def write_pruning_result(model, test_dl, rcm, sequence, save_name):
if rcm == 1:
result = evaluate(model, test_dl)
else:
result = distributed_evaluate(model, test_dl)
if save_name is not None:
if not (os.path.isdir("analysis")):
os.makedirs(os.path.join("analysis"))
with open('./analysis/{}_eval.csv'.format(save_name), 'a',
newline='') as csvfile:
acc_file = csv.writer(csvfile)
acc_file.writerow(
[sequence,
float(result["acc"]),
float(result["f1score"])])
def evaluation_rcm(option):
if option["model"] == "VGG16":
model = [
models.VGG(layers=16, classification=10)
for _ in range(option["rcm"])
]
elif option["model"] == "MobileNet":
model = [
models.MobileNet(alpha=option["alpha"], classification=10)
for _ in range(option["rcm"])
]
elif option["model"] == "LeNet5":
model = [
models.LeNet5(classification=10) for _ in range(option["rcm"])
]
elif option["model"] == "ConvNet":
model = [
models.ConvNet(classification=10) for _ in range(option["rcm"])
]
elif option["model"] == "ConvNet_s":
model = [
models.ConvNet_s(classification=10) for _ in range(option["rcm"])
]
if option["model"] == "LeNet5":
test_dl = load_mnist("test", 1, 1, option["batch"])
else:
test_dl = load_cifar10("test", 1, 1, option["batch"])
i = 0
for m in model:
load_model = torch.load(option["model_path"][i],
map_location=option["dev"])
m._modules = load_model['_modules']
m.load_state_dict(load_model['state_dict'])
m = m.to(option["dev"])
i = i + 1
result = distributed_evaluate(model, test_dl)
def train_rcm(option):
_save_name = option["save_name"]
if option["model"] == "VGG16":
model = [
models.VGG(layers=16, classification=10)
for _ in range(option["rcm"])
]
elif option["model"] == "MobileNet":
model = [
models.MobileNet(alpha=option["alpha"], classification=10)
for _ in range(option["rcm"])
]
if option["alpha"] == 1.0:
_save_name = _save_name + "x1.0"
else:
_save_name = _save_name + "x{:n}".format(option["alpha"])
elif option["model"] == "LeNet5":
model = [
models.LeNet5(classification=10) for _ in range(option["rcm"])
]
lenet_size = 32
elif option["model"] == "LeNet300100":
model = [models.LeNet300100() for _ in range(option["rcm"])]
lenet_size = 28
elif option["model"] == "ConvNet":
model = [
models.ConvNet(classification=10) for _ in range(option["rcm"])
]
elif option["model"] == "ConvNet_s":
model = [
models.ConvNet_s(classification=10) for _ in range(option["rcm"])
]
elif option["model"] == "ResNet20":
model = [
models.ResNet20(classification=10) for _ in range(option["rcm"])
]
_save_name = _save_name + "_rcm{:d}".format(option["rcm"])
if option["model"] == "LeNet5" or option["model"] == "LeNet300100":
train_dl = [
load_mnist("train", option["rcm"], i + 1, option["batch"],
lenet_size) for i in range(option["rcm"])
] # ([train_dl[0], ...], [valid_dl[0], ...])
test_dl = load_mnist("test", 1, 1, option["batch"], lenet_size)
else:
train_dl = [
load_cifar10("train", option["rcm"], i + 1, option["batch"])
for i in range(option["rcm"])
]
test_dl = load_cifar10("test", 1, 1, option["batch"])
load_model = torch.load(option["model_path"][0],
map_location=option["dev"])
for i in range(option["rcm"]):
print("Reduced Classification model #{:d}".format(i + 1))
classification = int(10 / option["rcm"]) + 1
model[i].load_state_dict(load_model["state_dict"])
last_in_features = model[i].classifier[-1].in_features
model[i].classifier[-1] = nn.Linear(in_features=last_in_features,
out_features=classification)
model[i] = model[i].to(option["dev"])
loss_fn = nn.CrossEntropyLoss()
opt = optim.Adam(model[i].parameters(), lr=option["lr"])
save_name = _save_name + "_{:d}".format(i + 1)
history = fit(model[i], train_dl[i][0], train_dl[i][1], loss_fn, opt,
option["epoch"], option["schedule"], save_name,
option["save_option"])
result = distributed_evaluate(model, test_dl)
def finetuning_rcm(option):
_save_name = option["save_name"]
if option["model"] == "VGG16":
model = [
models.VGG(layers=16, classification=10)
for _ in range(option["rcm"])
]
elif option["model"] == "MobileNet":
model = [
models.MobileNet(alpha=option["alpha"], classification=10)
for _ in range(option["rcm"])
]
if option["alpha"] == 1.0:
_save_name = _save_name + "x1.0"
else:
_save_name = _save_name + "x{:n}".format(option["alpha"])
elif option["model"] == "LeNet5":
model = [
models.LeNet5(classification=10) for _ in range(option["rcm"])
]
elif option["model"] == "ConvNet":
model = [
models.ConvNet(classification=10) for _ in range(option["rcm"])
]
elif option["model"] == "ConvNet_s":
model = [
models.ConvNet_s(classification=10) for _ in range(option["rcm"])
]
_save_name = _save_name + "_ft_rcm{:d}".format(option["rcm"])
if option["model"] == "LeNet5":
train_dl = [
load_mnist("train", option["rcm"], i + 1, option["batch"])
for i in range(option["rcm"])
] # ([train_dl[0], ...], [valid_dl[0], ...])
test_dl = load_mnist("test", 1, 1, option["batch"])
else:
train_dl = [
load_cifar10("train", option["rcm"], i + 1, option["batch"])
for i in range(option["rcm"])
]
test_dl = load_cifar10("test", 1, 1, option["batch"])
i = 0
for m in model:
load_model = torch.load(option["model_path"][i],
map_location=option["dev"])
m._modules = load_model['_modules']
m.load_state_dict(load_model['state_dict'])
m = m.to(option["dev"])
i = i + 1
for i in range(option["rcm"]):
print("Reduced Classification model #{:d}".format(i + 1))
loss_fn = nn.CrossEntropyLoss()
opt = optim.Adam(model[i].parameters(), lr=option["lr"])
save_name = _save_name + "_{:d}".format(i + 1)
history = fit(model[i], train_dl[i][0], train_dl[i][1], loss_fn, opt,
option["epoch"], option["schedule"], save_name,
option["save_option"])
result = distributed_evaluate(model, test_dl)
def eprune_rcm(option):
_save_name = option["save_name"]
classification = int(10 / option["rcm"]) + 1
if option["model"] == "VGG16":
model = [
models.VGG(layers=16, classification=classification)
for _ in range(option["rcm"])
]
elif option["model"] == "MobileNet":
model = [
models.MobileNet(alpha=option["alpha"],
classification=classification)
for _ in range(option["rcm"])
]
if option["alpha"] == 1.0:
_save_name = _save_name + "x1.0"
else:
_save_name = _save_name + "x{:n}".format(option["alpha"])
elif option["model"] == "LeNet5":
model = [
models.LeNet5(classification=classification)
for _ in range(option["rcm"])
]
elif option["model"] == "ConvNet":
model = [
models.ConvNet(classification=classification)
for _ in range(option["rcm"])
]
elif option["model"] == "ConvNet_s":
model = [
models.ConvNet_s(classification=classification)
for _ in range(option["rcm"])
]
_save_name = _save_name + "_ep_rcm{:d}".format(option["rcm"])
if option["model"] == "LeNet5":
train_dl = [
load_mnist("train", option["rcm"], i + 1, option["batch"])
for i in range(option["rcm"])
] # ([train_dl[0], ...], [valid_dl[0], ...])
test_dl = load_mnist("test", 1, 1, option["batch"])
else:
train_dl = [
load_cifar10("train", option["rcm"], i + 1, option["batch"])
for i in range(option["rcm"])
]
test_dl = load_cifar10("test", 1, 1, option["batch"])
i = 0
for m in model:
load_model = torch.load(option["model_path"][i],
map_location=option["dev"])
m._modules = load_model['_modules']
m.load_state_dict(load_model['state_dict'])
m = m.to(option["dev"])
i = i + 1
print("Element-wise Pruning - RCM{}".format(option["rcm"]))
for i in range(option["rcm"]):
print("Reduced Classification model #{:d}".format(i + 1))
print("\t{}".format(option["model_path"][i]))
pruning = ePruning(model[i], train_dl[i][0], train_dl[i][1],
option["threshold_ratio"])
save_name = _save_name + "_{:03d}_{:d}".format(
(option["threshold_ratio"]), i + 1)
history = pruning.iterative_pruning(finetuning=option["retraining"],
epoch=option["epoch"],
lr=option["lr"],
schedule=option["schedule"],
save_name=save_name,
save_mode=option["save_option"])
print("* Evaluating Test Sets")
result = distributed_evaluate(model, test_dl)