-
Notifications
You must be signed in to change notification settings - Fork 1
/
trainer.py
165 lines (142 loc) · 8.04 KB
/
trainer.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
import copy
import wandb
import numpy as np
import torch
from torch.utils.data import Dataset, DataLoader
import pytorch_lightning as pl
from pytorch_lightning.callbacks import LearningRateMonitor, ModelCheckpoint
from networks import _create_vision_transformer
def training(args, prototype_mode, taskid, train_loader, test_loader, known_classes=0, vitprompt=None, numclass=10):
trainer = pl.Trainer(default_root_dir='./checkpoints/'+wandb.run.name, accelerator="gpu", devices=1,
max_epochs=args.max_epochs, progress_bar_refresh_rate=1,
callbacks=[ModelCheckpoint(save_weights_only=True, mode="max", monitor="train_acc"),
LearningRateMonitor("epoch"),])
if taskid == 0:
model = prototype_mode(num_cls=numclass, lr=args.lr, max_epoch=args.max_epochs, weight_decay=0.0005,
known_classes=known_classes, freezep=False, using_prompt=args.using_prompt,
anchor_energy=args.anchor_energy, lamda=args.lamda, energy_beta=args.energy_beta)
else:
model = prototype_mode(num_cls=numclass, lr=args.lr, max_epoch=args.max_epochs, weight_decay=0.0005,
known_classes=known_classes, freezep=True, using_prompt=args.using_prompt,
anchor_energy=args.anchor_energy, lamda=args.lamda, energy_beta=args.energy_beta)
model.vitprompt = vitprompt
trainer.fit(model, train_loader)
if args.dataset == "core50":
val_result = [{'test_acc':0}]
else:
val_result = trainer.test(model, dataloaders=test_loader, verbose=False)
print(val_result)
return model, val_result
def eval(args, load_path, datamanage):
assembles = torch.load(load_path, map_location=torch.device('cpu'))
model_kwargs = dict(patch_size=16, embed_dim=768, depth=12, num_heads=12)
ptvit = _create_vision_transformer('vit_base_patch16_224', pretrained=True, **model_kwargs)
device = 'cuda:0'
ptvit = ptvit.to(device)
ptvit.eval()
all_tabs = assembles['all_tabs']
all_classifiers = assembles['all_classifiers']
all_tokens = assembles['all_tokens']
vitpromptlist = assembles['vitpromptlist']
all_tabs = [i.to(device) for i in all_tabs]
all_classifiers = [i.to(device) for i in all_classifiers]
all_tokens = [i.to(device) for i in all_tokens]
vitpromptlist = [i.to(device) for i in vitpromptlist]
_known_classes=0
# fast mode
candidata_temperatures = [0.001, 0.01, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0]
# slow mode
# candidata_temperatures = [i/1000 for i in range(1,1000,1)]
select_temperature = [0.001] # for first stage with no former data
accuracy_table, accs = [], []
print("Testing...")
for taskid in range(datamanage.nb_tasks):
_total_classes = _known_classes + datamanage.get_task_size(taskid)
test_till_now_dataset = datamanage.get_dataset(np.arange(0, _total_classes), source='test', mode='test')
test_till_now_loader = DataLoader(test_till_now_dataset, batch_size=128, shuffle=False, num_workers=8)
if taskid > 0:
classifiers = all_classifiers[:taskid+1]
task_tokens = all_tokens[:taskid+1]
tabs = all_tabs[:taskid+1]
current_dataset = datamanage.get_dataset(np.arange(_known_classes, _total_classes), source='train', mode='test')
current_dataloader = DataLoader(current_dataset, batch_size=128, shuffle=False, num_workers=8)
all_energies = {i: [] for i in candidata_temperatures}
with torch.no_grad():
for _, (_, inputs, targets) in enumerate(current_dataloader):
inputs = inputs.to(device)
targets = targets.to(device)
energys = {i: [] for i in candidata_temperatures}
image_features = ptvit(inputs, instance_tokens=vitpromptlist[taskid].weight, returnbeforepool=True)
B = image_features.shape[0]
for idx, fc in enumerate(classifiers):
task_token = task_tokens[idx].expand(B, -1, -1)
task_token, attn, v = tabs[idx](torch.cat((task_token, image_features), dim=1), mask_heads=None)
task_token = task_token[:, 0]
logit = fc(task_token)
for tem in candidata_temperatures:
energys[tem].append(torch.logsumexp(logit / tem, axis=-1))
energys = {i: torch.stack(energys[i]).T for i in candidata_temperatures}
for i in candidata_temperatures:
all_energies[i].append(energys[i])
all_energies = {i: torch.cat(all_energies[i]) for i in candidata_temperatures}
seperation_accuracy = []
for i in candidata_temperatures:
seperation_accuracy.append((sum(all_energies[i].max(1)[1]==(taskid))/len(all_energies[i])).item())
select_temperature.append(candidata_temperatures[np.array(seperation_accuracy).argmax()])
set_select_temperature = list(set(select_temperature))
y_pred, y_true = [], []
for _, (_, inputs, targets) in enumerate(test_till_now_loader):
inputs = inputs.to(device)
targets = targets.to(device)
candiatetask = {i: [] for i in set_select_temperature}
seperatePreds = []
with torch.no_grad():
image_features = ptvit(inputs, instance_tokens=vitpromptlist[taskid].weight, returnbeforepool=True)
B = image_features.shape[0]
for idx, fc in enumerate(all_classifiers[:taskid+1]):
task_token = all_tokens[:taskid+1][idx].expand(B, -1, -1)
task_token, attn, v = all_tabs[:taskid+1][idx](torch.cat((task_token, image_features), dim=1), mask_heads=None)
task_token = task_token[:, 0]
logit = fc(task_token)
for tem in set_select_temperature:
candiatetask[tem].append(torch.logsumexp(logit / tem, axis=-1))
seperatePreds.append(logit.max(1)[1]+idx*logit.shape[1])
candiatetask = {i: torch.stack(candiatetask[i]).T for i in set_select_temperature}
seperatePreds = torch.stack(seperatePreds).T
pred = []
for tem in set_select_temperature:
val, ind = candiatetask[tem].max(1)
pred.append(ind)
indexselection = torch.stack(pred, 1)
selectid = torch.mode(indexselection, dim=1, keepdim=False)[0]
outputs = []
for row, idx in enumerate(selectid):
outputs.append(seperatePreds[row][idx])
outputs = torch.stack(outputs)
y_pred.append(outputs.cpu().numpy())
y_true.append(targets.cpu().numpy())
y_pred = np.concatenate(y_pred)
y_true = np.concatenate(y_true)
if args.dil:
accs.append(np.around((y_pred.T%args.max_cls == y_true%args.max_cls).sum() * 100 / len(y_true), decimals=2))
else:
accs.append(np.around((y_pred.T == y_true).sum() * 100 / len(y_true), decimals=2))
tempacc = []
for class_id in range(0, np.max(y_true), _total_classes-_known_classes):
idxes = np.where(np.logical_and(y_true >= class_id, y_true < class_id + _total_classes-_known_classes))[0]
tempacc.append(np.around((y_pred[idxes] == y_true[idxes]).sum() * 100 / len(idxes), decimals=3))
accuracy_table.append(tempacc)
_known_classes = _total_classes
np_acctable = np.zeros([taskid + 1, taskid + 1])
for idxx, line in enumerate(accuracy_table):
idxy = len(line)
np_acctable[idxx, :idxy] = np.array(line)
# import pdb;pdb.set_trace()
np_acctable = np_acctable.T
print("Accuracy table:")
print(np_acctable)
print("Accuracy curve:")
print(accs)
print("FAA: {}".format(accs[-1]))
forgetting = np.mean((np.max(np_acctable, axis=1) - np_acctable[:, taskid])[:taskid])
print("FF: {}".format(forgetting))