def test_speedup_integration(self):
for model_name in [
'resnet18', 'squeezenet1_1', 'mobilenet_v2', 'densenet121',
'inception_v3'
]:
Model = getattr(models, model_name)
net = Model(pretrained=True, progress=False).to(device)
speedup_model = Model().to(device)
net.eval() # this line is necessary
speedup_model.eval()
# random generate the prune config for the pruner
cfgs = generate_random_sparsity(net)
pruner = L1FilterPruner(net, cfgs)
pruner.compress()
pruner.export_model(MODEL_FILE, MASK_FILE)
pruner._unwrap_model()
state_dict = torch.load(MODEL_FILE)
speedup_model.load_state_dict(state_dict)
zero_bn_bias(net)
zero_bn_bias(speedup_model)
data = torch.ones(BATCH_SIZE, 3, 224, 224).to(device)
ms = ModelSpeedup(speedup_model, data, MASK_FILE)
ms.speedup_model()
ori_out = net(data)
speeded_out = speedup_model(data)
ori_sum = torch.sum(ori_out).item()
speeded_sum = torch.sum(speeded_out).item()
print('Sum of the output of %s (before speedup):' % model_name,
ori_sum)
print('Sum of the output of %s (after speedup):' % model_name,
speeded_sum)
assert (abs(ori_sum - speeded_sum) / abs(ori_sum) < RELATIVE_THRESHOLD) or \
(abs(ori_sum - speeded_sum) < ABSOLUTE_THRESHOLD)
def build_model(model_class, channelcfg):
model = model_class(pretrained=True).cuda()
# dummy_input = torch.ones(16, 3, 224, 224).cuda()
cfglist = []
pos = 0
for name, module in model.named_modules():
if isinstance(module, nn.Conv2d):
sparsity = 1-channelcfg[pos]/module.out_channels
if sparsity > 0:
cfglist.append({'sparsity':sparsity + 1e-5, 'op_names':[name], 'op_types': ['Conv2d']})
pos += 1
pruner = L1FilterPruner(model, cfglist, dummy_input=dummy_input, dependency_aware=True)
pruner.compress()
pruner.export_model('./model.pth', './mask')
pruner._unwrap_model()
del pruner
ms = ModelSpeedup(model, dummy_input, './mask')
ms.speedup_model()
del ms
pos = 0
for name, module in model.named_modules():
if isinstance(module, nn.Conv2d):
print(module.out_channels, channelcfg[pos])
assert module.out_channels == channelcfg[pos]
pos += 1
# del dummy_input
torch.cuda.empty_cache()
return model
def test_mask_conflict(self):
outdir = os.path.join(prefix, 'masks')
os.makedirs(outdir, exist_ok=True)
for name in model_names:
print('Test mask conflict for %s' % name)
model = getattr(models, name)
net = model().to(device)
dummy_input = torch.ones(1, 3, 224, 224).to(device)
# random generate the prune sparsity for each layer
cfglist = []
for layername, layer in net.named_modules():
if isinstance(layer, nn.Conv2d):
# pruner cannot allow the sparsity to be 0 or 1
sparsity = np.random.uniform(0.01, 0.99)
cfg = {
'op_types': ['Conv2d'],
'op_names': [layername],
'sparsity': sparsity
}
cfglist.append(cfg)
pruner = L1FilterPruner(net, cfglist)
pruner.compress()
ck_file = os.path.join(outdir, '%s.pth' % name)
mask_file = os.path.join(outdir, '%s_mask' % name)
pruner.export_model(ck_file, mask_file)
pruner._unwrap_model()
# Fix the mask conflict
mf = MaskConflict(mask_file, net, dummy_input)
fixed_mask = mf.fix_mask_conflict()
mf.export(os.path.join(outdir, '%s_fixed_mask' % name))
# use the channel dependency groud truth to check if
# fix the mask conflict successfully
for dset in channel_dependency_ground_truth[name]:
lset = list(dset)
for i, _ in enumerate(lset):
assert fixed_mask[lset[0]]['weight'].size(0) == fixed_mask[
lset[i]]['weight'].size(0)
w_index1 = self.get_pruned_index(
fixed_mask[lset[0]]['weight'])
w_index2 = self.get_pruned_index(
fixed_mask[lset[i]]['weight'])
assert w_index1 == w_index2
if hasattr(fixed_mask[lset[0]], 'bias'):
b_index1 = self.get_pruned_index(
fixed_mask[lset[0]]['bias'])
b_index2 = self.get_pruned_index(
fixed_mask[lset[i]]['bias'])
assert b_index1 == b_index2
'resnet.layer1.2.conv1', 'resnet.layer1.2.conv2', 'resnet.layer2.0.conv1', 'resnet.layer2.0.conv2', \
'resnet.layer2.0.downsample.0', 'resnet.layer2.1.conv1', 'resnet.layer2.1.conv2', 'resnet.layer2.2.conv1', \
'resnet.layer2.2.conv2', 'resnet.layer2.3.conv1', 'resnet.layer2.3.conv2', 'resnet.layer3.0.conv1', \
'resnet.layer3.0.conv2', 'resnet.layer3.0.downsample.0', 'resnet.layer3.1.conv1', 'resnet.layer3.1.conv2', \
'resnet.layer3.2.conv1', 'resnet.layer3.2.conv2', 'resnet.layer3.3.conv1', 'resnet.layer3.3.conv2', \
'resnet.layer3.4.conv1', 'resnet.layer3.4.conv2', 'resnet.layer3.5.conv1', 'resnet.layer3.5.conv2', \
'resnet.layer4.0.conv1', 'resnet.layer4.0.conv2', 'resnet.layer4.0.downsample.0', 'resnet.layer4.1.conv1', \
'resnet.layer4.1.conv2', 'resnet.layer4.2.conv1', 'resnet.layer4.2.conv2',\
'center.0.0', 'center.1.0', 'decoder4.squeeze.0', 'decoder3.squeeze.0', 'decoder2.squeeze.0','decoder1.squeeze.0', \
]
}]
# Prune model and test accuracy without fine tuning.
# print('=' * 10 + 'Test on the pruned model before fine tune' + '=' * 10)
optimizer_finetune = optimizer
pruner = L1FilterPruner(model, configure_list, optimizer_finetune)
model = pruner.compress()
# Code for fots training
train_folder_syn = args.train_folder_syn
train_folder_sample = args.train_folder_sample
output_path = args.save_dir
data_set = datasets.MergeText(train_folder_syn, train_folder_sample, datasets.transform, train=True)
dl = torch.utils.data.DataLoader(data_set, batch_size=args.batch_size, shuffle=True,
sampler=None, batch_sampler=None, num_workers=args.num_workers)
dl_val = None
if args.val:
data_set_val = datasets.MergeText(train_folder_syn, train_folder_sample, datasets.transform, train=False)
dl_val = torch.utils.data.DataLoader(data_set_val, batch_size=1, shuffle=True,
sampler=None, batch_sampler=None, num_workers=args.num_workers)
max_batches_per_iter_cnt = 2
def prune_model_l1(model):
config_list = [{'sparsity': SPARSITY, 'op_types': ['Conv2d']}]
pruner = L1FilterPruner(model, config_list)
pruner.compress()
pruner.export_model(model_path=MODEL_FILE, mask_path=MASK_FILE)
def main():
torch.manual_seed(0)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
train_loader = torch.utils.data.DataLoader(datasets.CIFAR10(
'./data.cifar10',
train=True,
download=True,
transform=transforms.Compose([
transforms.Pad(4),
transforms.RandomCrop(32),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465),
(0.2023, 0.1994, 0.2010))
])),
batch_size=64,
shuffle=True)
test_loader = torch.utils.data.DataLoader(datasets.CIFAR10(
'./data.cifar10',
train=False,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465),
(0.2023, 0.1994, 0.2010))
])),
batch_size=200,
shuffle=False)
model = VGG(depth=16)
model.to(device)
# Train the base VGG-16 model
print('=' * 10 + 'Train the unpruned base model' + '=' * 10)
optimizer = torch.optim.SGD(model.parameters(),
lr=0.1,
momentum=0.9,
weight_decay=1e-4)
lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer, 160, 0)
for epoch in range(160):
print('# Epoch {} #'.format(epoch))
train(model, device, train_loader, optimizer)
test(model, device, test_loader)
lr_scheduler.step(epoch)
torch.save(model.state_dict(), 'vgg16_cifar10.pth')
# Test base model accuracy
print('=' * 10 + 'Test on the original model' + '=' * 10)
model.load_state_dict(torch.load('vgg16_cifar10.pth'))
test(model, device, test_loader)
# top1 = 93.51%
# Pruning Configuration, all convolution layers are pruned out 80% filters according to the L1 norm
configure_list = [{
'sparsity': 0.8,
'op_types': ['Conv2d'],
}]
# Prune model and test accuracy without fine tuning.
print('=' * 10 + 'Test on the pruned model before fine tune' + '=' * 10)
pruner = L1FilterPruner(model, configure_list)
model = pruner.compress()
test(model, device, test_loader)
# top1 = 10.00%
# Fine tune the pruned model for 40 epochs and test accuracy
print('=' * 10 + 'Fine tuning' + '=' * 10)
optimizer_finetune = torch.optim.SGD(model.parameters(),
lr=0.001,
momentum=0.9,
weight_decay=1e-4)
best_top1 = 0
kd_teacher_model = VGG(depth=16)
kd_teacher_model.to(device)
kd_teacher_model.load_state_dict(torch.load('vgg16_cifar10.pth'))
kd = KnowledgeDistill(kd_teacher_model, kd_T=5)
for epoch in range(40):
pruner.update_epoch(epoch)
print('# Epoch {} #'.format(epoch))
train(model, device, train_loader, optimizer_finetune, kd)
top1 = test(model, device, test_loader)
if top1 > best_top1:
best_top1 = top1
# Export the best model, 'model_path' stores state_dict of the pruned model,
# mask_path stores mask_dict of the pruned model
pruner.export_model(model_path='pruned_vgg16_cifar10.pth',
mask_path='mask_vgg16_cifar10.pth')
# Test the exported model
print('=' * 10 + 'Test on the pruned model after fine tune' + '=' * 10)
new_model = VGG(depth=16)
new_model.to(device)
new_model.load_state_dict(torch.load('pruned_vgg16_cifar10.pth'))
test(new_model, device, test_loader)
def main():
torch.manual_seed(0)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
train_loader = torch.utils.data.DataLoader(datasets.CIFAR10(
'./data.cifar10',
train=True,
download=True,
transform=transforms.Compose([
transforms.Pad(4),
transforms.RandomCrop(32),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465),
(0.2023, 0.1994, 0.2010))
])),
batch_size=64,
shuffle=True)
test_loader = torch.utils.data.DataLoader(datasets.CIFAR10(
'./data.cifar10',
train=False,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465),
(0.2023, 0.1994, 0.2010))
])),
batch_size=200,
shuffle=False)
model = VGG(depth=16)
model.to(device)
# Train the base VGG-16 model
print('=' * 10 + 'Train the unpruned base model' + '=' * 10)
optimizer = torch.optim.SGD(model.parameters(),
lr=0.1,
momentum=0.9,
weight_decay=1e-4)
lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer, 160, 0)
for epoch in range(160):
train(model, device, train_loader, optimizer)
test(model, device, test_loader)
lr_scheduler.step(epoch)
torch.save(model.state_dict(), 'vgg16_cifar10.pth')
# Test base model accuracy
print('=' * 10 + 'Test on the original model' + '=' * 10)
model.load_state_dict(torch.load('vgg16_cifar10.pth'))
test(model, device, test_loader)
# top1 = 93.51%
# Pruning Configuration, in paper 'PRUNING FILTERS FOR EFFICIENT CONVNETS',
# Conv_1, Conv_8, Conv_9, Conv_10, Conv_11, Conv_12 are pruned with 50% sparsity, as 'VGG-16-pruned-A'
configure_list = [{
'sparsity':
0.5,
'op_types': ['default'],
'op_names': [
'feature.0', 'feature.24', 'feature.27', 'feature.30',
'feature.34', 'feature.37'
]
}]
# Prune model and test accuracy without fine tuning.
print('=' * 10 + 'Test on the pruned model before fine tune' + '=' * 10)
pruner = L1FilterPruner(model, configure_list)
model = pruner.compress()
test(model, device, test_loader)
# top1 = 88.19%
# Fine tune the pruned model for 40 epochs and test accuracy
print('=' * 10 + 'Fine tuning' + '=' * 10)
optimizer_finetune = torch.optim.SGD(model.parameters(),
lr=0.001,
momentum=0.9,
weight_decay=1e-4)
best_top1 = 0
for epoch in range(40):
pruner.update_epoch(epoch)
print('# Epoch {} #'.format(epoch))
train(model, device, train_loader, optimizer_finetune)
top1 = test(model, device, test_loader)
if top1 > best_top1:
best_top1 = top1
# Export the best model, 'model_path' stores state_dict of the pruned model,
# mask_path stores mask_dict of the pruned model
pruner.export_model(model_path='pruned_vgg16_cifar10.pth',
mask_path='mask_vgg16_cifar10.pth')
# Test the exported model
print('=' * 10 + 'Test on the pruned model after fine tune' + '=' * 10)
new_model = VGG(depth=16)
new_model.to(device)
new_model.load_state_dict(torch.load('pruned_vgg16_cifar10.pth'))
test(new_model, device, test_loader)
def main(args):
# prepare dataset
torch.manual_seed(0)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
train_loader, val_loader, criterion = get_data(args)
model, optimizer = get_trained_model_optimizer(args, device, train_loader,
val_loader, criterion)
def short_term_fine_tuner(model, epochs=1):
for epoch in range(epochs):
train(args, model, device, train_loader, criterion, optimizer,
epoch)
def trainer(model, optimizer, criterion, epoch, callback):
return train(args,
model,
device,
train_loader,
criterion,
optimizer,
epoch=epoch,
callback=callback)
def evaluator(model):
return test(model, device, criterion, val_loader)
# used to save the performance of the original & pruned & finetuned models
result = {'flops': {}, 'params': {}, 'performance': {}}
flops, params = count_flops_params(model, get_input_size(args.dataset))
result['flops']['original'] = flops
result['params']['original'] = params
evaluation_result = evaluator(model)
print('Evaluation result (original model): %s' % evaluation_result)
result['performance']['original'] = evaluation_result
# module types to prune, only "Conv2d" supported for channel pruning
if args.base_algo in ['l1', 'l2']:
op_types = ['Conv2d']
elif args.base_algo == 'level':
op_types = ['default']
config_list = [{'sparsity': args.sparsity, 'op_types': op_types}]
dummy_input = get_dummy_input(args, device)
if args.pruner == 'L1FilterPruner':
pruner = L1FilterPruner(model, config_list)
elif args.pruner == 'L2FilterPruner':
pruner = L2FilterPruner(model, config_list)
elif args.pruner == 'ActivationMeanRankFilterPruner':
pruner = ActivationMeanRankFilterPruner(model, config_list)
elif args.pruner == 'ActivationAPoZRankFilterPruner':
pruner = ActivationAPoZRankFilterPruner(model, config_list)
elif args.pruner == 'NetAdaptPruner':
pruner = NetAdaptPruner(model,
config_list,
short_term_fine_tuner=short_term_fine_tuner,
evaluator=evaluator,
base_algo=args.base_algo,
experiment_data_dir=args.experiment_data_dir)
elif args.pruner == 'ADMMPruner':
# users are free to change the config here
if args.model == 'LeNet':
if args.base_algo in ['l1', 'l2']:
config_list = [{
'sparsity': 0.8,
'op_types': ['Conv2d'],
'op_names': ['conv1']
}, {
'sparsity': 0.92,
'op_types': ['Conv2d'],
'op_names': ['conv2']
}]
elif args.base_algo == 'level':
config_list = [{
'sparsity': 0.8,
'op_names': ['conv1']
}, {
'sparsity': 0.92,
'op_names': ['conv2']
}, {
'sparsity': 0.991,
'op_names': ['fc1']
}, {
'sparsity': 0.93,
'op_names': ['fc2']
}]
else:
raise ValueError('Example only implemented for LeNet.')
pruner = ADMMPruner(model,
config_list,
trainer=trainer,
num_iterations=2,
training_epochs=2)
elif args.pruner == 'SimulatedAnnealingPruner':
pruner = SimulatedAnnealingPruner(
model,
config_list,
evaluator=evaluator,
base_algo=args.base_algo,
cool_down_rate=args.cool_down_rate,
experiment_data_dir=args.experiment_data_dir)
elif args.pruner == 'AutoCompressPruner':
pruner = AutoCompressPruner(
model,
config_list,
trainer=trainer,
evaluator=evaluator,
dummy_input=dummy_input,
num_iterations=3,
optimize_mode='maximize',
base_algo=args.base_algo,
cool_down_rate=args.cool_down_rate,
admm_num_iterations=30,
admm_training_epochs=5,
experiment_data_dir=args.experiment_data_dir)
else:
raise ValueError("Pruner not supported.")
# Pruner.compress() returns the masked model
# but for AutoCompressPruner, Pruner.compress() returns directly the pruned model
model = pruner.compress()
evaluation_result = evaluator(model)
print('Evaluation result (masked model): %s' % evaluation_result)
result['performance']['pruned'] = evaluation_result
if args.save_model:
pruner.export_model(
os.path.join(args.experiment_data_dir, 'model_masked.pth'),
os.path.join(args.experiment_data_dir, 'mask.pth'))
print('Masked model saved to %s', args.experiment_data_dir)
# model speed up
if args.speed_up:
if args.pruner != 'AutoCompressPruner':
if args.model == 'LeNet':
model = LeNet().to(device)
elif args.model == 'vgg16':
model = VGG(depth=16).to(device)
elif args.model == 'resnet18':
model = ResNet18().to(device)
elif args.model == 'resnet50':
model = ResNet50().to(device)
elif args.model == 'mobilenet_v2':
model = models.mobilenet_v2(pretrained=False).to(device)
model.load_state_dict(
torch.load(
os.path.join(args.experiment_data_dir,
'model_masked.pth')))
masks_file = os.path.join(args.experiment_data_dir, 'mask.pth')
m_speedup = ModelSpeedup(model, dummy_input, masks_file, device)
m_speedup.speedup_model()
evaluation_result = evaluator(model)
print('Evaluation result (speed up model): %s' % evaluation_result)
result['performance']['speedup'] = evaluation_result
torch.save(
model.state_dict(),
os.path.join(args.experiment_data_dir, 'model_speed_up.pth'))
print('Speed up model saved to %s', args.experiment_data_dir)
flops, params = count_flops_params(model, get_input_size(args.dataset))
result['flops']['speedup'] = flops
result['params']['speedup'] = params
if args.fine_tune:
if args.dataset == 'mnist':
optimizer = torch.optim.Adadelta(model.parameters(), lr=1)
scheduler = StepLR(optimizer, step_size=1, gamma=0.7)
elif args.dataset == 'cifar10' and args.model == 'vgg16':
optimizer = torch.optim.SGD(model.parameters(),
lr=0.01,
momentum=0.9,
weight_decay=5e-4)
scheduler = MultiStepLR(optimizer,
milestones=[
int(args.fine_tune_epochs * 0.5),
int(args.fine_tune_epochs * 0.75)
],
gamma=0.1)
elif args.dataset == 'cifar10' and args.model == 'resnet18':
optimizer = torch.optim.SGD(model.parameters(),
lr=0.1,
momentum=0.9,
weight_decay=5e-4)
scheduler = MultiStepLR(optimizer,
milestones=[
int(args.fine_tune_epochs * 0.5),
int(args.fine_tune_epochs * 0.75)
],
gamma=0.1)
elif args.dataset == 'cifar10' and args.model == 'resnet50':
optimizer = torch.optim.SGD(model.parameters(),
lr=0.1,
momentum=0.9,
weight_decay=5e-4)
scheduler = MultiStepLR(optimizer,
milestones=[
int(args.fine_tune_epochs * 0.5),
int(args.fine_tune_epochs * 0.75)
],
gamma=0.1)
best_acc = 0
for epoch in range(args.fine_tune_epochs):
train(args, model, device, train_loader, criterion, optimizer,
epoch)
scheduler.step()
acc = evaluator(model)
if acc > best_acc:
best_acc = acc
torch.save(
model.state_dict(),
os.path.join(args.experiment_data_dir,
'model_fine_tuned.pth'))
print('Evaluation result (fine tuned): %s' % best_acc)
print('Fined tuned model saved to %s', args.experiment_data_dir)
result['performance']['finetuned'] = best_acc
with open(os.path.join(args.experiment_data_dir, 'result.json'),
'w+') as f:
json.dump(result, f)
result = {}
evaluation_result = evaluator(model)
print('Evaluation result (original model): %s' % evaluation_result)
result['original'] = evaluation_result
if args.pruning_mode == 'channel':
op_types = ['Conv2d']
elif args.pruning_mode == 'fine_grained':
op_types = ['default']
config_list = [{'sparsity': args.sparsity, 'op_types': op_types}]
if args.pruner == 'L1FilterPruner':
pruner = L1FilterPruner(model, config_list)
elif args.pruner == 'SimulatedAnnealingPruner':
pruner = SimulatedAnnealingPruner(
model,
config_list,
evaluator=evaluator,
cool_down_rate=args.cool_down_rate,
experiment_data_dir=args.experiment_data_dir)
elif args.pruner == 'NetAdaptPruner':
pruner = NetAdaptPruner(model,
config_list,
fine_tuner=fine_tuner,
evaluator=evaluator,
pruning_mode='channel',
pruning_step=args.pruning_step,
experiment_data_dir=args.experiment_data_dir)
def prune_model_l1(model):
config_list = [{'sparsity': SPARSITY, 'op_types': ['Conv2d']}]
pruner = L1FilterPruner(model, config_list)
pruner.compress()
pruner.export_model(model_path='./11_model.pth', mask_path='./l1_mask.pth')