-
Notifications
You must be signed in to change notification settings - Fork 1
/
evaluate_nmn_utils.py
280 lines (191 loc) · 8.35 KB
/
evaluate_nmn_utils.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
import utils
utils.set_seed(11)
import evaluate
import numpy as np
from io import StringIO
device = evaluate.device
import torch, json, copy, time, math, sys
import torch.optim as optim
import torch.nn.functional as F
import network, train, utils, dataset
import numpy as np
from spikingjelly.clock_driven import functional
config_file_path = "config.json"
with open(config_file_path) as f:
config = json.load(f)
batch_size = config["batch_size"]
n = config["n"]
nf = config["nf"]
ks = config["ks"]
n_fc = config["n_fc"]
T = config["T"]
lr = config["lr"]
loader_0 = evaluate.loader
loader_test_0 = evaluate.loader_test
net_init = network.SNN().to(device)
net_init.load_state_dict(torch.load("net_pretrain.pth"))
for param in net_init.conv.parameters():
param.requires_grad = False
for param in net_init.fc.parameters():
param.requires_grad = False
nm = utils.vect2net(np.load("x_opt.npy")).to(device)
def ev(seed, nm, perm_i, tasks, n=n, batch_size=batch_size, lr=lr, use_nm = True, device = device, loader = loader_0, loader_test = loader_test_0, net_init = net_init):
# We set all random seeds inside this method, so that we are guaranteed to have the same data generation etc (for every scenario)
utils.set_seed(seed)
n_tasks = len(tasks)
standard_stdout = sys.stdout
outBuffer = StringIO()
sys.stdout = outBuffer
d = [dataset.dataset_prepare_fewshot([x], n, train=True) for x in tasks]
loader = [torch.utils.data.DataLoader(d[i], batch_size, shuffle=False) for i in range(n_tasks)]
d_test = [dataset.dataset_prepare_fewshot([x], n, train=False) for x in tasks]
loader_test = [torch.utils.data.DataLoader(d_test[i], batch_size, shuffle=False) for i in range(n_tasks)]
sys.stdout = standard_stdout
net = copy.deepcopy(net_init)
if use_nm:
net.nm = copy.deepcopy(nm)
for param in net.nm.parameters():
param.requires_grad = False
accs = []
accs_batches = [[] for i in range(n_tasks)]
for i in range(perm_i, perm_i+n_tasks):
with torch.no_grad():
net.output[0].weight[tasks[i%n_tasks]] = copy.deepcopy(net_init.output[0].weight[tasks[i%n_tasks]])
opt = optim.SGD(net.parameters(), lr = lr)
net.train()
correct_pred = 0
for ii, inp in enumerate(loader[i%n_tasks]):
x, label = inp[0].float().to(device), inp[1].to(device)
opt.zero_grad()
y = net(x, use_nm = use_nm)
label = F.one_hot(label, 11).float()
loss = F.mse_loss(y, label)
correct_pred += (y.argmax(dim=1) == label.argmax(dim=1)).sum().item()
loss.backward()
opt.step()
functional.reset_net(net)
if use_nm: functional.reset_net(net.nm)
#for j in range(perm_i, i+1):
# accs_batches[j - perm_i].append(train.get_acc(net, loader_test[j%n_tasks], device, use_nm))
for j in range(perm_i, i+1):
accs.append(train.get_acc(net, loader_test[j%n_tasks], device, use_nm))
return np.array(accs), accs_batches
def ev_ewc(ewc_lambda, seed, nm, perm_i, tasks, n=n, batch_size=batch_size, lr=lr, device = device, loader = loader_0, loader_test = loader_test_0, net_init = net_init):
utils.set_seed(0)
n_tasks = len(tasks)
standard_stdout = sys.stdout
outBuffer = StringIO()
sys.stdout = outBuffer
d = [dataset.dataset_prepare_fewshot([x], n, train=True) for x in tasks]
loader = [torch.utils.data.DataLoader(d[i], batch_size, shuffle=False) for i in range(n_tasks)]
d_test = [dataset.dataset_prepare_fewshot([x], n, train=False) for x in tasks]
loader_test = [torch.utils.data.DataLoader(d_test[i], batch_size, shuffle=False) for i in range(n_tasks)]
sys.stdout = standard_stdout
net = copy.deepcopy(net_init)
accs = []
accs_batches = [[] for i in range(n_tasks)]
fisher_dict = {}
optpar_dict = {}
def on_task_update(task_id,opt):
net.train()
opt.zero_grad()
# accumulating gradients
for ii, inp in enumerate(loader[task_id%n_tasks]):
x, label = inp[0].float().to(device), inp[1].to(device)
y = net(x, use_nm = False)
label = F.one_hot(label, 11).float()
loss = F.mse_loss(y, label)
loss.backward()
functional.reset_net(net)
fisher_dict[task_id] = {}
optpar_dict[task_id] = {}
# gradients accumulated can be used to calculate fisher
for name, param in net.named_parameters():
if param.requires_grad:
optpar_dict[task_id][name] = param.data.clone()
fisher_dict[task_id][name] = param.grad.data.clone().pow(2)
def train_ewc(task_id, opt):
net.train()
for ii, inp in enumerate(loader[task_id%n_tasks]):
x, label = inp[0].float().to(device), inp[1].to(device)
opt.zero_grad()
y = net(x, use_nm = False)
label = F.one_hot(label, 11).float()
loss = F.mse_loss(y, label)
for task in range(perm_i, task_id):
for name, param in net.named_parameters():
if param.requires_grad:
fisher = fisher_dict[task][name]
optpar = optpar_dict[task][name]
loss += (fisher * (optpar - param).pow(2)).sum() * ewc_lambda
loss.backward()
opt.step()
functional.reset_net(net)
for i in range(perm_i, perm_i+n_tasks):
with torch.no_grad():
net.output[0].weight[tasks[i%n_tasks]] = copy.deepcopy(net_init.output[0].weight[tasks[i%n_tasks]])
opt = optim.SGD(net.parameters(), lr = lr)
train_ewc(i, opt)
on_task_update(i, opt)
for j in range(perm_i, i+1):
accs.append(train.get_acc(net, loader_test[j%n_tasks], device, False))
'''
net.eval()
for j in range(perm_i, i+1):
correct_pred = 0
for ii, inp in enumerate(loader_test[j%n_tasks]):
x, label = inp[0].float().to(device), inp[1].to(device)
y = net(x, use_nm = use_nm)
label = F.one_hot(label, 11).float()
correct_pred += (y.argmax(dim=1) == label.argmax(dim=1)).sum().item()
functional.reset_net(net)
if use_nm: functional.reset_net(net.nm)
acc = 100. * correct_pred / len(loader_test[j%n_tasks].dataset)
accs.append(acc)
'''
return np.array(accs), accs_batches
def test_nmn(seeds, tasks, n, batch_size, use_nm=True, ewc = 0, nm=nm, lr=7e-2, show_perm_details = True, return_acc=False):
acc = []
accs_mean_overall = 0
all_accs_nonm = []
n_tasks = len(tasks)
acc_batches_mean_nonm = [0 for i in range(n_tasks)]
cnt_accs_nonm = {x:0 for x in range(0, 100, 5)}
for perm_id in range(0,n_tasks):
accs_mean = 0
for seed in seeds:
if ewc == 0:
acc_curr, acc_batches = ev(seed, nm, perm_id, tasks=tasks, n=n, batch_size=batch_size, lr = lr, use_nm=use_nm, net_init=net_init)
else:
acc_curr, acc_batches = ev_ewc(ewc, seed, nm, perm_id, tasks=tasks, n=n, batch_size=batch_size, lr = lr, net_init=net_init)
for th in range(0, 100, 5):
for accc in acc_curr:
if accc <= th:
cnt_accs_nonm[th] += 1
accs_mean += acc_curr
accs_mean_overall += acc_curr
all_accs_nonm.append(acc_curr)
for i in range(n_tasks):
acc_batches_mean_nonm[i] += np.array(acc_batches[i])
if show_perm_details:
cnt = 0
for i in range(0, n_tasks):
for j in range(0, i+1):
print(f"{accs_mean[cnt]/len(seeds):6.2f} | ", end='')
cnt += 1
print()
acc.append((accs_mean/len(seeds)).mean())
print(f"\nPermutation {perm_id} mean ACC = {(accs_mean/len(seeds)).mean()}")
print(f"=====================================================\n")
for i in range(n_tasks):
acc_batches_mean_nonm[i] /= (len(seeds)*n_tasks)
cnt = 0
print("\n===============================================================================\n")
for i in range(0, n_tasks):
for j in range(0, i+1):
print(f"{accs_mean_overall[cnt]/(len(seeds)*n_tasks):6.2f} | ", end='')
cnt += 1
print()
acc.append((accs_mean_overall/(len(seeds)*n_tasks)).mean())
print("\n==> Total ACC mean = ", (accs_mean_overall/(len(seeds)*n_tasks)).mean())
if return_acc: return (accs_mean_overall/(len(seeds)*n_tasks)).mean()