-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathparameters_linear.py
223 lines (174 loc) · 7.6 KB
/
parameters_linear.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
### Authors: Nicolas Y. Masse, Gregory D. Grant
import numpy as np
import pickle
print('\n--> Loading parameters...')
##############################
### Independent parameters ###
##############################
par = {
# Setup parameters
'savedir' : './savedir/',
'save_fn' : '80_tasks_v1',
'train' : True,
'save_weights' : True,
'learning_method' : 'RL', # 'RL' or 'SL'
'LSTM_init' : 0.02,
'w_init' : 0.02,
# Network shape
'num_motion_tuned' : 32,
'num_fix_tuned' : 1,
'num_rule_tuned' : 0,
'n_hidden' : 250,
'n_val' : 4,
'n_modules' : 4,
'n_module_out' : 10,
'n_ff0' : 200,
'n_ff1' : 200,
'top_down' : True,
# Encoder configuration
'n_latent' : 200,
'enc_activity_cost' : 0.1,
'enc_weight_cost' : 0.05,
'internal_sampling' : True,
# Cortex configuration
'sample_step' : 5,
# read/write configuration
'A_alpha_init' : 0.99995,
'A_beta_init' : 1.5,
'inner_steps' : 1,
'batch_norm_inner' : False,
# Timings and rates
'learning_rate' : 5e-4,
'drop_rate' : 0.5,
# Variance values
'input_mean' : 0.0,
'noise_in' : 0.1,
'noise_rnn' : 0.05,
'n_filters' : 1,
# Task specs
'task' : 'multistim',
'trial_length' : 2000,
'mask_duration' : 0,
'dead_time' : 200,
'dt' : 100,
'trials_per_seq' : 70,
'task_list' : [a for a in range(1,81)], #[a for a in range(49)],
'dead_trials' : 60,
# RL parameters
'fix_break_penalty' : -1.,
'wrong_choice_penalty' : -0.01,
'correct_choice_reward' : 1.,
'discount_rate' : 0.9,
# Tuning function data
'num_motion_dirs' : 8,
'tuning_height' : 4.0,
# Cost values
'sparsity_cost' : 1e-2, # was 1e-2
'rec_cost' : 1e-3, # was 1e-2
'weight_cost' : 1e-5, # was 1e-6
'entropy_cost' : 0.01,
'val_cost' : 0.01,
'stim_cost' : 1e-1,
# Training specs
'batch_size' : 128,
'n_batches' : 3000000, # 1500 to train straight cortex
}
############################
### Dependent parameters ###
############################
def update_parameters(updates, verbose=True, update_deps=True):
""" Updates parameters based on a provided
dictionary, then updates dependencies """
par.update(updates)
if verbose:
print('Updating parameters:')
for (key, val) in updates.items():
print('{:<24} --> {}'.format(key, val))
if update_deps:
update_dependencies()
def load_encoder_weights():
fn = './savedir/80_tasks_v0_model_weights.pkl'
results = pickle.load(open(fn, 'rb'))
print('Weight keys ', results['weights'].keys())
par['W0_init'] = results['weights']['W0']
par['W1_init'] = results['weights']['W1']
def update_weights(var_dict):
print('Setting weight values manually; disabling training and weight saving.')
par['train'] = False
par['save_weights'] = False
for key, val in var_dict['weights'].items():
print(key, val.shape)
if not 'A_' in key:
par[key+'_init'] = val
def update_dependencies():
""" Updates all parameter dependencies """
# Reward map, for hippocampus reward one-hot conversion
par['reward_map'] = {
par['fix_break_penalty'] : 0,
par['wrong_choice_penalty'] : 1,
0. : 2,
par['correct_choice_reward'] : 3
}
par['num_reward_types'] = len(par['reward_map'].keys())
par['reward_map_matrix'] = np.zeros([par['num_reward_types'],1]).astype(np.float32)
for key, val in par['reward_map'].items():
par['reward_map_matrix'][val,:] = key
# Set input and output sizes
par['n_input'] = par['num_motion_tuned'] + par['num_fix_tuned'] + par['num_rule_tuned']
par['n_pol'] = par['num_motion_dirs'] + 1
# Set trial step length
par['num_time_steps'] = par['trial_length']//par['dt']
# Set up standard LSTM weights and biases
LSTM_weight = lambda size : np.random.uniform(-par['LSTM_init'], par['LSTM_init'], size=size).astype(np.float32)
# option 1
#for p in ['Wf', 'Wi', 'Wo', 'Wc']: par[p+'_init'] = LSTM_weight([par['n_input']+par['n_hidden'], par['n_hidden']])
#par['W_write_init'] = LSTM_weight([par['n_input']+par['n_val']+par['n_pol'], par['n_latent']])
# option 2
#for p in ['Wf', 'Wi', 'Wo', 'Wc']: par[p+'_init'] = LSTM_weight([par['n_input'], par['n_hidden'][0]])
#for j in range(par['n_modules']):
# for p in ['Wf', 'Wi', 'Wo', 'Wc']: par[p+str(j)+'_init'] = LSTM_weight([par['n_input'], par['n_hidden'][j]])
# for p in ['Uf', 'Ui', 'Uo', 'Uc']: par[p+str(j)+'_init'] = LSTM_weight([par['n_hidden'][0], par['n_hidden'][j]])
# for p in ['bf', 'bi', 'bo', 'bc']: par[p+str(j)+'_init'] = np.zeros([1, par['n_hidden'][j]], dtype=np.float32)
par['W0_init'] = np.random.uniform(-0.02, 0.02, size=[par['n_input']*par['n_filters'], par['n_ff0']]).astype(np.float32)
#par['W1_init'] = np.random.uniform(-1., 1., size=[par['n_ff0'], par['n_ff1']]).astype(np.float32)
par['W1_init'] = np.random.uniform(-0.02, 0.02, size=[par['n_ff0'], par['n_input']*par['n_filters']]).astype(np.float32)
par['W2_init'] = np.random.uniform(-0.02, 0.02, size=[par['n_ff0'], par['n_ff1']]).astype(np.float32)
par['W_td_init'] = np.random.uniform(-0.02, 0.02, size=[par['n_hidden'], par['n_ff0']]).astype(np.float32)
par['b0_init'] = np.zeros([1, par['n_ff0']], dtype=np.float32)
par['b2_init'] = np.zeros([1, par['n_ff1']], dtype=np.float32)
# V0
n_input_ctl = par['n_pol']*par['n_val'] + par['n_pol'] + par['n_val'] + par['n_input']*par['n_filters']
#n_input_ctl = 33 + par['n_pol'] + par['n_val'] + par['n_pol']*par['n_val']
# V1
#n_input_ctl = par['n_module_out']*par['n_modules'] + par['n_pol'] + par['n_val'] + par['n_pol']*par['n_val']
#n_input_ctl = par['n_input'] + par['n_module_out']*par['n_modules'] + par['n_pol']*par['n_val']
for p in ['Wf', 'Wi', 'Wo', 'Wc']: par[p+'_init'] = LSTM_weight([n_input_ctl, par['n_hidden']])
for p in ['Uf', 'Ui', 'Uo', 'Uc']: par[p+'_init'] = LSTM_weight([par['n_hidden'], par['n_hidden']])
for p in ['bf', 'bi', 'bo', 'bc']: par[p+'_init'] = np.zeros([1, par['n_hidden']], dtype=np.float32)
# LSTM posterior distribution weights
#for p in ['Pf', 'Pi', 'Po', 'Pc']: par[p+'_init'] = LSTM_weight([par['n_tasks'], par['n_hidden']])
# Cortex RL weights and biases
par['W_pol_init'] = np.random.uniform(-par['w_init'], par['w_init'], size=[par['n_hidden'], par['n_pol']]).astype(np.float32)
par['b_pol_init'] = np.zeros([1,par['n_pol']], dtype=np.float32)
par['W_val_init'] = np.random.uniform(-par['w_init'], par['w_init'], size=[par['n_hidden'], 1]).astype(np.float32)
par['b_val_init'] = np.zeros([1,1], dtype=np.float32)
par['W_norm'] = np.zeros((par['n_ff1'], par['n_ff1']), dtype=np.float32)
for i in range(par['n_ff1']):
u = np.arange(i, i + 50)%par['n_ff1']
par['W_norm'][i, u] = 1.
# Gate weights and biases
"""
par['W_pos_gate_init'] = np.random.uniform(-par['w_init'], par['w_init'], size=[par['n_tasks'],1]).astype(np.float32)
par['W_cor_gate_init'] = np.random.uniform(-par['w_init'], par['w_init'], size=[par['n_pol'],1]).astype(np.float32)
par['W_hip_gate_init'] = np.random.uniform(-par['w_init'], par['w_init'], size=[par['n_pol'],1]).astype(np.float32)
par['W_cor_gate_val_init'] = np.random.uniform(-par['w_init'], par['w_init'], size=[par['n_val'],1]).astype(np.float32)
par['W_hip_gate_val_init'] = np.random.uniform(-par['w_init'], par['w_init'], size=[par['n_val'],1]).astype(np.float32)
par['b_act_gate_init'] = np.ones([1,1], dtype=np.float32)
par['b_pos_gate_init'] = np.ones([1,1], dtype=np.float32)
par['encoder_weight_file'] = './datadir/gotask_50unit_input_encoder_weights.pkl'
print('--> Loading encoder from {}.'.format(par['encoder_weight_file']))
par['encoder_init'] = pickle.load(open(par['encoder_weight_file'], 'rb'))['weights']['W']
"""
load_encoder_weights()
update_dependencies()
print('--> Parameters successfully loaded.\n')