def __init__(self, root, type='train', transform=None):
self.root = os.path.expanduser(root)
self.transform = transform
self.type = type
all_train_image_list = misc.load_pickle(
os.path.join(self.root, 'train_img_list.pkl'))
all_test_image_list = misc.load_pickle(
os.path.join(self.root, 'val_img_list.pkl'))
self.train_image_list = []
self.train_labels = []
self.val_image_list = []
self.val_labels = []
self.test_image_list = []
self.test_labels = []
for i in range(1000):
self.train_image_list += all_train_image_list[i][:-50]
self.train_labels += [i] * len(all_train_image_list[i][:-50])
self.val_image_list += all_train_image_list[i][-50:]
self.val_labels += [i] * 50
self.test_image_list += all_test_image_list[i]
self.test_labels += [i] * 50
if self.type == 'train':
self.data = self.train_image_list
self.labels = self.train_labels
elif self.type == 'val':
self.data = self.val_image_list
self.labels = self.val_labels
elif self.type == 'train+val':
self.data = self.train_image_list + self.val_image_list
self.labels = self.train_labels + self.val_labels
elif self.type == 'test':
self.data = self.test_image_list
self.labels = self.test_labels
def get_film_reviews():
"""Gets the film reviews, building them if they are not saved as a pickle file.
@returns {List} reviews"""
if os.path.isfile("{}/{}".format(os.environ['PICKLE_DIRECTORY'], 'movie_reviews.pkl')):
print("Movie Reviews found, using pickle file.")
return load_pickle('movie_reviews.pkl')
else:
print("Building movie reviews.")
return build_film_reviews()
def get_classifier():
"""Gets the film classifier - loading it from memory if
stored as a pickle file, else creating one if not stored.
@returns {Classifier} Film Classifier"""
if os.path.isfile("{}/{}".format(os.environ['PICKLE_DIRECTORY'], 'filmClassifier.pkl')):
print("Classifier found, using pickle file")
return load_pickle('filmClassifier.pkl')
else:
print("Classifier not found, creating classifier")
return build_classifier()
def __init__(self, images_list_path, start_class, end_class, num_per_class, random_order=False, transform=None):
self.images_list = load_pickle(images_list_path)
self.transform = transform
self.targets = []
self.images = []
classes = list(range(1000))[start_class: end_class]
for i in classes:
self.targets.extend([i] * num_per_class)
imgs = self.images_list[i]
if random_order:
random.shuffle(imgs)
self.images.extend(imgs[:num_per_class])
else:
self.images.extend(imgs[:num_per_class])
from sklearn import linear_model
from sklearn.model_selection import cross_val_score
import numpy as np
import misc
print("Loading data...")
training_data = misc.load_pickle('train')
print(training_data.shape)
lr = linear_model.LinearRegression()
X = training_data[:, 1:training_data.shape[1]]
Y = training_data[:, 0]
print("Fitting...")
lr.fit(X, Y)
trainloader = torch.utils.data.DataLoader(trainset,
batch_size=args.train_batch_size,
shuffle=True,
num_workers=2)
testset = datasets.CIFAR10(root='./data/cifar10',
type='test',
transform=transform_val)
testloader = torch.utils.data.DataLoader(testset,
batch_size=100,
shuffle=False,
num_workers=2)
print('==> Initializing model...')
pruned_cfg = misc.load_pickle(
'logs/seed-%d/%s-%s/channel-%d-sparsity-%.2f/pruned_cfg-%.2f.pkl' %
(args.seed, args.dataset, args.arch, args.expanded_inchannel,
args.sparsity_level, args.pruned_ratio))
model = models.__dict__[args.arch](args.num_classes, args.expanded_inchannel,
pruned_cfg)
model = model.to(args.device)
optimizer = torch.optim.SGD(model.parameters(),
lr=args.lr,
momentum=args.mm,
weight_decay=args.wd)
scheduler = torch.optim.lr_scheduler.MultiStepLR(
optimizer,
milestones=[int(args.epochs * 0.5),
int(args.epochs * 0.75)],
return loss
if __name__ == "__main__":
expanded_inchannel = 80
pruned_ratio = 0.7
sparsity_level = 0.5
arch = 'resnet50'
multiplier = 1.0
logdir = 'imagenet-%s/channel-%d-pruned-%.2f' % (arch, expanded_inchannel,
pruned_ratio)
import misc
import models
print("=> Using model {}".format(arch))
pruned_cfg = misc.load_pickle(
'/home/victorfang/pruning-from-scratch/script/logs/imagenet-%s/channel-%d-sparsity-%.2f/pruned_cfg-%.2f.pkl'
% (arch, expanded_inchannel, sparsity_level, pruned_ratio))
model = models.__dict__[arch](1000, expanded_inchannel, multiplier,
pruned_cfg)
# checkpoint=torch.load('/home/victorfang/pruning-from-scratch/data/model_saved/imagenet-resnet50-sparsity-0.50/channel-80-pruned-0.70/checkpoint/flop=1262676704,accuracy=0.69436.tar')
# model.load_state_dict(checkpoint['state_dict'])
# model=nn.DataParallel(model)
model.cuda()
train(net=model,
net_name=arch,
exp_name='imagenet-%s-sparsity-%.2f/channel-%d-pruned-%.2f_new' %
(arch, sparsity_level, expanded_inchannel, pruned_ratio),
learning_rate=0.1,
learning_rate_decay_epoch=[2 * i for i in [30, 60, 90]],
num_epochs=2 * 100,
parser.add_argument('--seed', default=4699, type=int)
args = parser.parse_args()
args.seed = misc.set_seed(args.seed)
args.device = 'cuda'
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
args.eps = 0.001
args.num_classes = 100
args.logdir = 'logs/seed-%d/%s-%s/channel-%d-sparsity-%.2f' % (
args.seed, args.dataset, args.arch, args.expanded_inchannel,
args.sparsity_level)
gates_params = misc.load_pickle(os.path.join(args.logdir, 'channel_gates.pkl'))
def calculate_flops(model, input_size=(1, 3, 32, 32)):
model = flop_counter.add_flops_counting_methods(model)
model.eval().start_flops_count()
inp = torch.randn(*input_size)
out = model(inp)
flops = model.compute_average_flops_cost()
return flops
print('==> Initializing full model...')
model = models.__dict__[args.arch](args.num_classes)
full_flops = calculate_flops(model)
args.gpu = 0
args.world_size = 1
if args.distributed:
args.gpu = args.local_rank % torch.cuda.device_count()
torch.cuda.set_device(args.gpu)
torch.distributed.init_process_group(backend='nccl', init_method='env://')
args.world_size = torch.distributed.get_world_size()
if args.local_rank == 0:
misc.prepare_logging(args)
print("=> Using model {}".format(args.arch))
pruned_cfg = misc.load_pickle(
'logs/imagenet-%s/channel-%d-sparsity-%.2f/pruned_cfg-%.2f.pkl' %
(args.arch, args.expanded_inchannel, args.sparsity_level,
args.pruned_ratio))
model = models.__dict__[args.arch](1000, args.expanded_inchannel,
args.multiplier, pruned_cfg)
model = model.cuda()
model = DDP(model, delay_allreduce=True)
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().cuda()
if args.label_smooth:
class CrossEntropyLabelSmooth(nn.Module):
def __init__(self, num_classes, epsilon):
super(CrossEntropyLabelSmooth, self).__init__()
self.num_classes = num_classes
self.epsilon = epsilon
import cv2
imagenet_urls = [
'http://ml.cs.tsinghua.edu.cn/~chenxi/dataset/val224_compressed.pkl'
]
parser = argparse.ArgumentParser(
description='Extract the ILSVRC2012 val dataset')
parser.add_argument('--in_file',
default='val224_compressed.pkl',
help='input file path')
parser.add_argument('--out_root',
default='~/public_dataset/pytorch/imagenet-data/',
help='output file path')
args = parser.parse_args()
d = misc.load_pickle(args.in_file)
assert len(d['data']) == 50000, len(d['data'])
assert len(d['target']) == 50000, len(d['target'])
data224 = []
data299 = []
for img, target in tqdm.tqdm(zip(d['data'], d['target']), total=50000):
img224 = misc.str2img(img)
img299 = cv2.resize(img224, (299, 299))
data224.append(img224)
data299.append(img299)
data_dict224 = dict(data=np.array(data224).transpose(0, 3, 1, 2),
target=d['target'])
data_dict299 = dict(data=np.array(data299).transpose(0, 3, 1, 2),
target=d['target'])
def get_collaborative_recommender():
"""Loads the algorithm's pickle file from memory
@returns {Algorithm Object} algorithm"""
return load_pickle('collaborative.pkl')