def test_active_subnet_with_config_and_save_subnet_result():
# config 에 따라 top1 acc 를 뽑을 수 있음
# N_network_try = 100
# exmple net_config
net_config = {'wid': None,
'ks': [7, 5, 5, 3, 7, 5, 3, 5, 7, 5, 3, 7, 5, 7, 7, 3, 5, 3, 7, 5],
'e': [4, 4, 6, 3, 3, 4, 4, 4, 6, 6, 4, 3, 6, 3, 3, 3, 6, 6, 6, 3],
'd': [2, 3, 3, 4, 4],
'r': [208]}
# results = []
arch_manager = ArchManager()
with torch.no_grad():
# for frame in range(N_network_try):
ofa_network = ofa_net('ofa_mbv3_d234_e346_k357_w1.2', pretrained=True)
torch.cuda.empty_cache()
net_config['r'] = [random.choice(arch_manager.resolutions)]
for i in range(len(net_config['e'])):
net_config['ks'][i] = random.choice(arch_manager.kernel_sizes)
net_config['e'][i] = random.choice(arch_manager.expand_ratios)
for i in range(len(net_config['d'])):
net_config['d'][i] = random.choice(arch_manager.depths)
ofa_network.set_active_subnet(ks=net_config['ks'], d=net_config['d'], e=net_config['e'])
imagenet_path = "../once-for-all/tutorial/.imagenet_fake/"
data_loader = get_data_loader(imagenet_path)
# print(ofa_network.module_str)
start = time.time()
top1 = evaluAate_ofa_subnet(
ofa_network,
imagenet_path,
net_config,
data_loader,
batch_size=250,
device='cuda' if torch.cuda.is_available() else 'cpu')
# device = f'cuda:{random.choice([0, 1, 2, 3])}' if torch.cuda.is_available() else 'cpu')
latency = time.time() - start
# results.append({**net_config, 'acc': top1, 'latency': latency})
# del ofa_network
# save
with open(f'../results/acc_{top1}_latency_{latency}.pickle', 'wb') as f:
pickle.dump({**net_config, 'acc': top1, 'latency': latency}, f, pickle.HIGHEST_PROTOCOL)
#os.environ['CUDA_VISIBLE_DEVICES'] = '0'
cuda_available = torch.cuda.is_available()
if cuda_available:
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = True
torch.cuda.manual_seed(random_seed)
print('Using GPU.')
else:
print('Using CPU.')
# %% [markdown]
# Good! Now you have successfully configured the environment! It's time to import the **OFA network** for the following experiments.
# The OFA network used in this tutorial is built upon MobileNetV3 with width multiplier 1.2, supporting elastic depth (2, 3, 4) per stage, elastic expand ratio (3, 4, 6), and elastic kernel size (3, 5 7) per block.
# %%
ofa_network = ofa_net('ofa_mbv3_d234_e346_k357_w1.2', pretrained=True)
print('The OFA Network is ready.')
# %% [markdown]
# Now, let's build the ImageNet dataset and the corresponding dataloader. Notice that **if you're using the CPU,
# we will skip ImageNet evaluation by default** since it will be very slow.
# If you are using the GPU, in case you don't have the full dataset,
# we will download a subset of ImageNet which contains 2,000 images (~250M) for testing.
# If you do have the full ImageNet dataset on your machine, just specify it in `imagenet_data_path` and the downloading script will be skipped.
# %%
if cuda_available:
# path to the ImageNet dataset
print("Please input the path to the ImageNet dataset.\n")
imagenet_data_path = './imagenet_1k' #input()
'ofa_mbv3_d234_e346_k357_w1.2',
'ofa_proxyless_d234_e346_k357_w1.3'
],
help='OFA networks')
args = parser.parse_args()
if args.gpu == 'all':
device_list = range(torch.cuda.device_count())
args.gpu = ','.join(str(_) for _ in device_list)
else:
device_list = [int(_) for _ in args.gpu.split(',')]
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
args.batch_size = args.batch_size * max(len(device_list), 1)
ImagenetDataProvider.DEFAULT_PATH = args.path
ofa_network = ofa_net(args.net, pretrained=True)
run_config = ImagenetRunConfig(test_batch_size=args.batch_size,
n_worker=args.workers)
""" Randomly sample a sub-network,
you can also manually set the sub-network using:
ofa_network.set_active_subnet(ks=7, e=6, d=4)
"""
ofa_network.sample_active_subnet()
subnet = ofa_network.get_active_subnet(preserve_weight=True)
""" Test sampled subnet
"""
run_manager = RunManager('.tmp/eval_subnet', subnet, run_config, init=False)
# assign image size: 128, 132, ..., 224
run_config.data_provider.assign_active_img_size(224)
run_manager.reset_running_statistics(net=subnet)
def quantization(title='optimize', model_name='', file_path=''):
data_dir = args.data_dir
quant_mode = args.quant_mode
finetune = args.fast_finetune
deploy = args.deploy
batch_size = args.batch_size
subset_len = args.subset_len
if quant_mode != 'test' and deploy:
deploy = False
print(
r'Warning: Exporting xmodel needs to be done in quantization test mode, turn off it in this running!'
)
if deploy and (batch_size != 1 or subset_len != 1):
print(
r'Warning: Exporting xmodel needs batch size to be 1 and only 1 iteration of inference, change them automatically!'
)
batch_size = 1
subset_len = 1
# model = resnet50().cpu()
ofa_network = ofa_net('ofa_resnet50', pretrained=True)
with open(args.model_dir + 'net_config_' + model_name + '.pickle',
'rb') as handle:
arch_dict = pickle.load(handle)
ofa_network.set_active_subnet(**arch_dict)
model = ofa_network.get_active_subnet()
model.load_state_dict(torch.load(file_path))
size = args.image_size
input = torch.randn([batch_size, 3, size, size])
if quant_mode == 'float':
quant_model = model
else:
## new api
####################################################################################
quantizer = torch_quantizer(quant_mode, model, (input), device=device)
quant_model = quantizer.quant_model
#####################################################################################
# to get loss value after evaluation
loss_fn = torch.nn.CrossEntropyLoss().to(device)
val_loader, _ = load_data(subset_len=subset_len,
train=False,
batch_size=batch_size,
sample_method='random',
data_dir=data_dir,
model_name=model_name)
# fast finetune model or load finetuned parameter before test
if finetune == True:
ft_loader, _ = load_data(subset_len=1024,
train=False,
batch_size=batch_size,
sample_method=None,
data_dir=data_dir,
model_name=model_name)
if quant_mode == 'calib':
quantizer.fast_finetune(evaluate,
(quant_model, ft_loader, loss_fn))
elif quant_mode == 'test':
quantizer.load_ft_param()
# record modules float model accuracy
# add modules float model accuracy here
acc_org1 = 0.0
acc_org5 = 0.0
loss_org = 0.0
#register_modification_hooks(model_gen, train=False)
acc1_gen, acc5_gen, loss_gen = evaluate(quant_model, val_loader, loss_fn)
# logging accuracy
print('loss: %g' % (loss_gen))
print('top-1 / top-5 accuracy: %g / %g' % (acc1_gen, acc5_gen))
# handle quantization result
if quant_mode == 'calib':
quantizer.export_quant_config()
if deploy:
quantizer.export_xmodel(deploy_check=False)
random.seed(random_seed)
np.random.seed(random_seed)
torch.manual_seed(random_seed)
print('Successfully imported all packages and configured random seed to %d!' %
random_seed)
#os.environ['CUDA_VISIBLE_DEVICES'] = '0'
cuda_available = torch.cuda.is_available()
if cuda_available:
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = True
torch.cuda.manual_seed(random_seed)
print('Using GPU.')
else:
print('Using CPU.')
# ofa_network = ofa_net('ofa_mbv3_d234_e346_k357_w1.2', pretrained=True)
ofa_network = ofa_net('ofa_resnet50', pretrained=True)
print('The OFA Network is ready.')
if cuda_available:
# path to the ImageNet dataset
print("Please input the path to the ImageNet dataset.\n")
imagenet_data_path = "/gdata/ImageNet2012"
print(os.path.isdir(imagenet_data_path))
# if 'imagenet_data_path' is empty, download a subset of ImageNet containing 2000 images (~250M) for test
# if not os.path.isdir(imagenet_data_path):
# os.makedirs(imagenet_data_path, exist_ok=True)
# download_url('https://hanlab.mit.edu/files/OnceForAll/ofa_cvpr_tutorial/imagenet_1k.zip', model_dir='data')
# ! cd data && unzip imagenet_1k 1>/dev/null && cd ..
# ! cp -r data/imagenet_1k/* $imagenet_data_path
# ! rm -rf data
# print('%s is empty. Download a subset of ImageNet for test.' % imagenet_data_path)
def make_db(self):
self.ofa_network = ofa_net('ofa_mbv3_d234_e346_k357_w1.0', pretrained=True)
self.run_config = ImagenetRunConfig(test_batch_size=self.args.batch_size,
n_worker=20)
database = []
st_time = time.time()
f = open(f'{self.path}/txt_{self.index}.txt', 'w')
for dn in range(10000):
best_pp = -1
best_info = None
dls = None
with torch.no_grad():
if self.model_name == 'generator':
for i in range(10):
net_setting = self.ofa_network.sample_active_subnet()
subnet = self.ofa_network.get_active_subnet(preserve_weight=True)
if i == 0:
run_manager = RunManager('.tmp/eval_subnet', self.args, subnet,
self.run_config, init=False, pp=self.predictor)
self.run_config.data_provider.assign_active_img_size(224)
dls = {j: copy.deepcopy(run_manager.data_loader) for j in range(1, 10)}
else:
run_manager = RunManager('.tmp/eval_subnet', self.args, subnet,
self.run_config,
init=False, data_loader=dls[i], pp=self.predictor)
run_manager.reset_running_statistics(net=subnet)
loss, (top1, top5), pred_acc \
= run_manager.validate(net=subnet, net_setting=net_setting)
if best_pp < pred_acc:
best_pp = pred_acc
print('[%d] class=%d,\t loss=%.5f,\t top1=%.1f,\t top5=%.1f' % (
dn, len(run_manager.cls_lst), loss, top1, top5))
info_dict = {'loss': loss,
'top1': top1,
'top5': top5,
'net': net_setting,
'class': run_manager.cls_lst,
'params': run_manager.net_info['params'],
'flops': run_manager.net_info['flops'],
'test_transform': run_manager.test_transform
}
best_info = info_dict
elif self.model_name == 'predictor':
net_setting = self.ofa_network.sample_active_subnet()
subnet = self.ofa_network.get_active_subnet(preserve_weight=True)
run_manager = RunManager('.tmp/eval_subnet', self.args, subnet, self.run_config, init=False)
self.run_config.data_provider.assign_active_img_size(224)
run_manager.reset_running_statistics(net=subnet)
loss, (top1, top5), _ = run_manager.validate(net=subnet)
print('[%d] class=%d,\t loss=%.5f,\t top1=%.1f,\t top5=%.1f' % (
dn, len(run_manager.cls_lst), loss, top1, top5))
best_info = {'loss': loss,
'top1': top1,
'top5': top5,
'net': net_setting,
'class': run_manager.cls_lst,
'params': run_manager.net_info['params'],
'flops': run_manager.net_info['flops'],
'test_transform': run_manager.test_transform
}
database.append(best_info)
if (len(database)) % 10 == 0:
msg = f'{(time.time() - st_time) / 60.0:0.2f}(min) save {len(database)} database, {self.index} id'
print(msg)
f.write(msg + '\n')
f.flush()
torch.save(database, f'{self.path}/database_{self.index}.pt')
