def main(data_shape, config_file, mobile_name):
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
torch.backends.cudnn.deterministic = False
torch.backends.cudnn.benchmark = True
cfg.merge_from_file(config_file)
gpu_device = torch.device('cuda:0')
cpu_device = torch.device('cpu')
model = build_recognizer(cfg, cpu_device)
model.eval()
data = torch.randn(data_shape).to(device=cpu_device, non_blocking=True)
GFlops, params_size = compute_num_flops(model, data)
print(f'{mobile_name} ' + '*' * 10)
print(f'device: {cpu_device}')
print(f'GFlops: {GFlops:.3f}G')
print(f'Params Size: {params_size:.3f}MB')
model = build_recognizer(cfg, cpu_device)
model.eval()
print(f'compute cpu infer time')
compute_model_time(data_shape, model, cpu_device)
print(f'compute gpu infer time')
compute_model_time(data_shape, model, gpu_device)
del model
torch.cuda.empty_cache()
def rotnet(pretrained=False, **kwargs):
cfg.merge_from_file(config_file)
model = build_recognizer(cfg, torch.device('cpu'))
if pretrained:
checkpoint = torch.load('weights/model.pth')
model.load_state_dict(checkpoint)
return model
def test_mnasnet():
cfg.merge_from_file(
'configs/benchmarks/lightweight/mnasnet_a1_1_3_cifar100_224_e100.yaml')
data = torch.randn(1, 3, 224, 224)
for wm in [0.5, 0.75, 1.0, 1.3]:
cfg.MODEL.COMPRESSION.WIDTH_MULTIPLIER = wm
model = MNASNet(cfg)
print(model)
outputs = model(data)[KEY_OUTPUT]
print(outputs.shape)
assert outputs.shape == (1, 100)
cfg.merge_from_file(
'configs/benchmarks/lightweight/mnasnet_b1_1_3_torchvision_cifar100_224_e100_sgd.yaml'
)
for wm in [0.5, 0.75, 1.0, 1.3]:
cfg.MODEL.COMPRESSION.WIDTH_MULTIPLIER = wm
model = build_torchvision_mnasnet(cfg)
print(model)
outputs = model(data)[KEY_OUTPUT]
print(outputs.shape)
assert outputs.shape == (1, 100)
def test_sknet():
config_file = 'configs/benchmarks/resnet/sknet50_cifar100_224_e100_rmsprop.yaml'
cfg.merge_from_file(config_file)
model = ResNet(cfg)
print(model)
test_data(model, (3, 3, 224, 224), (3, 100))
def load_config(args):
if args.config_file:
cfg.merge_from_file(args.config_file)
if args.pretrained:
cfg.MODEL.RECOGNIZER.PRELOADED = args.pretrained
if args.output_dir:
cfg.OUTPUT_DIR = args.output_dir
if args.log_step != -1:
cfg.TRAIN.LOG_STEP = args.log_step
if args.save_step != -1:
cfg.TRAIN.SAVE_STEP = args.save_step
if args.eval_step != -1:
cfg.TRAIN.EVAL_STEP = args.eval_step
if args.resume:
cfg.TRAIN.RESUME = True
if not args.use_tensorboard:
cfg.TRAIN.USE_TENSORBOARD = False
if args.gpus != -1:
cfg.NUM_GPUS = args.gpus
if args.nodes != -1:
cfg.NUM_NODES = args.nodes
if args.nr != -1:
cfg.RANK_ID = args.nr
if args.init_method:
cfg.INIT_METHOD = args.init_method
cfg.merge_from_list(args.opts)
cfg.freeze()
return cfg
def load_train_config(args):
if args.config_file:
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
if args.log_step != -1:
cfg.TRAIN.LOG_STEP = args.log_step
if args.save_step != -1:
cfg.TRAIN.SAVE_STEP = args.save_step
if args.eval_step != -1:
cfg.TRAIN.EVAL_STEP = args.eval_step
if args.resume:
cfg.TRAIN.RESUME = True
if not args.use_tensorboard:
cfg.TRAIN.USE_TENSORBOARD = False
if args.gpus != -1:
cfg.NUM_GPUS = args.gpus
if args.nodes != -1:
cfg.NUM_NODES = args.nodes
if args.nr != -1:
cfg.RANK_ID = args.nr
cfg.freeze()
if not os.path.exists(cfg.OUTPUT_DIR):
os.makedirs(cfg.OUTPUT_DIR)
return cfg
def test_dataloader():
config_file = 'configs/imagenet/rxtd50_32x4d_imagenet_224_e100_sgd_mslr_e100_g2.yaml'
cfg.merge_from_file(config_file)
dataloader = build_dataloader(cfg, is_train=True)
print(dataloader)
te = iter(dataloader)
print(te)
images, targets = te.__next__()
print(images.shape)
print(targets)
def test_config():
config_file = 'configs/benchmarks/rxd50_32x4d_avg_custom_cifar100_224_e100_rmsprop.yaml'
cfg.merge_from_file(config_file)
model = build_resnet(cfg)
print(model)
test_data(model, (1, 3, 224, 224), (1, 100))
config_file = 'configs/benchmarks/rxd50_32x4d_fast_avg_custom_cifar100_224_e100_rmsprop.yaml'
cfg.merge_from_file(config_file)
model = build_resnet(cfg)
print(model)
test_data(model, (1, 3, 224, 224), (1, 100))
def test_resnetd():
config_file = 'configs/benchmarks/resnet/rd50_cifar100_224_e100_rmsprop.yaml'
cfg.merge_from_file(config_file)
model = ResNet(cfg)
print(model)
test_data(model, (1, 3, 224, 224), (1, 100))
config_file = 'configs/benchmarks/resnet/rxtd50_32x4d_cifar100_224_e100_rmsprop.yaml'
cfg.merge_from_file(config_file)
model = ResNet(cfg)
print(model)
test_data(model, (1, 3, 224, 224), (1, 100))
config_file = 'configs/benchmarks/resnet/rxtd50_32x4d_fast_avg_cifar100_224_e100_rmsprop.yaml'
cfg.merge_from_file(config_file)
model = ResNet(cfg)
print(model)
test_data(model, (1, 3, 224, 224), (1, 100))
config_file = 'configs/benchmarks/resnet/rxtd50_32x4d_avg_cifar100_224_e100_rmsprop.yaml'
cfg.merge_from_file(config_file)
model = ResNet(cfg)
print(model)
test_data(model, (1, 3, 224, 224), (1, 100))
def test_resnet():
config_file = 'configs/benchmarks/resnet/r50_cifar100_224_e100_rmsprop.yaml'
cfg.merge_from_file(config_file)
model = ResNet(cfg)
print(model)
test_data(model, (1, 3, 224, 224), (1, 100))
config_file = 'configs/benchmarks/resnet/rxt50_32x4d_cifar100_224_e100_rmsprop.yaml'
cfg.merge_from_file(config_file)
model = ResNet(cfg)
print(model)
test_data(model, (1, 3, 224, 224), (1, 100))
config_file = 'configs/benchmarks/resnet/r50_torchvision_cifar100_224_e100_rmsprop.yaml'
cfg.merge_from_file(config_file)
model = build_torchvision_resnet(cfg)
print(model)
test_data(model, (1, 3, 224, 224), (1, 100))
config_file = 'configs/benchmarks/resnet/rxt50_32x4d_torchvision_cifar100_224_e100_rmsprop.yaml'
cfg.merge_from_file(config_file)
model = build_torchvision_resnet(cfg)
print(model)
test_data(model, (1, 3, 224, 224), (1, 100))
def test_shufflenet_v2():
cfg.merge_from_file(
'configs/benchmarks/lightweight/sfv2_x2_0_cifar100_224_e100.yaml')
print(cfg)
model = ShuffleNetV2(cfg)
print(model)
test_data(model)
cfg.merge_from_file(
'configs/benchmarks/lightweight/sfv2_torchvision_cifar100_224_e100.yaml'
)
model = build_torchvision_sfv2(cfg)
print(model)
test_data(model)
def test_cfg():
cfg_file = 'configs/cifar/mbv3_large_se_hsigmoid_c100_224_e100_rmsprop_mslr_g1.yaml'
cfg.merge_from_file(cfg_file)
print(cfg.TRANSFORM)
res = build_transform(cfg, is_train=True)
data = Image.fromarray(torch.randn(234, 134, 3).numpy().astype(np.uint8))
print(res(data).shape)
cfg_file = 'configs/cifar/rxtd50_32x4d_c100_224_e100_sgd_mslr_g1.yaml'
cfg.merge_from_file(cfg_file)
print(cfg.TRANSFORM)
res = build_transform(cfg, is_train=True)
data = Image.fromarray(torch.randn(234, 134, 3).numpy().astype(np.uint8))
print(res(data).shape)
def test_prefetcher():
config_file = 'configs/cifar/mbv3_large_se_hsigmoid_c10_224_e100_rmsprop_mslr_g1.yaml'
cfg.merge_from_file(config_file)
data_loader = build_dataloader(cfg, is_train=False)
print(data_loader.__len__())
print('enumerate')
model = Prefetcher(data_loader)
print(model)
for i, item in enumerate(model):
print(i)
print('tqdm')
model = Prefetcher(data_loader)
print(model)
for images, targets in tqdm(model):
time.sleep(0.1)
print('done')
def test_mobilenet_v2():
for s in [224, 192, 160, 128]:
for wm in [1.0, 0.75, 0.5, 0.25]:
cfg.merge_from_file('configs/benchmarks/lightweight/mbv2_cifar100_224_e100.yaml')
cfg.MODEL.COMPRESSION.WIDTH_MULTIPLIER = wm
round_nearest = cfg.MODEL.COMPRESSION.ROUND_NEAREST
feature_dims = make_divisible(cfg.MODEL.HEAD.FEATURE_DIMS * wm, round_nearest)
cfg.MODEL.HEAD.FEATURE_DIMS = feature_dims
print(f's: {s}, wn: {wm}, feature_dims: {feature_dims}')
model = MobileNetV2(cfg)
# print(model)
data = torch.randn(1, 3, s, s)
outputs = model(data)[KEY_OUTPUT]
print(outputs.shape)
assert outputs.shape == (1, 100)
def test_regvgg():
cfg.merge_from_file('configs/benchmarks/repvgg/repvgg_b2g4_cifar100_224_e100_sgd_calr.yaml')
model = RepVGG(cfg)
model.eval()
print(model)
data = torch.randn(1, 3, 224, 224)
insert_repvgg_block(model)
model.eval()
train_outputs = model(data)[KEY_OUTPUT]
print(model)
fuse_repvgg_block(model)
model.eval()
eval_outputs = model(data)[KEY_OUTPUT]
print(model)
print(torch.sqrt(torch.sum((train_outputs - eval_outputs) ** 2)))
print(torch.allclose(train_outputs, eval_outputs, atol=1e-8))
assert torch.allclose(train_outputs, eval_outputs, atol=1e-8)
def load_test_config(args):
if not os.path.isfile(args.config_file) or not os.path.isfile(
args.pretrained):
raise ValueError('需要输入配置文件和预训练模型路径')
cfg.merge_from_file(args.config_file)
cfg.MODEL.RECOGNIZER.PRELOADED = args.pretrained
cfg.OUTPUT_DIR = args.output
if args.gpus != -1:
cfg.NUM_GPUS = args.gpus
if args.nodes != -1:
cfg.NODES = args.nodes
if args.nr != -1:
cfg.RANK = args.nr
cfg.freeze()
if not os.path.exists(cfg.OUTPUT_DIR):
os.makedirs(cfg.OUTPUT_DIR)
return cfg
def main():
args = parse_args()
print(args)
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
if args.ckpt is not None:
cfg.MODEL.RECOGNIZER.PRELOADED = args.ckpt
cfg.freeze()
if args.config_file:
print("Loaded configuration file {}".format(args.config_file))
with open(args.config_file, "r") as cf:
config_str = "\n" + cf.read()
print(config_str)
# print("Running with config:\n{}".format(cfg))
run_demo(cfg=cfg,
images_dir=args.images_dir,
rotate_dir=args.rotate_dir,
output_dir=args.output_dir)
def test_config():
data = torch.randn(1, 3, 224, 224)
config_file = 'configs/benchmarks/mnasnet_a1_1_3_se_custom_cifar100_224_e50.yaml'
cfg.merge_from_file(config_file)
model = build_mnasnet(cfg)
print(model)
outputs = model(data)[KEY_OUTPUT]
print(outputs.shape)
assert outputs.shape == (1, 100)
config_file = 'configs/benchmarks/mnasnet_a1_1_3_custom_cifar100_224_e50.yaml'
cfg.merge_from_file(config_file)
model = build_mnasnet(cfg)
print(model)
outputs = model(data)[KEY_OUTPUT]
print(outputs.shape)
assert outputs.shape == (1, 100)
config_file = 'configs/benchmarks/mnasnet_b1_1_3_custom_cifar100_224_e50.yaml'
cfg.merge_from_file(config_file)
model = build_mnasnet(cfg)
print(model)
outputs = model(data)[KEY_OUTPUT]
print(outputs.shape)
assert outputs.shape == (1, 100)
config_file = 'configs/benchmarks/mnasnet_b1_1_3_torchvision_cifar100_224_e50.yaml'
cfg.merge_from_file(config_file)
model = build_mnasnet(cfg)
print(model)
outputs = model(data)[KEY_OUTPUT]
print(outputs.shape)
assert outputs.shape == (1, 100)
def test_resnest():
config_file = 'configs/benchmarks/resnet/rstd50_2s2x40d_cifar100_224_e100_rmsprop.yaml'
cfg.merge_from_file(config_file)
model = ResNet(cfg)
print(model)
test_data(model, (3, 3, 224, 224), (3, 100))
config_file = 'configs/benchmarks/resnet/rstd50_2s2x40d_fast_cifar100_224_e100_rmsprop.yaml'
cfg.merge_from_file(config_file)
model = ResNet(cfg)
print(model)
test_data(model, (3, 3, 224, 224), (3, 100))
config_file = 'configs/benchmarks/resnet/rstd50_2s2x40d_fast_official_cifar100_224_e100_rmsprop.yaml'
cfg.merge_from_file(config_file)
model = ResNet(cfg)
print(model)
test_data(model, (3, 3, 224, 224), (3, 100))
def test_config():
data = torch.randn(1, 3, 224, 224)
config_file = 'configs/benchmarks/mbv3_large_se_custom_cifar100_224_e50.yaml'
cfg.merge_from_file(config_file)
model = build_mobilenet_v3(cfg)
print(model)
outputs = model(data)
outputs = model(data)[KEY_OUTPUT]
print(outputs.shape)
assert outputs.shape == (1, 100)
config_file = 'configs/benchmarks/mbv3_large_se_hsigmoid_custom_cifar100_224_e50.yaml'
cfg.merge_from_file(config_file)
model = build_mobilenet_v3(cfg)
print(model)
outputs = model(data)
outputs = model(data)[KEY_OUTPUT]
print(outputs.shape)
assert outputs.shape == (1, 100)
config_file = 'configs/benchmarks/mbv3_large_custom_cifar100_224_e50.yaml'
cfg.merge_from_file(config_file)
model = build_mobilenet_v3(cfg)
print(model)
outputs = model(data)
outputs = model(data)[KEY_OUTPUT]
print(outputs.shape)
assert outputs.shape == (1, 100)
config_file = 'configs/benchmarks/mbv3_small_se_custom_cifar100_224_e50.yaml'
cfg.merge_from_file(config_file)
model = build_mobilenet_v3(cfg)
print(model)
outputs = model(data)
outputs = model(data)[KEY_OUTPUT]
print(outputs.shape)
assert outputs.shape == (1, 100)
config_file = 'configs/benchmarks/mbv3_small_se_hsigmoid_custom_cifar100_224_e50.yaml'
cfg.merge_from_file(config_file)
model = build_mobilenet_v3(cfg)
print(model)
outputs = model(data)
outputs = model(data)[KEY_OUTPUT]
print(outputs.shape)
assert outputs.shape == (1, 100)
config_file = 'configs/benchmarks/mbv3_small_custom_cifar100_224_e50.yaml'
cfg.merge_from_file(config_file)
model = build_mobilenet_v3(cfg)
print(model)
outputs = model(data)
outputs = model(data)[KEY_OUTPUT]
print(outputs.shape)
assert outputs.shape == (1, 100)
def test_config():
config_file = "configs/benchmarks/mbv1_custom_cifar100_224.yaml"
cfg.merge_from_file(config_file)
model = build_mobilenet_v1(cfg)
print(model)
def test_config_file():
data = torch.randn(3, 3, 224, 224)
print('repvgg_b2g4_custom_cifar100_224_e100_sgd')
config_file = "configs/benchmarks/repvgg/repvgg_b2g4_cifar100_224_e100_sgd_calr.yaml"
cfg.merge_from_file(config_file)
device = torch.device('cpu')
model = build_recognizer(cfg, device)
print(model)
outputs = model(data)[KEY_OUTPUT]
assert outputs.shape == (3, 100)
fuse_repvgg_block(model)
print(model)
outputs = model(data)[KEY_OUTPUT]
assert outputs.shape == (3, 100)
print('repvgg_b2g4_acb_custom_cifar100_224_e100_sgd')
config_file = "configs/benchmarks/repvgg/repvgg_b2g4_acb_cifar100_224_e100_sgd_calr.yaml"
cfg.merge_from_file(config_file)
device = torch.device('cpu')
model = build_recognizer(cfg, device)
print(model)
outputs = model(data)[KEY_OUTPUT]
assert outputs.shape == (3, 100)
# 注意：如果在RepVGG中嵌入了ACBlock，融合时应该先acb再regvgg
fuse_acblock(model)
print(model)
fuse_repvgg_block(model)
print(model)
outputs = model(data)[KEY_OUTPUT]
assert outputs.shape == (3, 100)
print('acb_repvgg_b2g4_custom_cifar100_224_e100_sgd')
config_file = "configs/benchmarks/repvgg/acb_repvgg_b2g4_cifar100_224_e100_sgd_calr.yaml"
cfg.merge_from_file(config_file)
device = torch.device('cpu')
model = build_recognizer(cfg, device)
print(model)
outputs = model(data)[KEY_OUTPUT]
assert outputs.shape == (3, 100)
# 注意：如果先嵌入ACBlock再嵌入RepVGGBlock，那么融合时应该先repvgg_block再acblock
fuse_repvgg_block(model)
print(model)
fuse_acblock(model)
print(model)
outputs = model(data)[KEY_OUTPUT]
assert outputs.shape == (3, 100)
print('rxtd50_32x4d_acb_rvb_custom_cifar100_224_e100_sgd')
config_file = "configs/benchmarks/repvgg/rxtd50_32x4d_acb_rvb_cifar100_224_e100_sgd_calr.yaml"
cfg.merge_from_file(config_file)
device = torch.device('cpu')
model = build_recognizer(cfg, device)
print(model)
outputs = model(data)[KEY_OUTPUT]
assert outputs.shape == (3, 100)
# 注意：如果先嵌入ACBlock再嵌入RepVGGBlock，那么融合时应该先repvgg_block再acblock
fuse_repvgg_block(model)
print(model)
fuse_acblock(model)
print(model)
outputs = model(data)[KEY_OUTPUT]
assert outputs.shape == (3, 100)
print('rxtd50_32x4d_rvb_acb_custom_cifar100_224_e100_sgd')
config_file = "configs/benchmarks/repvgg/rxtd50_32x4d_rvb_acb_cifar100_224_e100_sgd_calr.yaml"
cfg.merge_from_file(config_file)
device = torch.device('cpu')
model = build_recognizer(cfg, device)
print(model)
outputs = model(data)[KEY_OUTPUT]
assert outputs.shape == (3, 100)
# 注意：如果先嵌入RepVGGBlock再嵌入ACBlock，那么逆序融合
fuse_acblock(model)
print(model)
fuse_repvgg_block(model)
print(model)
outputs = model(data)[KEY_OUTPUT]
assert outputs.shape == (3, 100)
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)
