def build_recognizer(cfg, device):
world_size = du.get_world_size()
model = registry.RECOGNIZER[cfg.MODEL.RECOGNIZER.TYPE](cfg)
if cfg.MODEL.NORM.SYNC_BN and world_size > 1:
logger.info("start sync BN on the process group of {}".format(
du.LOCAL_RANK_GROUP))
convert_sync_bn(model, du.LOCAL_PROCESS_GROUP)
preloaded = cfg.MODEL.RECOGNIZER.PRELOADED
if preloaded != "":
logger.info(f'load pretrained: {preloaded}')
check_pointer = CheckPointer(model)
check_pointer.load(preloaded, map_location=device)
logger.info("finish loading model weights")
if cfg.MODEL.CONV.ADD_BLOCKS is not None:
assert isinstance(cfg.MODEL.CONV.ADD_BLOCKS, tuple)
for add_block in cfg.MODEL.CONV.ADD_BLOCKS:
if add_block == 'RepVGGBlock':
insert_repvgg_block(model)
if add_block == 'ACBlock':
insert_acblock(model)
model = model.to(device=device)
if du.get_world_size() > 1:
model = DDP(model,
device_ids=[device],
output_device=device,
find_unused_parameters=True)
return model
def test_regvgg_recognizer():
data = torch.randn(1, 3, 224, 224)
for key in arch_settings.keys():
print('*' * 10, key)
model = RepVGGRecognizer(arch=key)
# print(model)
outputs = model(data)[KEY_OUTPUT]
assert outputs.shape == (1, 1000)
print('insert_regvgg_block -> fuse_regvgg_block')
insert_repvgg_block(model)
# print(model)
model.eval()
outputs_insert = model(data)[KEY_OUTPUT]
fuse_repvgg_block(model)
# print(model)
model.eval()
outputs_fuse = model(data)[KEY_OUTPUT]
# print(outputs_insert)
# print(outputs_fuse)
print(torch.sqrt(torch.sum((outputs_insert - outputs_fuse)**2)))
print(torch.allclose(outputs_insert, outputs_fuse, atol=1e-8))
assert torch.allclose(outputs_insert, outputs_fuse, atol=1e-8)
print(
'insert_regvgg_block -> insert_acblock -> fuse_acblock -> fuse_regvgg_block'
)
insert_repvgg_block(model)
insert_acblock(model)
# print(model)
model.eval()
outputs_insert = model(data)[KEY_OUTPUT]
fuse_acblock(model)
fuse_repvgg_block(model)
# print(model)
model.eval()
outputs_fuse = model(data)[KEY_OUTPUT]
print(torch.sqrt(torch.sum((outputs_insert - outputs_fuse)**2)))
print(torch.allclose(outputs_insert, outputs_fuse, atol=1e-7))
assert torch.allclose(outputs_insert, outputs_fuse, atol=1e-7)
print(
'insert_acblock -> insert_regvgg_block -> fuse_regvgg_block -> fuse_acblock'
)
insert_repvgg_block(model)
insert_acblock(model)
# print(model)
model.eval()
outputs_insert = model(data)[KEY_OUTPUT]
fuse_acblock(model)
fuse_repvgg_block(model)
# print(model)
model.eval()
outputs_fuse = model(data)[KEY_OUTPUT]
print(torch.sqrt(torch.sum((outputs_insert - outputs_fuse)**2)))
print(torch.allclose(outputs_insert, outputs_fuse, atol=1e-7))
assert torch.allclose(outputs_insert, outputs_fuse, atol=1e-7)
def test_resnet50_acb():
model = resnet50()
model.eval()
# print(model)
data = torch.randn(1, 3, 224, 224)
insert_acblock(model)
model.eval()
train_outputs = model(data)
# print(model)
fuse_acblock(model, eps=1e-5)
model.eval()
eval_outputs = model(data)
# print(model)
print(torch.sqrt(torch.sum((train_outputs - eval_outputs)**2)))
print(torch.allclose(train_outputs, eval_outputs, atol=1e-7))
assert torch.allclose(train_outputs, eval_outputs, atol=1e-7)
def test_acb_helper():
in_channels = 32
out_channels = 64
dilation = 1
# 下采样 + 分组卷积
kernel_size = 3
stride = 2
padding = 1
groups = 8
model = nn.Sequential(
nn.Sequential(
nn.Conv2d(in_channels,
out_channels,
kernel_size,
stride=stride,
padding=padding,
dilation=dilation,
groups=groups), nn.BatchNorm2d(out_channels),
nn.ReLU(inplace=True)), nn.BatchNorm2d(out_channels),
nn.ReLU(inplace=True))
model.eval()
print(model)
data = torch.randn(1, in_channels, 56, 56)
insert_acblock(model)
model.eval()
train_outputs = model(data)
print(model)
fuse_acblock(model, eps=1e-5)
model.eval()
eval_outputs = model(data)
print(model)
print(torch.sqrt(torch.sum((train_outputs - eval_outputs)**2)))
print(torch.allclose(train_outputs, eval_outputs, atol=1e-6))
assert torch.allclose(train_outputs, eval_outputs, atol=1e-6)
def test_regvgg():
data = torch.randn(1, 3, 224, 224)
for key in arch_settings.keys():
print('*' * 10, key)
cfg.merge_from_file(
'configs/benchmarks/repvgg/repvgg_b2g4_cifar100_224_e100_sgd_calr.yaml'
)
model = RepVGG(cfg)
# print(model)
outputs = model(data)[KEY_OUTPUT]
assert outputs.shape == (1, 100)
print('insert_regvgg_block -> fuse_regvgg_block')
insert_repvgg_block(model)
# print(model)
model.eval()
outputs_insert = model(data)[KEY_OUTPUT]
fuse_repvgg_block(model)
# print(model)
model.eval()
outputs_fuse = model(data)[KEY_OUTPUT]
# print(outputs_insert)
# print(outputs_fuse)
print(torch.sqrt(torch.sum((outputs_insert - outputs_fuse)**2)))
print(torch.allclose(outputs_insert, outputs_fuse, atol=1e-8))
assert torch.allclose(outputs_insert, outputs_fuse, atol=1e-8)
print(
'insert_regvgg_block -> insert_acblock -> fuse_acblock -> fuse_regvgg_block'
)
insert_repvgg_block(model)
insert_acblock(model)
# print(model)
model.eval()
outputs_insert = model(data)[KEY_OUTPUT]
fuse_acblock(model)
fuse_repvgg_block(model)
# print(model)
model.eval()
outputs_fuse = model(data)[KEY_OUTPUT]
print(torch.sqrt(torch.sum((outputs_insert - outputs_fuse)**2)))
print(torch.allclose(outputs_insert, outputs_fuse, atol=1e-6))
assert torch.allclose(outputs_insert, outputs_fuse, atol=1e-6)
print(
'insert_acblock -> insert_regvgg_block -> fuse_regvgg_block -> fuse_acblock'
)
insert_repvgg_block(model)
insert_acblock(model)
# print(model)
model.eval()
outputs_insert = model(data)[KEY_OUTPUT]
fuse_acblock(model)
fuse_repvgg_block(model)
# print(model)
model.eval()
outputs_fuse = model(data)[KEY_OUTPUT]
print(torch.sqrt(torch.sum((outputs_insert - outputs_fuse)**2)))
print(torch.allclose(outputs_insert, outputs_fuse, atol=1e-6))
assert torch.allclose(outputs_insert, outputs_fuse, atol=1e-6)