def run_10_times_CAKT():
list_best_auc = []
H = 15
data_source = "synthetic"
knowledge_length = 50
for i in range(10):
best_test_auc = CAKT_main(
model_name="CAKT",
env='CAKT',
data_source=data_source,
k_frames=4,
batch_size=80,
num_layers=1,
input_dim=2 * knowledge_length,
H=H,
embed_dim=H * H,
output_dim=knowledge_length,
weight_decay=1e-5,
max_epoch=30,
# lr=0.001,
# lr_decay=0.5,
# decay_every_epoch=10,
cv_times=1,
plot_every_iter=5,
vis=False,
issave=False)
list_best_auc.append(best_test_auc)
print(list_best_auc)
def run_one_time_CAKT():
H = 15
knowledge_length = 110
best_test_auc = CAKT_main(
model_name="CAKT",
env='CAKT',
data_source="assist2009",
k_frames=4,
batch_size=80,
num_layers=1,
input_dim=2 * knowledge_length,
H=H,
embed_dim=H * H,
output_dim=knowledge_length,
weight_decay=1e-5,
max_epoch=30,
# lr=0.001,
# lr_decay=0.5,
# decay_every_epoch=10,
cv_times=1,
plot_every_iter=5,
vis=False,
issave=False)
print(best_test_auc)
def run_sensi_H():
'''
要控制H的话要同时控制三个参数：H, embed_dim = hidden_dim = H*H
'''
list_datasets = [
# ("assist2009", 110),
# ("assist2015", 100),
# ("assist2017", 102),
("synthetic", 50),
("statics", 1223)
]
k_frames = 4
batch_size = 64
params_H = [11, 13, 15, 17, 19]
results = {}
for data_source, knowledge_length in list_datasets:
for H in params_H[:3]:
print("==============================")
print(H, data_source, knowledge_length)
# run CAKT_WWW
best_test_auc = CAKT_main(model_name="CAKT",
env='CAKT',
data_source=data_source,
k_frames=k_frames,
batch_size=batch_size,
num_layers=1,
input_dim=2 * knowledge_length,
H=H,
embed_dim=H * H,
hidden_dim=H * H,
output_dim=knowledge_length,
weight_decay=1e-5,
max_epoch=20,
cv_times=1,
plot_every_iter=5,
vis=False,
issave=False)
##########################
results["H={}_{}".format(H, data_source)] = best_test_auc
print(results)
print(results)
def run_ablation():
list_datasets = [("assist2009", 110), ("assist2015", 100),
("assist2017", 102), ("synthetic", 50), ("statics", 1223)]
k_frames = 4
batch_size = 80 * torch.cuda.device_count() if torch.cuda.device_count(
) > 1 else 80
H = 15
# ablation_list = ["LSTM_RECENT", "FC_POOLING", "FC_REAR", "WEIGHT_SUM"]
ablation_list = ["NO_EXP_DECAY"]
results = {}
for data_source, knowledge_length in list_datasets:
for ablation_option in ablation_list:
print("=================================")
print(ablation_option, data_source, knowledge_length)
# run CAKT_ablation
best_test_auc = CAKT_main(model_name="CAKT_ablation",
env='CAKT',
ablation=ablation_option,
data_source=data_source,
k_frames=k_frames,
batch_size=batch_size,
num_layers=1,
input_dim=2 * knowledge_length,
H=H,
embed_dim=H * H,
output_dim=knowledge_length,
weight_decay=1e-5,
max_epoch=30,
cv_times=1,
plot_every_iter=5,
vis=False,
issave=False)
##########################
results["ablation={}_{}".format(ablation_option,
data_source)] = best_test_auc
print(results)
print(results)
def run_sensi_b():
list_datasets = [
# ("assist2009", 110),
# ("assist2015", 100),
("assist2017", 102),
("synthetic", 50),
("statics", 1223)
]
k_frames = 4
params_batch_size = [8, 16, 32, 48, 64, 80, 96]
H = 15
results = {}
for data_source, knowledge_length in list_datasets[::-1]:
for b in params_batch_size[:3]:
print("==============================")
print(b, data_source, knowledge_length)
# run CAKT_WWW
best_test_auc = CAKT_main(model_name="CAKT",
env='CAKT',
data_source=data_source,
k_frames=k_frames,
batch_size=b * torch.cuda.device_count(),
num_layers=1,
input_dim=2 * knowledge_length,
H=H,
embed_dim=H * H,
hidden_dim=H * H,
output_dim=knowledge_length,
weight_decay=1e-5,
max_epoch=20,
cv_times=1,
plot_every_iter=5,
vis=False,
issave=False)
##########################
results["b={}_{}".format(b, data_source)] = best_test_auc
print(results)
print(results)
def run_sensi_k():
list_datasets = [
# ("assist2009", 110),
# ("assist2015", 100),
("assist2017", 102),
# ("synthetic", 50),
# ("statics", 1223)
]
params_k_frames = [2, 4, 6, 8, 10, 12, 14, 16]
batch_size = 64
H = 15
results = {}
for data_source, knowledge_length in list_datasets:
for k in params_k_frames[5:]:
print("==============================")
print(k, data_source, knowledge_length)
# run CAKT_WWW
best_test_auc = CAKT_main(model_name="CAKT",
env='CAKT',
data_source=data_source,
k_frames=k,
batch_size=batch_size *
torch.cuda.device_count() // 2,
num_layers=1,
input_dim=2 * knowledge_length,
H=H,
embed_dim=H * H,
output_dim=knowledge_length,
weight_decay=1e-5,
max_epoch=20,
cv_times=1,
plot_every_iter=5,
vis=False,
issave=False)
##########################
results["k={}_{}".format(k, data_source)] = best_test_auc
print(results)
print(results)
def run_params_combination():
list_datasets = [("assist2009", 110), ("assist2015", 100),
("assist2017", 102), ("synthetic", 50), ("statics", 1223)]
params_k_frames = [4, 8, 16, 32]
params_batch_size = [80, 64, 48, 32]
H = 15
results = {}
for data_source, knowledge_length in list_datasets:
for k, b in zip(params_k_frames, params_batch_size):
# if data_source == "assist2015" and (k == 4 or k == 8 or k == 16):
# continue
print("=================================")
print(k, b, data_source, knowledge_length)
# run CAKT_WWW
best_test_auc = CAKT_main(model_name="CAKT",
env='CAKT',
data_source=data_source,
k_frames=k,
batch_size=b,
num_layers=1,
input_dim=2 * knowledge_length,
H=H,
embed_dim=H * H,
output_dim=knowledge_length,
weight_decay=1e-5,
max_epoch=30,
cv_times=1,
plot_every_iter=5,
vis=False,
issave=False)
##########################
results["k={}_b={}_{}".format(k, b, data_source)] = best_test_auc
print(results)
print(results)