-
Notifications
You must be signed in to change notification settings - Fork 13
/
pipeline.py
494 lines (432 loc) · 20.5 KB
/
pipeline.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
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
import random
import datetime
import numpy as np
import torch
import importlib
from dgllife.model import load_pretrained
from dgllife.utils import EarlyStopping, Meter
from torch.nn import BCEWithLogitsLoss
from torch.utils.data import ConcatDataset,DataLoader
import copy
from utils import collate_molgraphs, load_model, GraphLevelDataset
import os
import errno
import pickle
joint_alias = ['joint', 'Joint', 'joint_replay_all', 'jointtrain', 'jointreplay']
def assign_hyp_param(args, params):
if args['method']=='lwf':
args['lwf_args'] = params
if args['method'] == 'bare':
args['bare_args'] = params
if args['method'] == 'gem':
args['gem_args'] = params
if args['method'] == 'ewc':
args['ewc_args'] = params
if args['method'] == 'mas':
args['mas_args'] = params
if args['method'] == 'twp':
args['twp_args'] = params
if args['method'] in joint_alias:
args['joint_args'] = params
def str2dict(s):
# accepts a str like " 'k1':v1; ...; 'km':vm ", values (v1,...,vm) can be single values or lists (for hyperparameter tuning)
output = dict()
kv_pairs = s.replace(' ','').replace("'",'').split(';')
for kv in kv_pairs:
key = kv.split(':')[0]
v_ = kv.split(':')[1]
if '[' in v_:
# transform list of values
v_list = v_.replace('[','').replace(']','').split(',')
vs=[]
for v__ in v_list:
try:
# if the parameter is float
vs.append(float(v__))
except:
# if the parameter is str
vs.append(str(v__))
output.update({key:vs})
else:
try:
output.update({key: float(v_)})
except:
output.update({key: str(v_)})
return output
def compose_hyper_params(hyp_params):
hyp_param_list = [{}]
for hk in hyp_params:
hyp_param_list_ = []
hyp_p_current = hyp_params[hk] if isinstance(hyp_params[hk],list) else [hyp_params[hk]]
for v in hyp_p_current:
for hk_ in hyp_param_list:
hk__ = copy.deepcopy(hk_)
hk__.update({hk: v})
hyp_param_list_.append(hk__)
hyp_param_list = hyp_param_list_
return hyp_param_list
def remove_illegal_characters(name, replacement='_'):
# replace any potential illegal characters with 'replacement'
for c in ['-', '[' ,']' ,'{', '}', "'", ',', ':', ' ']:
name = name.replace(c,replacement)
return name
def mkdir_if_missing(directory):
if not os.path.exists(directory):
try:
os.makedirs(directory)
except OSError as e:
if e.errno != errno.EEXIST:
raise
def predict(args, model, bg, task_i=None):
node_feats = bg.ndata.pop(args['node_data_field']).cuda()
if args.get('edge_featurizer', None) is not None:
edge_feats = bg.edata.pop(args['edge_data_field']).cuda()
if args['backbone'] in ['GCN', 'GAT', 'Weave']:
return model(bg.to(f"cuda:{args['gpu']}"), node_feats, edge_feats)
else:
return model(bg.to(f"cuda:{args['gpu']}"), node_feats, edge_feats, task_i)
else:
if args['backbone'] in ['GCN', 'GAT', 'Weave']:
return model(bg.to(f"cuda:{args['gpu']}"), node_feats)
else:
return model(bg.to(f"cuda:{args['gpu']}"), node_feats, task_i)
def run_a_train_epoch(args, epoch, model, data_loader, loss_criterion, optimizer, task_i):
model.train()
train_meter = Meter()
for batch_id, batch_data in enumerate(data_loader):
smiles, bg, labels, masks = batch_data
labels, masks = labels.cuda(), masks.cuda()
logits = predict(args, model, bg, task_i)
if isinstance(logits, tuple):
logits = logits[0]
# Mask non-existing labels
loss = loss_criterion(logits, labels) * (masks != 0).float()
loss = loss[:,task_i].mean()
optimizer.zero_grad()
loss.backward()
optimizer.step()
train_meter.update(logits, labels, masks)
def eval_single_task_multi_label(args, model, data_loader, task_i):
# only returns the performance of the given task
model.eval()
eval_meter = Meter()
with torch.no_grad():
for batch_id, batch_data in enumerate(data_loader):
smiles, bg, labels, masks = batch_data
labels = labels.cuda()
logits = predict(args, model, bg, task_i)
if isinstance(logits, tuple):
logits = logits[0]
eval_meter.update(logits, labels, masks)
return eval_meter.compute_metric(args['metric_name'])[task_i]
def eval_all_learnt_task_multi_label(args, model, data_loader, task_i):
model.eval()
eval_meter = Meter()
with torch.no_grad():
if args['classifier_increase']:
t_end = task_i + 1
else:
t_end = args['n_tasks']
for batch_id, batch_data in enumerate(data_loader):
smiles, bg, labels, masks = batch_data
labels = labels.cuda()
logits = predict(args, model, bg)
if isinstance(logits, tuple):
logits = logits[0]
eval_meter.update(logits[:,0:t_end], labels[:,0:t_end], masks)
test_score = eval_meter.compute_metric(args['metric_name'])
score_mean = round(np.mean(test_score),4)
for t in range(t_end):
score = test_score[t]
print(f"T{t:02d} {score:.4f}|", end="")
print(f"score_mean: {score_mean}", end="")
print()
return test_score
def eval_single_task_multiclass_tskIL(args, model, data_loader, task_i):
model.eval()
y_pred = []
y_true = []
with torch.no_grad():
for batch_id, batch_data in enumerate(data_loader[task_i]):
smiles, bg, labels, masks = batch_data
logits = predict(args, model, bg, task_i)
if isinstance(logits, tuple):
logits = logits[0]
logits = logits[:, args['tasks'][task_i]]
y_pred.append(logits.detach().cpu())
y_true.append(labels.detach().cpu())
y_pred = torch.cat(y_pred, dim=0)
y_pred = y_pred.argmax(-1)
y_true = torch.cat(y_true, dim=0)
for i, c in enumerate(args['tasks'][task_i]):
y_true[y_true == c] = i
ids_per_cls = [(y_true == i).nonzero().view(-1).tolist() for i in y_true.int().unique().tolist()]
acc_per_cls = [(y_pred[ids] == y_true[ids]).sum() / len(ids) for ids in ids_per_cls]
return sum(acc_per_cls).item() / len(acc_per_cls)
def eval_all_learnt_task_multiclass_tskIL(args, model, data_loaders, task_i):
model.eval()
acc_learnt_tsk = np.zeros(args['n_tasks'])
with torch.no_grad():
t_end = task_i + 1
#t_end = args['n_tasks']
for tid, data_loader in enumerate(data_loaders):
y_pred = []
y_true = []
for batch_id, batch_data in enumerate(data_loader):
smiles, bg, labels, masks = batch_data
logits = predict(args, model, bg, task_i)
if isinstance(logits, tuple):
logits = logits[0]
logits = logits[:, args['tasks'][tid]]
y_pred.append(logits.detach().cpu())
y_true.append(labels.detach().cpu())
y_pred = torch.cat(y_pred, dim=0)
y_pred = y_pred.argmax(-1)
y_true = torch.cat(y_true, dim=0)
for i, c in enumerate(args['tasks'][tid]):
y_true[y_true == c] = i
ids_per_cls = [(y_true == i).nonzero().view(-1).tolist() for i in y_true.int().unique().tolist()]
acc_per_cls = [(y_pred[ids] == y_true[ids]).sum() / len(ids) for ids in ids_per_cls]
acc_learnt_tsk[tid] = sum(acc_per_cls).item() / len(acc_per_cls)
score_mean = round(np.mean(acc_learnt_tsk[0:t_end]), 4)
for t in range(t_end):
score = acc_learnt_tsk[t]
print(f"T{t:02d} {score:.4f}|", end="")
print(f"score_mean: {score_mean}", end="")
print()
return acc_learnt_tsk
def eval_single_task_multiclass_clsIL(args, model, data_loader, task_i):
# calculate the AP of tasks learnt so far
model.eval()
y_pred = []
y_true = []
learnt_cls = []
for tid in range(task_i+1):
learnt_cls.extend(args['tasks'][tid])
with torch.no_grad():
for batch_id, batch_data in enumerate(data_loader[task_i]):
smiles, bg, labels, masks = batch_data
logits = predict(args, model, bg, task_i)
if isinstance(logits, tuple):
logits = logits[0]
logits = logits[:, learnt_cls]
y_pred.append(logits.detach().cpu())
y_true.append(labels.detach().cpu())
y_pred = torch.cat(y_pred, dim=0)
y_pred = y_pred.argmax(-1)
y_true = torch.cat(y_true, dim=0)
for i, c in enumerate(learnt_cls):
y_true[y_true == c] = i
ids_per_cls = [(y_true == i).nonzero().view(-1).tolist() for i in y_true.int().unique().tolist()]
acc_per_cls = [(y_pred[ids] == y_true[ids]).sum() / len(ids) for ids in ids_per_cls]
return sum(acc_per_cls).item() / len(acc_per_cls)
def eval_all_learnt_task_multiclass_clsIL(args, model, data_loaders, task_i):
model.eval()
acc_learnt_tsk = np.zeros(args['n_tasks'])
learnt_cls,allclss = [],[]
for tid in range(task_i + 1):
learnt_cls.extend(args['tasks'][tid])
for tid in args['tasks']:
allclss.extend(tid)
with torch.no_grad():
#selected_clss = allclss if args['method'] is 'jointtrain' else learnt_cls
#t_end = task_i + 1 if args['method'] is not 'jointtrain' else args['n_tasks']
selected_clss = learnt_cls
t_end = task_i + 1
for tid, data_loader in enumerate(data_loaders[0:t_end]):
y_pred,y_true = [],[]
for batch_id, batch_data in enumerate(data_loader):
smiles, bg, labels, masks = batch_data
logits = predict(args, model, bg, task_i)
if isinstance(logits, tuple):
logits = logits[0]
logits = logits[:, selected_clss]
y_pred.append(logits.detach().cpu())
y_true.append(labels.detach().cpu())
y_pred = torch.cat(y_pred, dim=0)
y_pred = y_pred.argmax(-1)
y_true = torch.cat(y_true, dim=0)
for i, c in enumerate(selected_clss):
y_true[y_true == c] = i
ids_per_cls = {i:(y_true == i).nonzero().view(-1).tolist() for i in y_true.int().unique().tolist()}
acc_per_cls = [(y_pred[ids_per_cls[ids]] == y_true[ids_per_cls[ids]]).sum().item() / len(ids_per_cls[ids]) for ids in ids_per_cls]
acc_learnt_tsk[tid] = sum(acc_per_cls) / len(acc_per_cls)
score_mean = round(np.mean(acc_learnt_tsk[0:t_end]), 4)
for t in range(t_end):
score = acc_learnt_tsk[t]
print(f"T{t:02d} {score:.4f}|", end="")
print(f"score_mean: {score_mean}", end="")
print()
return acc_learnt_tsk
def get_pipeline(args):
if args['dataset'] in ['SIDER-tIL', 'Tox21-tIL']:
# for multi-label datasets
return pipeline_multi_label
else:
return pipeline_multi_class
def pipeline_multi_label(args, valid=False):
'''
:param args: arguments specified in train.py
:valid: whether to use validation set or testing set to evaluate the model. If True, the model is in training mode. If false, the model is in testing mode, and the training epochs will be 0
'''
epochs = args['num_epochs'] if valid else 0
torch.cuda.set_device(args['gpu'])
# set_random_seed(args['random_seed'])
G = GraphLevelDataset(args)
dataset, train_set, val_set, test_set = G.dataset, G.train_set, G.val_set, G.test_set
args['n_cls'] = dataset.labels.shape[1]
args['tasks'] = [list(range(i, i + args['n_cls_per_task'])) for i in
range(0, args['n_cls'], args['n_cls_per_task'])]
train_loader = DataLoader(train_set, batch_size=args['batch_size'],
collate_fn=collate_molgraphs, shuffle=True)
val_loader = DataLoader(val_set, batch_size=args['batch_size'],
collate_fn=collate_molgraphs)
test_loader_ = DataLoader(test_set, batch_size=args['batch_size'],
collate_fn=collate_molgraphs)
args['d_data'] = dataset.graphs[0].ndata['h'].shape[1]
test_loader = val_loader if valid else test_loader_
if args['pre_trained']:
args['num_epochs'] = 0
model = load_pretrained(args['exp'])
else:
args['n_tasks'] = len(args['tasks'])
args['n_outheads'] = args['n_tasks'] # for multi-label tasks, #output heads equals #tasks
model = load_model(args)
life_model = importlib.import_module(f"Baselines.{args['method']}_model")
life_model_ins = life_model.NET(model, args) if valid else None
data_loader = DataLoader(train_set, batch_size=len(train_set),collate_fn=collate_molgraphs, shuffle=True)
if life_model_ins is not None:
life_model_ins.data_loader = data_loader
loss_criterion = BCEWithLogitsLoss(pos_weight=dataset.task_pos_weights(train_set.indices).cuda(),
reduction='none')
model.cuda(args['gpu'])
#score_mean = []
score_matrix = np.zeros([args['n_tasks'], args['n_tasks']])
prev_model = None
val_func, test_func = eval_single_task_multi_label, eval_all_learnt_task_multi_label
for tid, task_i in enumerate(args['tasks']):
if args['method'] == 'jointtrain' and life_model_ins is not None:
if args['joint_args']['reset_param']:
# reset the model for joint train
model = load_model(args)
model.cuda(args['gpu'])
life_model_ins.change_model(model, args)
name, ite = args['current_model_save_path']
config_name = name.split('/')[-1]
subfolder_c = name.split(config_name)[-2]
save_model_name = f'{config_name}_{ite}_{task_i}'
save_model_path = f"{args['result_path']}/{subfolder_c}val_models/{save_model_name}.pkl"
print('\n********' + str([tid, task_i]))
dt = datetime.datetime.now()
mkdir_if_missing(f"{args['result_path']}/early_stop")
stopper = EarlyStopping(patience=args['patience'],filename='{}/early_stop/{}_{:02d}-{:02d}-{:02d}-{}.pth'.format(args['result_path'],
dt.date(), dt.hour, dt.minute, dt.second,str(random.randint(100,999))))
for epoch in range(epochs):
# Train
if args['method'] == 'lwf':
life_model_ins.observe(train_loader, loss_criterion, tid, args, prev_model)
else:
life_model_ins.observe(train_loader, loss_criterion, tid, args)
# Validation and early stop
val_score = val_func(args, model, val_loader, tid)
early_stop = stopper.step(val_score, model)
if early_stop and args['early_stop']:
print(epoch)
break
if not args['pre_trained'] and valid and args['early_stop']:
stopper.load_checkpoint(model)
if not valid:
model = pickle.load(open(save_model_path,'rb')).cuda(args['gpu'])
score_matrix[tid] = test_func(args, model, test_loader, tid)
if valid:
mkdir_if_missing(f"{args['result_path']}/{subfolder_c}/val_models")
with open(save_model_path, 'wb') as f:
pickle.dump(model, f)
if args['method'] == 'lwf':
prev_model = copy.deepcopy(life_model_ins).cuda(args['gpu']) if valid else None
AP = round(np.mean(score_matrix[-1, :]), 4)
print('AP: ', round(np.mean(score_matrix[-1, :]), 4))
backward = []
for t in range(args['n_tasks'] - 1):
b = score_matrix[args['n_tasks'] - 1][t] - score_matrix[t][t]
backward.append(round(b, 4))
mean_backward = round(np.mean(backward), 4)
print('AF: ', mean_backward)
return AP, mean_backward, score_matrix
def pipeline_multi_class(args, valid=False):
epochs = args['num_epochs'] if valid else 0
torch.cuda.set_device(args['gpu'])
G = GraphLevelDataset(args)
dataset, train_set, val_set, test_set = G.dataset, G.train_set, G.val_set, G.test_set # cls-IL and tsk-IL have different test_set, and different test_func
train_set_joint = [ConcatDataset(train_set[0:i]) for i in range(1,len(train_set)+1)] # for tsk_IL only, meaningless when cls-IL since train_set already concats learnt tasks when cls-IL
coll_f = collate_molgraphs
args['n_cls'] = dataset.labels.max().int().item() + 1
args['n_outheads'] = args['n_cls']
train_loader = [DataLoader(s, batch_size=args['batch_size'], collate_fn=coll_f, shuffle=True) for s in train_set] # train_set, a list of data for each task
train_loader_joint = [DataLoader(s, batch_size=args['batch_size'], collate_fn=coll_f, shuffle=True) for s in train_set_joint]
val_loader = [DataLoader(s, batch_size=args['batch_size'], collate_fn=coll_f) for s in val_set]
test_loader_ = [DataLoader(s, batch_size=args['batch_size'],collate_fn=coll_f) for s in test_set]
test_loader = val_loader if valid else test_loader_
args['d_data'] = dataset.graphs[0].ndata['h'].shape[1]
if args['pre_trained']:
args['num_epochs'] = 0
model = load_pretrained(args['exp'])
else:
args['n_tasks'] = len(args['tasks']) # dataset.n_tasks
model = load_model(args)
life_model = importlib.import_module(f"Baselines.{args['method']}_model")
life_model_ins = life_model.NET(model, args) if valid else None
dataloader_shuffle = False if args['method'] is 'gem' else True # to ensure indices stored by gem always refer to same data in the dataloader
data_loader = [DataLoader(s, batch_size=len(s), collate_fn=coll_f, shuffle=dataloader_shuffle) for s in train_set]
if life_model_ins is not None:
life_model_ins.data_loader = data_loader
loss_criterion = torch.nn.functional.cross_entropy
model.cuda(args['gpu'])
score_matrix = np.zeros([args['n_tasks'], args['n_tasks']])
prev_model = None
if args['clsIL']:
train_func = life_model_ins.observe_clsIL if valid else None
val_func, test_func = eval_single_task_multiclass_clsIL, eval_all_learnt_task_multiclass_clsIL
elif not args['clsIL']:
train_func = life_model_ins.observe_tskIL_multicls if valid else None
val_func, test_func = eval_single_task_multiclass_tskIL, eval_all_learnt_task_multiclass_tskIL
for tid, task_i in enumerate(args['tasks']):
if args['method'] == 'jointtrain' and life_model_ins is not None:
if args['joint_args']['reset_param']:
# reset the model for joint train
model = load_model(args)
model.cuda(args['gpu'])
life_model_ins.change_model(model, args)
name, ite = args['current_model_save_path']
config_name = name.split('/')[-1]
subfolder_c = name.split(config_name)[-2]
save_model_name = f'{config_name}_{ite}_{task_i}'
save_model_path = f"{args['result_path']}/{subfolder_c}val_models/{save_model_name}.pkl"
print('\n********' + str([tid, task_i]),{i:args['n_per_cls'][i] for i in task_i})
for epoch in range(epochs):
# Train
if args['method'] == 'lwf':
train_func(train_loader, loss_criterion, tid, args, prev_model)
elif args['method'] == 'jointtrain' and args['clsIL'] == False:
train_func(train_loader_joint, loss_criterion, tid, args)
else:
train_func(train_loader, loss_criterion, tid, args)
if not valid:
# if testing, load the trained model
model = pickle.load(open(save_model_path,'rb')).cuda(args['gpu'])
score_matrix[tid] = test_func(args, model, test_loader, tid)
if valid:
mkdir_if_missing(f"{args['result_path']}/{subfolder_c}/val_models")
with open(save_model_path, 'wb') as f:
pickle.dump(model, f)
if args['method'] == 'lwf':
prev_model = copy.deepcopy(life_model_ins).cuda(args['gpu']) if valid else None
AP = round(np.mean(score_matrix[-1, :]), 4)
print('AP: ', AP)
backward = []
for t in range(args['n_tasks'] - 1):
b = score_matrix[args['n_tasks'] - 1][t] - score_matrix[t][t]
backward.append(round(b, 4))
mean_backward = round(np.mean(backward), 4)
print('AF: ', mean_backward)
return AP, mean_backward, score_matrix