def test_DML():
student_params = [4, 4, 4, 4, 4]
student_model_1 = ResNet50(student_params, 1, 10)
student_model_2 = ResNet18(student_params, 1, 10)
student_cohort = (student_model_1, student_model_2)
s_optimizer_1 = optim.SGD(student_model_1.parameters(), 0.01)
s_optimizer_2 = optim.SGD(student_model_2.parameters(), 0.01)
student_optimizers = (s_optimizer_1, s_optimizer_2)
distiller = DML(
student_cohort,
train_loader,
test_loader,
student_optimizers,
log=True,
logdir=".",
)
distiller.train_students(epochs=1,
plot_losses=True,
save_model=True,
save_model_path="./student.pt")
distiller.evaluate()
distiller.get_parameters()
del student_model_1, student_model_2, distiller, s_optimizer_1, s_optimizer_2
def test_NoisyTeacher():
teacher_params = [4, 4, 8, 4, 4]
student_params = [4, 4, 4, 4, 4]
teacher_model = ResNet50(teacher_params, 1, 10)
student_model = ResNet18(student_params, 1, 10)
t_optimizer = optim.SGD(teacher_model.parameters(), 0.01)
s_optimizer = optim.SGD(student_model.parameters(), 0.01)
experiment = NoisyTeacher(
teacher_model,
student_model,
train_loader,
test_loader,
t_optimizer,
s_optimizer,
alpha=0.4,
noise_variance=0.2,
device="cpu",
)
experiment.train_teacher(epochs=1, plot_losses=False, save_model=False)
experiment.train_student(epochs=1, plot_losses=False, save_model=False)
experiment.evaluate(teacher=False)
experiment.get_parameters()
del teacher_model, student_model, experiment, t_optimizer, s_optimizer
def test_RCO():
teacher_params = [4, 4, 8, 4, 4]
student_params = [4, 4, 4, 4, 4]
teacher_model = ResNet50(teacher_params, 1, 10)
student_model = ResNet18(student_params, 1, 10)
t_optimizer = optim.SGD(teacher_model.parameters(), 0.01)
s_optimizer = optim.SGD(student_model.parameters(), 0.01)
distiller = RCO(
teacher_model,
student_model,
train_loader,
test_loader,
t_optimizer,
s_optimizer,
)
distiller.train_teacher(epochs=1, plot_losses=False, save_model=False)
distiller.train_student(epochs=1, plot_losses=False, save_model=False)
distiller.evaluate()
distiller.get_parameters()
del teacher_model, student_model, distiller, t_optimizer, s_optimizer
def test_dynamic_quantization():
model_params = [4, 4, 8, 4, 4]
model = ResNet50(model_params, 1, 10, True)
quantizer = Dynamic_Quantizer(model, test_loader, {torch.nn.Linear})
quantized_model = quantizer.quantize()
quantizer.get_model_sizes()
quantizer.get_performance_statistics()
def test_resnet():
params = [4, 4, 8, 8, 16]
ResNet18(params)
ResNet34(params)
ResNet50(params)
ResNet101(params)
ResNet152(params)
def test_lottery_tickets():
teacher_params = [4, 4, 8, 4, 4]
teacher_model = ResNet50(teacher_params, 1, 10, True)
pruner = Lottery_Tickets_Pruner(teacher_model, train_loader, test_loader)
pruner.prune(num_iterations=0,
train_iterations=0,
valid_freq=1,
print_freq=1)
def test_BANN():
params = [4, 4, 4, 4, 4]
model = ResNet50(params, 1, 10)
optimizer = optim.SGD(model.parameters(), 0.01)
distiller = BANN(model, train_loader, test_loader, optimizer, num_gen=2)
distiller.train_student(epochs=0, plot_losses=False, save_model=False)
def test_DML():
student_params = [4, 4, 4, 4, 4]
student_model_1 = ResNet50(student_params, 1, 10)
student_model_2 = ResNet18(student_params, 1, 10)
student_cohort = (student_model_1, student_model_2)
s_optimizer_1 = optim.SGD(student_model_1.parameters(), 0.01)
s_optimizer_2 = optim.SGD(student_model_2.parameters(), 0.01)
student_optimizers = (s_optimizer_1, s_optimizer_2)
distiller = DML(student_cohort, train_loader, test_loader, student_optimizers)
distiller.train_students(epochs=1, plot_losses=False, save_model=False)
distiller.evaluate()
distiller.get_parameters()
def test_mean_teacher():
teacher_params = [4, 4, 8, 4, 4]
student_params = [4, 4, 4, 4, 4]
teacher_model = ResNet50(teacher_params, 1, 10, mean=True)
student_model = ResNet18(student_params, 1, 10, mean=True)
t_optimizer = optim.SGD(teacher_model.parameters(), 0.01)
s_optimizer = optim.SGD(student_model.parameters(), 0.01)
mt = MeanTeacher(
teacher_model,
student_model,
train_loader,
test_loader,
t_optimizer,
s_optimizer,
)
mt.train_teacher(epochs=1, plot_losses=False, save_model=False)
mt.train_student(epochs=1, plot_losses=False, save_model=False)
mt.evaluate()
mt.get_parameters()
def test_attention():
teacher_params = [4, 4, 8, 4, 4]
student_params = [4, 4, 4, 4, 4]
teacher_model = ResNet50(teacher_params, 1, 10, True)
student_model = ResNet18(student_params, 1, 10, True)
t_optimizer = optim.SGD(teacher_model.parameters(), 0.01)
s_optimizer = optim.SGD(student_model.parameters(), 0.01)
att = Attention(
teacher_model,
student_model,
train_loader,
test_loader,
t_optimizer,
s_optimizer,
)
att.train_teacher(epochs=1, plot_losses=False, save_model=False)
att.train_student(epochs=1, plot_losses=False, save_model=False)
att.evaluate(teacher=False)
att.get_parameters()