:
accelerators = [
'{}{}'.format(args.accelerator_prefix, a)
for a in args.accelerator_ids
]
assert cfg.SEARCH.CHILDREN % cfg.SEARCH.PARENTS == 0, \
'Number of children must be divisible by number of parents.'
assert cfg.SEARCH.CHILDREN % len(accelerators) == 0, \
