def FISHNets(arch='fish99', cfg='', model_path='', **kwargs):
"""
自定义接口 for model_factory
指定arch=官方模型，即可从map中调用官方模型的名称name，配置cfg，以及预训练参数ckpt
指定arch=自定义模型，即可使用传入的模型名称name，配置cfg，以及预训练参数ckpt
"""
model_name_map = {
'fish99': 'fishnet99',
'fish150': 'fishnet150',
'fish201': 'fishnet201',
}
model_cfg_map = {
'fish99': fish99_cfg,
'fish150': fish150_cfg,
'fish201': fish201_cfg,
}
model_ckpt_map = {
'fish99': 'fishnet99_ckpt.tar',
'fish150': 'fishnet150_ckpt.tar',
'fish201': 'fishnet20x_ckpt_welltrain-==fishnet201.tar',
}
try:
# 调用官方模型
name = model_name_map[arch]
except:
# 使用自定义模型，如fish33, fish55
name = arch
if cfg == '':
# 调用官方配置
cfg = model_cfg_map[arch]
else:
# 使用自定义配置
assert isinstance(cfg, dict)
cfg = cfg
cfg = {**cfg, **kwargs}
model = FishNet(Bottleneck, **cfg)
model_dir = xtils.get_pretrained_models()
if os.path.isfile(model_path):
model = xtils.load_ckpt_weights(model, model_path, device='cpu', mgpus_to_sxpu='auto')
elif model_path == 'local':
model_path = os.path.join(model_dir, model_ckpt_map[arch])
model = xtils.load_ckpt_weights(model, model_path, device='cpu', mgpus_to_sxpu='auto')
elif model_path == 'download':
# model_url_map = {}
# import torch.utils.model_zoo as model_zoo
# model.load_state_dict(model_zoo.load_url(model_url_map['arch'], model_dir))
raise NotImplementedError
else:
assert model_path == '', '<model_path> must refer to "local" or "download" or "" or "model.ckpt".'
return model
class Config(object):
gpu_ids = [0, 1, 2, 3]
device = torch.device('cuda:{}'.format(gpu_ids[0]))
dataset = 'imagenet'
data_root = xtils.get_data_root(data='imagenet')
data_augment = {'train': 'rotate-rresize-1crop', 'val': '1resize-1crop',
'imsize': 256, 'insize': 224, 'color': True,
'degree': (0, 0), 'scale': (0.08, 1), 'ratio': (3. / 4, 4. / 3)}
arch_name = ['hrnet', 'resnet50', 'resnet152', 'fishnet150', 'scalenet'][-1]
arch_kwargs = [hrw18, hrw30, hrw32, vo69, vo72][-2]
ckpt_file = os.path.join(xtils.get_pretrained_models(),
['hrnetv2_w18_imagenet_pretrained.pth',
'fishnet150_ckpt.tar',
'resnet50-19c8e357.pth', 'resnet152-b121ed2d.pth',
'scale103-5.09M-71.63-vo69.pth',
'scale59-30.52M-74.86-vo72.pth'][-2])
mgpus_to_sxpu = ['m2s', 's2m', 'none'][-1]
image_size = 256
input_size = int(0.875 * image_size)
bsize_val = 256
num_workers = 8
print_freq = 20
valid_total_time = 0
random_seed = random.randint(0, 100) or None
# 特殊模型的补充字段
mode_custom = False
xfc_which = -1
def vgg19_bn(pretrained=False, **kwargs):
"""VGG 19-layer model (configuration 'E') with batch normalization
Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
"""
if pretrained:
kwargs['init_weights'] = False
model = VGG(make_layers(cfg['E'], batch_norm=True), **kwargs)
if pretrained:
model.load_state_dict(model_zoo.load_url(model_urls['vgg19_bn']))
return model
# 统一接口 interface for model_factory ##########################################
model_dir = xtils.get_pretrained_models()
def VGGNets(arch, model_path=None):
pretrained = True if model_path else False
if arch == 'vgg19':
return vgg19(pretrained, model_path)
elif arch == 'vgg16':
return vgg16(pretrained, model_path)
elif arch == 'vgg19_bn':
return vgg19_bn(pretrained)
else:
raise NotImplementedError
batch_size_val = 4
batch_nums = math.ceil(train_size / batch_size)
BN = batch_nums # =>> Unit #5005
cfgmsnet = {
# experiment config
'exp_version': 'exp.xxx',
'train_val_test': (False, True, False),
# device config
'gpu_ids': [0, 1, 2, 3, 4, 5, 6, 7, 8][0:4],
# model config
'arch_name': 'msnet',
'arch_kwargs': {},
'resume': os.path.join(xtils.get_pretrained_models(), ''),
'resume_config': False,
'resume_optimizer': False,
'mgpus_to_sxpu': ['m2s', 's2m', 'none', 'auto'][3],
# data config
'dataset': 'imagenet',
'data_info': {'train_size': train_size, 'val_size': 50000, 'test_size': 50000},
'data_root': xtils.get_data_root(data='imagenet'),
'data_augment': {'train': 'rotate-rresize-1crop', 'val': '1resize-1crop',
'imsize': 256, 'insize': 224, 'color': True,
'degree': (0, 0), 'scale': (0.08, 1), 'ratio': (3. / 4, 4. / 3)},
'data_kwargs': {},
'data_workers': 12,
# path config
def RESNets(arch, cfg, model_path=''):
"""
自定义接口 for model_factory
:param arch: res18, res34, res50, res101, res152
:param cfg: arch configs
:param model_path: state_dict.pth
:return: a blank model or pre-trained model
"""
_block = {'BasicBlock': BasicBlock, 'Bottleneck': Bottleneck}
model_name_map = {
'res18': 'resnet18',
'res34': 'resnet34',
'res50': 'resnet50',
'res101': 'resnet101',
'res152': 'resnet152',
}
model_cfg_map = {
'resnet18': res18['cfg'],
'resnet34': res34['cfg'],
'resnet50': res50['cfg'],
'resnet101': res101['cfg'],
'resnet152': res152['cfg'],
}
model_ckpt_map = {
'resnet18': 'resnet18-5c106cde.pth',
'resnet34': 'resnet34-333f7ec4.pth',
'resnet50': 'resnet50-19c8e357.pth',
'resnet101': 'resnet101-5d3b4d8f.pth',
'resnet152': 'resnet152-b121ed2d.pth',
}
try:
# 调用官方模型
name = model_name_map[arch]
except:
# 使用自定义模型，如fish33, fish55
name = arch
if cfg == '':
# 调用官方配置
cfg = model_cfg_map[name]
else:
# 使用自定义配置
assert isinstance(cfg, dict)
cfg = cfg
if cfg['block'] in _block.keys():
cfg['block'] = _block[cfg['block']]
model = ResNet(**cfg)
model_dir = xtils.get_pretrained_models()
if os.path.isfile(model_path):
print('\n=> loading model.pth from %s.' % model_path)
model.load_state_dict(torch.load(model_path))
print('\nSuccess: loaded model.pth from %s.\n' % model_path)
elif model_path == 'local':
model_path = os.path.join(model_dir, model_ckpt_map[name])
print('\n=> loading model.pth from %s.' % model_path)
model.load_state_dict(torch.load(model_path))
print('\nSuccess: loaded model.pth from %s.\n' % model_path)
elif model_path == 'download':
print('\n=> downloading model.pth from %s.' % model_urls[name])
model.load_state_dict(model_zoo.load_url(model_urls[name], model_dir))
print('\nSuccess: downloaded model.pth from %s.\n' % model_urls[name])
else:
assert model_path == '', '<model_path> must refer to valid-model.ckpt || ' 'download' ' || "".'
return model
def MobileV3(arch='mbvdl', cfg='', model_path=''):
"""
自定义接口 for model_factory
指定arch=官方模型，即可从map中调用官方模型的名称name，配置cfg，以及预训练参数ckpt
指定arch=自定义模型，即可使用传入的模型名称name，配置cfg，以及预训练参数ckpt
"""
model_name_map = {
'mbvdl': 'MobileV3D-large',
'mbvds': 'MobileV3D-small',
'mbvxl': 'MobileV3X-large',
'mbvxs': 'MobileV3X-small',
'mbvyl': 'MobileV3Y-large',
'mbvys': 'MobileV3Y-small',
}
model_cfg_map = {
'mbvdl': 'large',
'mbvds': 'small',
'mbvxl': 'large',
'mbvxs': 'small',
'mbvyl': 'large',
'mbvys': 'small',
}
model_ckpt_map = {
'mbvdl': 'mobilev3_d_large.pth.tar',
'mbvds': 'mobilev3_d_small.pth.tar',
'mbvxl': 'mobilev3_x_large.pth.tar',
'mbvxs': 'mobilev3_x_small.pth.tar',
'mbvyl': 'mobilev3_y_large.pth.tar',
'mbvys': 'mobilev3_y_small.pth.tar',
}
try:
# 调用官方模型
name = model_name_map[arch]
except:
# 使用自定义模型，如mbvd33, mbvd22
name = arch
if cfg == '':
# 调用官方配置
cfg = model_cfg_map[arch]
else:
# 使用自定义配置
assert isinstance(cfg, dict)
raise NotImplementedError
if name.startswith('MobileV3D'):
model = MobileV3D(cfg)
elif name.startswith('MobileV3X'):
model = MobileV3X(cfg)
elif name.startswith('MobileV3Y'):
model = MobileV3Y(cfg)
elif name.startswith('MobileV3Z'):
model = MobileV3Z(cfg)
else:
raise NotImplementedError
model_dir = xtils.get_pretrained_models()
if os.path.isfile(model_path):
model = xtils.load_ckpt_weights(model,
model_path,
device='cpu',
mgpus_to_sxpu='auto')
elif model_path == 'local':
model_path = os.path.join(model_dir, model_ckpt_map[arch])
model = xtils.load_ckpt_weights(model,
model_path,
device='cpu',
mgpus_to_sxpu='auto')
elif model_path == 'download':
# model_url_map = {}
# import torch.utils.model_zoo as model_zoo
# model.load_state_dict(model_zoo.load_url(model_url_map['arch'], model_dir))
raise NotImplementedError
else:
assert model_path == '', '<model_path> must refer to "local" or "download" or "" or "model.ckpt".'
return model
batch_size_val = 4
batch_nums = math.ceil(train_size / batch_size)
BN = batch_nums # =>> Unit #5005
cfgscale = {
# experiment config
'exp_version': 'exp.xxx',
'train_val_test': (False, True, False),
# device config
'gpu_ids': [0, 1, 2, 3, 4, 5, 6, 7, 8][0:4],
# model config
'arch_name': 'scalenet',
'arch_kwargs': {},
'resume': os.path.join(xtils.get_pretrained_models(), 'scale103-5.09M-71.63-vo69.pth'),
'resume_config': False,
'resume_optimizer': False,
'mgpus_to_sxpu': ['m2s', 's2m', 'none', 'auto'][3],
# data config
'dataset': 'imagenet',
'data_info': {'train_size': train_size, 'val_size': 50000, 'test_size': 50000},
'data_root': xtils.get_data_root(data='imagenet'),
'data_augment': {'train': 'rotate-rresize-1crop', 'val': '1resize-1crop',
'imsize': 256, 'insize': 224, 'color': True,
'degree': (0, 0), 'scale': (0.08, 1), 'ratio': (3. / 4, 4. / 3)},
'data_kwargs': {},
'data_workers': 4,
# path config
def EFFNets(arch='effb0', cfg='', model_path='', override_params=None):
"""
自定义接口 for model_factory
指定arch=官方模型，即可从map中调用官方模型的名称name，配置cfg，以及预训练参数ckpt
指定arch=自定义模型，即可使用传入的模型名称name，配置cfg，以及预训练参数ckpt
"""
model_name_map = {
'effb0': 'efficientnet-b0',
'effb1': 'efficientnet-b1',
'effb2': 'efficientnet-b2',
'effb3': 'efficientnet-b3',
'effb4': 'efficientnet-b4',
'effb5': 'efficientnet-b5',
'effb6': 'efficientnet-b6',
'effb7': 'efficientnet-b7',
}
model_cfg_map = {"""no-use"""}
model_ckpt_map = {
'effb0': 'efficientnet-b0-08094119.pth',
'effb1': 'efficientnet-b1-dbc7070a.pth',
'effb2': 'efficientnet-b2-27687264.pth',
'effb3': 'efficientnet-b3-c8376fa2.pth',
'effb4': '',
'effb5': '',
'effb6': '',
'effb7': '',
}
try:
# 调用官方模型
model_name = model_name_map[arch]
except:
# 使用自定义模型，如effb11, effb888
model_name = arch
if cfg == '':
# 调用官方配置
# cfg = model_cfg_map[arch]
pass
else:
# 使用自定义配置
assert isinstance(cfg, dict)
# cfg = cfg
blocks_args, global_params = get_model_params(model_name, override_params)
model = EfficientNet(blocks_args, global_params, model_name)
model_dir = xtils.get_pretrained_models()
if os.path.isfile(model_path):
model = xtils.load_ckpt_weights(model,
model_path,
device='cpu',
mgpus_to_sxpu='auto')
elif model_path == 'local':
model_path = os.path.join(model_dir, model_ckpt_map[arch])
model = xtils.load_ckpt_weights(model,
model_path,
device='cpu',
mgpus_to_sxpu='auto')
elif model_path == 'download':
model_url_map = {
'efficientnet-b0':
'http://storage.googleapis.com/public-models/efficientnet-b0-08094119.pth',
'efficientnet-b1':
'http://storage.googleapis.com/public-models/efficientnet-b1-dbc7070a.pth',
'efficientnet-b2':
'http://storage.googleapis.com/public-models/efficientnet-b2-27687264.pth',
'efficientnet-b3':
'http://storage.googleapis.com/public-models/efficientnet-b3-c8376fa2.pth',
}
import torch.utils.model_zoo as model_zoo
model.load_state_dict(
model_zoo.load_url(model_url_map[model_name], model_dir))
else:
assert model_path == '', '<model_path> must refer to "local" or "download" or "" or "model.ckpt".'
return model