-
Notifications
You must be signed in to change notification settings - Fork 1
/
lambda_search.py
218 lines (142 loc) · 8.21 KB
/
lambda_search.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
from pathlib import Path
import pickle
import torch
from dataset_graph import Dataset, load_processed_dataset
from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score
from configs import EncoderConfig, SearchParams, Type3Config, Type4Config, Type12Config, TypeInput, Type4Input
from graph_models import Type3, Type12, Type4
from trainers import TypeInput, TypeTrainer, Type4Trainer
from utils_funcs import get_A_s, get_variables,get_loaders,set_seeds,results_dict
from anndata import AnnData
import scanpy as sc
import seaborn as sns
import matplotlib.pyplot as plt
from sklearn.metrics import confusion_matrix
import os
import gc
import numpy as np
# Set the device
device = torch.device('cuda:1' if torch.cuda.is_available() else 'cpu')
def run_type12(dataset_name: str, model_type: str, path: str, params: SearchParams):
data = load_processed_dataset(dataset_name)
n_class = data.y.cpu().view(-1).unique().shape[0]
print(f"Number of unique classes is {n_class}")
x, _, fan_in, _ = get_variables(model_type, path, data)
config = Type12Config(fan_in=fan_in, fan_mid=params.fan_mid, fan_out=n_class, dropout=params.gcn_p)
model = Type12(config).to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=params.gcn_lr, weight_decay=params.wd)
t_input = TypeInput(x, get_A_s(data, path), data.y.to(device), data.train_ids, data.test_ids, data.valid_ids)
trainer = TypeTrainer(model, optimizer, t_input)
trainer.pipeline(params.max_epochs, params.patience)
return trainer
def run_type3(dataset_name: str, model_type: str, path: str, params: SearchParams):
data = load_processed_dataset(dataset_name)
n_class = data.y.cpu().view(-1).unique().shape[0]
print(f"Number of unique classes is {n_class}")
x, cls_logit, fan_in, _ = get_variables(model_type, path, data)
gcn_config = Type12Config(fan_in=fan_in, fan_mid=params.fan_mid, fan_out=n_class, dropout=params.gcn_p)
config = Type3Config(type12_config=gcn_config, cls_logit=cls_logit, lmbd=params.lmbd)
model = Type3(config).to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=params.gcn_lr, weight_decay=params.wd)
t_input = TypeInput(x, get_A_s(data, path), data.y.to(device), data.train_ids, data.test_ids, data.valid_ids)
trainer = TypeTrainer(model, optimizer, t_input)
trainer.pipeline(params.max_epochs, params.patience)
return trainer
def run_type4(dataset_name: str, model_type: str, path: str, params: SearchParams):
data = load_processed_dataset(dataset_name)
n_class = data.y.cpu().view(-1).unique().shape[0]
print(f"Number of unique classes is {n_class}")
x, cls_logits , fan_in, update_cls = get_variables(model_type, path, data)
gcn_config = Type12Config(fan_in=fan_in, fan_mid=params.fan_mid, fan_out=n_class, dropout=params.gcn_p)
encoder_config = EncoderConfig(model_name="scgpt", dataset_name=dataset_name, n_class=n_class, CLS=True, dropout=params.encoder_p)
config = Type4Config(type12_config=gcn_config, encoder_config=encoder_config, lmbd=params.lmbd,batch_size=params.batch_size)
model = Type4(config).to(device)
loaders= get_loaders(dataset_name, config.batch_size)
optimizer = torch.optim.Adam(
[
{"params": model.encoder.parameters(), "lr": params.encoder_lr},
{"params": model.gcn.parameters(), "lr": params.gcn_lr},
],
weight_decay=params.wd,
)
t_input = Type4Input(x=x, A_s=get_A_s(data, path), loaders=loaders, train_ids=data.train_ids, valid_ids=data.valid_ids, test_ids=data.test_ids,y=data.y)
trainer = Type4Trainer(model, optimizer, t_input, update_cls=update_cls)
trainer.pipeline(params.max_epochs, params.patience)
return trainer
def run_type(dataset_name: str, model_type: str, path: str, params: SearchParams):
if model_type == "type1" or model_type == "type2":
return run_type12(dataset_name, model_type, path, params)
elif model_type == "type3":
return run_type3(dataset_name, model_type, path, params)
elif model_type == "type4":
return run_type4(dataset_name, model_type, path, params)
else:
raise ValueError("Undefined type!")
if __name__=="__main__":
cur_dir= os. getcwd()
save_dir= cur_dir+"/scgnn_lambda"
if not os.path.exists(path=save_dir):
save_dir= os.path.join(cur_dir, 'scgnn_lambda')
os.makedirs(save_dir)
print(f"Created directory: {save_dir}.")
else:
print(f"{save_dir} is already created.")
dataset_name="myeloid"
type_name="type3"
path_list= ["GG-CG","GC-CG","CG-CC","CC-CC"]
#### Take results from the save transformer model
file_path = os.path.join(f"/auto/k2/aykut3/scgpt/scGPT/scgpt_gcn/save_scgcn/scgpt_{dataset_name}_median/results.pkl")
with open(file_path, "rb") as file:
results= pickle.load(file)
seed_list=results["seed_numbers"]
lambda_list= [1.0,0.9,0.8,0.7,0.6,0.5,0.4,0.3,0.2,0.1,0.0] #0.7 is by default performed in earlier experiments
for l in lambda_list:
params = SearchParams(
fan_mid=256,
lmbd=l,
gcn_lr=0.0001,
gcn_p=0.2,
wd=0.00005,
patience=10,
max_epochs=3000,
encoder_lr=0.00001,
encoder_p=0.2, batch_size=16
)
##### CREATE DICTIONARY TO SAVE RESULTS
for pt in path_list:
print(pt)
d = results_dict()
d["path"].append(pt)
d["type"].append(type_name)
d["dataset"].append(dataset_name)
for i, seed in enumerate(seed_list):
if seed==15:
seed=0
set_seeds(seed)
trainer=run_type(dataset_name= dataset_name, model_type=type_name, path=pt, params=params)
d["test_acc"].append(100*trainer.metrics["test"]["acc"])
d["test_f1"].append(100*trainer.metrics["test"]["macro"])
d["test_precision"].append(100*trainer.metrics["test"]["precision"])
d["test_recall"].append(100*trainer.metrics["test"]["recall"])
d["avg_epoch_time"].append(trainer.avg_epoch_time)
d["test_preds"].append(trainer.y_test_preds) # these are numpyed values
d["test_true"].append(trainer.y_test_true) # these are numpyed values
# These chek and save block can be slided in left to not to save for each seed iteration, last one can be saved normally.
# However, it can stay as it is now.
load_dir= save_dir+"/"+f"{dataset_name}/{type_name}/lamda_{str(l)}/{pt}/"
if not os.path.exists(path=load_dir):
load_dir= os.path.join(cur_dir, load_dir)
os.makedirs(load_dir)
print(f"Created directory: {load_dir}.")
else:
print(f"{load_dir} is already created.")
result_dir= os.path.join(load_dir, f"dname_{dataset_name}_path_[{pt}]_type_{type_name}_seedid_{str(i)}_seed_{seed}")
# Save dictionary using pickle
with open(result_dir, 'wb') as pickle_file:
pickle.dump(d, pickle_file)
equal = np.array_equal(results["labels"],trainer.y_test_true)
assert equal
# Free up memory
del trainer
torch.cuda.empty_cache()
gc.collect()