def initialize_from(self, old_model_path, ctx, *args):
if os.path.exists(os.path.join(old_model_path, 'ModelAchi.pkl')):
with open(os.path.join(old_model_path, 'ModelAchi.pkl'),
'rb') as f:
archi_dict = pickle.load(f)
else:
archi_dict = self.default_params
old = models.SphereNet20(archi_dict=archi_dict, use_custom_relu=True)
old.load_params(os.path.join(old_model_path, 'model'))
for name, block in self.features._children.items():
old_block = old.features._children[name]
if isinstance(block, qDense):
block.weight.initialize(init=mx.init.Constant(
old_block.weight._data[0]),
force_reinit=True,
ctx=ctx)
block.bias.initialize(init=mx.init.Constant(
old_block.bias._data[0]),
force_reinit=True,
ctx=ctx)
else:
if not old_block.same_shape:
block.conv0.weight.initialize(init=mx.init.Constant(
old_block.conv0.conv.weight._data[0]),
force_reinit=True,
ctx=ctx)
block.conv0.bias.initialize(init=mx.init.Constant(
old_block.conv0.conv.bias._data[0]),
force_reinit=True,
ctx=ctx)
block.a0.initialize(init=mx.init.Constant(
old_block.a0.alpha._data[0]),
force_reinit=True,
ctx=ctx)
block.conv1.weight.initialize(init=mx.init.Constant(
old_block.conv1.conv.weight._data[0]),
force_reinit=True,
ctx=ctx)
block.conv1.bias.initialize(init=mx.init.Constant(
old_block.conv1.conv.bias._data[0]),
force_reinit=True,
ctx=ctx)
block.a1.initialize(init=mx.init.Constant(
old_block.a1.alpha._data[0]),
force_reinit=True,
ctx=ctx)
block.conv2.weight.initialize(init=mx.init.Constant(
old_block.conv2.conv.weight._data[0]),
force_reinit=True,
ctx=ctx)
block.conv2.bias.initialize(init=mx.init.Constant(
old_block.conv2.conv.bias._data[0]),
force_reinit=True,
ctx=ctx)
block.a2.initialize(init=mx.init.Constant(
old_block.a2.alpha._data[0]),
force_reinit=True,
ctx=ctx)
def cal_my_acc(test_files, target_files):
'''
this method is deprecated
:param test_files:
:param target_files:
:return:
'''
mTransform = MTransform()
normalize = transforms.Normalize(mean=0.5, std=0.5)
transform = transforms.Compose([
transforms.Resize((96, 112)),
transforms.ToTensor(),
normalize,
# mTransform,
])
model = models.SphereNet20()
model.load_params("spherenet_model", ctx=mx.gpu())
correct = 0
total = 0
target_emb = {}
for target_file in target_files:
target_image = transform(nd.array(Image.open(target_file))).as_in_context(mx.gpu())
target_image = nd.expand_dims(target_image, axis=0)
target_label = ''.join(target_file.split('/')[-1].split('.')[:-1])
target_out = model(target_image)
target_emb[target_label] = target_out
test_emb = {}
for test_file in test_files:
test_image = Image.open(test_file)
test_image = nd.expand_dims(transform(nd.array(test_image)), axis=0).as_in_context(mx.gpu())
test_label = ''.join(test_file.split('/')[-1].split('.')[:-1])
test_out = model(test_image)
max_s = mx.nd.zeros(1, ctx=mx.gpu())
max_label = ''
sims = {}
for target_label, target_out in target_emb.items():
similarity = nd.sum(test_out * target_out) / \
(nd.norm(test_out) * nd.norm(target_out))
sims[target_label] = similarity.asscalar()
if max_s < similarity:
max_s = similarity
max_label = target_label
if ''.join(max_label.split('_')[:-1]) == ''.join(test_label.split('_')[:-1]):
correct += 1
else:
print test_label, max_s.asscalar(), max_label
total += 1
test_emb[test_label] = test_out
# print correct, total, float(correct)/total
return float(correct) / total, test_emb, target_emb
def export_sym(model_path, ctx=mx.gpu()):
if os.path.exists(os.path.join(model_path, 'ModelAchi.pkl')):
with open(os.path.join(model_path, 'ModelAchi.pkl'), 'rb') as f:
archi_dict = pickle.load(f)
else:
archi_dict = None
model = models.SphereNet20(archi_dict=archi_dict)
model.load_params(os.path.join(model_path, 'model'), ctx=ctx)
train_data_loader, _, _ = dataset.train_valid_test_loader(
'/home1/CASIA-WebFace/aligned_Webface-112X96', (0.9, 0.05),
batch_size=1)
# run forward once
model.hybridize()
model.get_feature = False
for batch, label in train_data_loader:
batch = batch.as_in_context(ctx)
model(batch)
break
model.export(os.path.join(model_path, 'model_symbol'))
os.system("mkdir " + self.log_dir)
os.system("touch " + log_path)
with open(os.path.join(self.log_dir, "log.txt"), "a") as f:
f.write(str(mstr) + "\n")
if __name__ == '__main__':
args = get_args()
# gpus = [0,1]
# ctx = [mx.gpu(ii) for ii in gpus]
ctx = mx.gpu(args.device_id)
time_info = time.strftime('%Y-%m-%d_%H%M%S', time.localtime(time.time()))
log_dir = os.path.abspath("./log/")
if not args.prune:
model = models.SphereNet20(use_custom_relu=False)
model.load_params('spherenet_model2', ctx=ctx)
if args.log:
log_dir = os.path.abspath("./log/train-" + time_info + "/")
else:
model = models.SphereNet20()
model.load_params(args.model_path, ctx=ctx)
if args.log:
log_dir = os.path.abspath("./log/prune-" + time_info + "/")
p = Printer(log_dir, args.log)
p.log(ctx)
msg = "doing fine tuning(train)" if not args.prune else "doing pruning, using model " + \
args.model_path
p.log(msg)
p.log(model)
if __name__ == "__main__":
# gpus = [0,1]
# ctx = [mx.gpu(ii) for ii in gpus]
model_path1 = "./log/prune-2018-07-13_121936"
model_path2 = "./log/prune-2018-07-14_003715"
ctx = mx.gpu()
archi_dict = {
0:[32,40],
1:[73,74],
2:[75,74],
3:[108,115],
4:[97,115],
5:[106,115],
6:[110,115],
7:[143,205]
}
with open(os.path.join(model_path1,"ModelAchi.pkl"),"rb") as f:
archi_dict1 = pickle.load(f)
# with open(os.path.join(model_path1,"ModelAchi.pkl"),"wb") as f:
# pickle.dump(archi_dict,f)
with open(os.path.join(model_path2,"ModelAchi.pkl"),"rb") as f:
archi_dict2 = pickle.load(f)
model1 = models.SphereNet20(archi_dict=archi_dict1,use_custom_relu=True)
model1.load_params(os.path.join(model_path1,"model"), ctx=ctx)
model2 = models.SphereNet20(archi_dict=archi_dict2,use_custom_relu=True)
model2.load_params(os.path.join(model_path2,"model"), ctx=ctx)
print(test_on_LFW(model1))
print(test_on_LFW(model2))
test_speed(model1)
test_speed(model2)
print 'over'