-
Notifications
You must be signed in to change notification settings - Fork 0
/
generate_SHD_data.py
275 lines (180 loc) · 8.69 KB
/
generate_SHD_data.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
# In[1]:
import os
os.environ['TF_CUDNN_DETERMINISTIC']='1'
import jax
import jax.numpy as jnp
import optax
from OTPE import OSTL, OTTT, OTPE, Approx_OTPE
from jax.tree_util import Partial, tree_map, tree_leaves, tree_structure, tree_unflatten
import spiking_learning as sl
import numpy as np
from utils import cos_sim_train_func, online_sim_train_func, test_func, custom_snn, bp_snn
import torch
import tonic
# In[2]:
torch.manual_seed(0)
np.random.seed(0)
output_size = 20
nlayers = 3
seq_len = 50
slope = 25
lr = '001'
init_seed_val = 0
init_seed = jax.random.split(jax.random.PRNGKey(init_seed_val))[0]
dtype = jnp.float32
tau = dtype(2.)
batch_sz = 128
spike_fn = sl.fs(slope)
n_iter = 10000
layer_sz = 512
update_time = 'offline'
optimizer = optax.adamax(dtype(0.001))
# In[3]:
sensor_size = tonic.datasets.SHD.sensor_size
train = tonic.datasets.SHD('data',train=True,transform=tonic.transforms.ToFrame(sensor_size=sensor_size,n_time_bins=seq_len))
test = tonic.datasets.SHD('data',train=False,transform=tonic.transforms.ToFrame(sensor_size=sensor_size,n_time_bins=seq_len))
# In[5]:
train_data = []
train_labels = []
for i in range(len(train)):
d,l = train[i]
train_data.append(d)
train_labels.append(l)
train_data = dtype(jnp.concatenate(train_data,axis=1))
train_labels = dtype(jax.nn.one_hot(jnp.stack(train_labels),output_size))
train_labels = jnp.tile(jnp.expand_dims(train_labels,axis=2),seq_len).transpose(2,0,1)
val_data = train_data[:,0:len(train)//10]
val_labels = train_labels[:,0:len(train)//10]
train_data = train_data[:,len(train)//10:]
train_labels = train_labels[:,len(train)//10:]
# In[6]:
def gen_data(seed2):
inds = jnp.arange(train_labels.shape[1])
inds = jax.random.permutation(seed2,inds,independent=True)
data = train_data[:,inds[0:batch_sz]]
labels = train_labels[:,inds[0:batch_sz]]
return data, labels
# In[7]:
test_data = []
test_labels = []
for i in range(len(test)):
d,l = test[i]
test_data.append(d)
test_labels.append(l)
test_data = dtype(jnp.concatenate(test_data,axis=1))
test_labels = dtype(jax.nn.one_hot(jnp.stack(test_labels),output_size))
test_labels = jnp.tile(jnp.expand_dims(test_labels,axis=2),seq_len).transpose(2,0,1)
# In[10]:
OTTTmodel = custom_snn(output_sz=output_size, n_layers=nlayers, mod1=OTTT, mod2=OTTT, spike_fn=spike_fn, layer_sz=layer_sz, dtype=dtype)
OSTLmodel = custom_snn(output_sz=output_size, n_layers=nlayers, mod1=OSTL, mod2=OSTL, spike_fn=spike_fn, layer_sz=layer_sz, dtype=dtype)
OTPEmodel = custom_snn(output_sz=output_size, n_layers=nlayers, mod1=OSTL, mod2=OTPE, spike_fn=spike_fn, layer_sz=layer_sz, dtype=dtype)
Approx_OTPEmodel = custom_snn(output_sz=output_size, n_layers=nlayers, mod1=OTTT, mod2=Approx_OTPE, spike_fn=spike_fn, layer_sz=layer_sz, dtype=dtype)
carry = [OTPE.initialize_carry(dtype=dtype)]*nlayers
params = OTPEmodel.init(init_seed,carry,train_data[0,:batch_sz])
carry,s = OTPEmodel.apply(params,carry,train_data[0,:batch_sz])
opt_state = optimizer.init(params)
orig_params = params
# In[11]:
val_carry = [OTPE.test_carry()]*nlayers
val_carry,_ = OTPEmodel.apply(params,val_carry,val_data[0])
test_carry = [OTPE.test_carry()]*nlayers
test_carry,_ = OTPEmodel.apply(params,test_carry,test_data[0])
# In[12]:
bp_model = bp_snn(output_sz=output_size, n_layers=nlayers, spike_fn=spike_fn, layer_sz=layer_sz, dtype=dtype)
bp_carry = carry
bp_params = bp_model.init(init_seed,bp_carry,train_data[0,:batch_sz])
struct = tree_structure(bp_params)
bp_params = tree_unflatten(struct,tree_leaves(orig_params))
bp_carry,s = bp_model.apply(bp_params,bp_carry,train_data[0,:batch_sz])
bp_opt_state = optimizer.init(bp_params)
# In[13]:
carry = tree_map(lambda x: jnp.zeros_like(x,dtype),carry)
val_carry = tree_map(lambda x: jnp.zeros_like(x,dtype),val_carry)
test_carry = tree_map(lambda x: jnp.zeros_like(x,dtype),test_carry)
# In[14]:
key = jax.random.split(init_seed)[0]
cosine_similarity = []
cosine_similarity_layer = []
val_acc = []
train_loss = []
all_params = [params]*4
all_params.append(bp_params)
all_opt_states = [opt_state]*4
all_opt_states.append(bp_opt_state)
best_params = tree_map(jnp.zeros_like,all_params)
best_val = [0]*4
if update_time == 'offline':
best_val.append(0)
# In[15]:
offline_training = jax.jit(Partial(cos_sim_train_func,OTTTmodel,
Approx_OTPEmodel,
OSTLmodel,
OTPEmodel,
bp_model,
optimizer,
carry,
val_carry,
val_data,
val_labels,
batch_sz,
gen_data
))
# In[16]:
online_training = jax.jit(Partial(online_sim_train_func,OTTTmodel,
Approx_OTPEmodel,
OSTLmodel,
OTPEmodel,
optimizer,
carry,
val_carry,
val_data,
val_labels,
batch_sz,
gen_data
))
# In[17]:
#-------------- uncomment to save inital model parameters --------------#
# with open('SHD_data/models/model_{}layer_{}_{}seqlen_{}iter_{}seed_{}_sub_{}fs_adamax_lr{}'.format(nlayers,layer_sz,seq_len,0,init_seed_val,update_time,slope,lr),'wb') as file:
# pickle.dump(tree_map(jnp.float32,all_params),file,protocol=pickle.HIGHEST_PROTOCOL)
#-----------------------------------------------------------------------#
# In[18]:
for epoch in range(n_iter):
if update_time == 'offline':
all_loss, all_cosines, all_cosines_per, all_acc, all_params, all_opt_states, key = offline_training(all_params,all_opt_states,key)
cosine_similarity.append(np.stack(list(tree_map(jnp.float32,all_cosines))))
cosine_similarity_layer.append(np.stack(list((tree_map(jnp.float32,all_cosines_per)))))
elif update_time == 'online':
all_loss, all_acc, all_params, all_opt_states, key = online_training(all_params,all_opt_states,key)
val_acc.append(np.stack(list(tree_map(jnp.float32,all_acc))))
train_loss.append(np.stack(list(tree_map(jnp.float32,all_loss))))
print(val_acc[-1])
is_best = np.greater(val_acc[-1],best_val).squeeze()
best_val = np.where(is_best,val_acc[-1],best_val)
for i in range(len(best_val)):
if is_best[i]:
best_params[i] = all_params[i]
#-------------- uncomment for saving model parameters --------------#
# if (epoch+1)%200 == 0: #200
# with open('SHD_data/models/model_{}layer_{}_{}seqlen_{}iter_{}seed_{}_sub_{}fs_adamax_lr{}'.format(nlayers,layer_sz,seq_len,epoch+1,init_seed_val,update_time,slope,lr),'wb') as file:
# pickle.dump(tree_map(jnp.float32,all_params),file,protocol=pickle.HIGHEST_PROTOCOL)
#-------------------------------------------------------------------#
# In[19]:
#----------------- uncomment to save best model --------------------#
# with open('SHD_data/models/model_{}layer_{}_{}seqlen_best_{}seed_{}_sub_{}fs_adamax_lr{}'.format(nlayers,layer_sz,seq_len,init_seed_val,t_name,slope,lr),'wb') as file:
# pickle.dump(tree_map(jnp.float32,best_params),file,protocol=pickle.HIGHEST_PROTOCOL)
#-------------------------------------------------------------------#
# In[20]:
OTTT_acc = test_func(OTTTmodel,best_params[0],test_carry,(test_data,test_labels))
Approx_OTPE_acc = test_func(Approx_OTPEmodel,best_params[1],test_carry,(test_data,test_labels))
OSTL_acc = test_func(OSTLmodel,best_params[2],test_carry,(test_data,test_labels))
OTPE_acc = test_func(OTPEmodel,best_params[3],test_carry,(test_data,test_labels))
bp_acc = test_func(bp_model,best_params[4],test_carry,(test_data,test_labels))
all_acc = (OTTT_acc,Approx_OTPE_acc,OSTL_acc,OTPE_acc,bp_acc)
val_acc.append(np.stack(list(tree_map(jnp.float32,all_acc))))
if update_time == 'online':
val_acc[-1] = val_acc[-1][0:4]
# In[21]:
np.save('SHD_data/layer_cosine_similarity/sim_{}layer_{}_{}seqlen_{}iter_{}_sub_{}fs_adamax_lr{}_{}seed'.format(nlayers,layer_sz,seq_len,n_iter,update_time,slope,lr,init_seed_val),cosine_similarity_layer)
np.save('SHD_data/model_cosine_similarity/sim_{}layer_{}_{}seqlen_{}iter_{}_sub_{}fs_adamax_lr{}_{}seed'.format(nlayers,layer_sz,seq_len,n_iter,update_time,slope,lr,init_seed_val),cosine_similarity)
np.save('SHD_data/accuracy/sim_{}layer_{}_{}seqlen_{}iter_{}_sub_{}fs_adamax_lr{}_{}seed'.format(nlayers,layer_sz,seq_len,n_iter,update_time,slope,lr,init_seed_val),val_acc)
np.save('SHD_data/loss/sim_{}layer_{}_{}seqlen_{}iter_{}_sub_{}fs_adamax_lr{}_{}seed'.format(nlayers,layer_sz,seq_len,n_iter,update_time,slope,lr,init_seed_val),train_loss)