-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathphase_mlp_actor.py
295 lines (236 loc) · 13.7 KB
/
phase_mlp_actor.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
import math
import torch
import numpy as np
import torch.nn as nn
from torch.autograd import Variable
import torch.nn.functional as F
import torch.nn.init as init
def kn(p, n):
return ((np.floor((4*p)/(2*math.pi)) + n - 1) % 4).astype(int)
def spline_w(p):
return ((4*p)/(2*math.pi)) % 1
def compute_multipliers(w,p):
list_of_w = [torch.zeros(p.shape), torch.zeros(p.shape), torch.zeros(p.shape), torch.zeros(p.shape)]
w2 = torch.pow(w,2)
w3 = torch.pow(w,3)
for n in range(4):
ind = kn(p,n)
if n == 0:
weight = w2 - 0.5*w - 0.5*w3
elif n == 1:
weight = 1 - 2.5*w2 + 1.5*w3
elif n == 2:
weight = 0.5*w + 2*w2 - 1.5*w3
elif n == 3:
weight = 0.5*(w3 - w2)
for j in range(len(ind)):
list_of_w[ind[j,0]][j,0] = weight[j,0]
return list_of_w[0], list_of_w[1], list_of_w[2], list_of_w[3]
def init_fanin(tensor):
fanin = tensor.size(1)
v = 1.0 / np.sqrt(fanin)
init.uniform(tensor, -v, v)
class PMLP(nn.Module):
def __init__(self, input_size, output_size, hidden_size, dtype, n_layers=1, batch_size=1, scale=1.0, tanh_flag=0):
super(PMLP, self).__init__()
self.input_size = input_size
self.output_size = output_size
self.hidden_size_1 = hidden_size[0]
self.hidden_size_2 = hidden_size[-1]
self.n_layers = n_layers
self.batch_size = batch_size
self.scale = scale # scale output of actor from [-1,1] to range of action space [-scale,scale]. set to 1 for critic
self.tanh_flag = tanh_flag # 1 for actor, 0 for critic (since critic range need not be restricted to [-1,1])
self.dtype = dtype
#self.control_gru_list = []
self.control_hidden_list = []
self.control_h2o_list = []
#self.gru_00 = nn.GRUCell(self.input_size, self.hidden_size)
self.l_00 = nn.Linear(self.input_size, self.hidden_size_1).type(dtype)
self.h2o_0 = nn.Linear(self.hidden_size_2, self.output_size).type(dtype)
#self.gru_10 = nn.GRUCell(self.input_size, self.hidden_size)
self.l_10 = nn.Linear(self.input_size, self.hidden_size_1).type(dtype)
self.h2o_1 = nn.Linear(self.hidden_size_2, self.output_size).type(dtype)
#self.gru_20 = nn.GRUCell(self.input_size, self.hidden_size)
self.l_20 = nn.Linear(self.input_size, self.hidden_size_1).type(dtype)
self.h2o_2 = nn.Linear(self.hidden_size_2, self.output_size).type(dtype)
#self.gru_30 = nn.GRUCell(self.input_size, self.hidden_size)
self.l_30 = nn.Linear(self.input_size, self.hidden_size_1).type(dtype)
self.h2o_3 = nn.Linear(self.hidden_size_2, self.output_size).type(dtype)
init_fanin(self.l_00.weight)
init_fanin(self.l_10.weight)
init_fanin(self.l_20.weight)
init_fanin(self.l_30.weight)
init.uniform(self.h2o_0.weight,-3e-3, 3e-3)
init.uniform(self.h2o_0.bias,-3e-3, 3e-3)
init.uniform(self.h2o_1.weight,-3e-3, 3e-3)
init.uniform(self.h2o_1.bias,-3e-3, 3e-3)
init.uniform(self.h2o_2.weight,-3e-3, 3e-3)
init.uniform(self.h2o_2.bias,-3e-3, 3e-3)
init.uniform(self.h2o_3.weight,-3e-3, 3e-3)
init.uniform(self.h2o_3.bias,-3e-3, 3e-3)
if n_layers == 2:
#self.gru_01 = nn.GRUCell(self.hidden_size, self.hidden_size)
#self.gru_11 = nn.GRUCell(self.hidden_size, self.hidden_size)
#self.gru_21 = nn.GRUCell(self.hidden_size, self.hidden_size)
#self.gru_31 = nn.GRUCell(self.hidden_size, self.hidden_size)
self.l_01 = nn.Linear(self.hidden_size_1, self.hidden_size_2).type(dtype)
self.l_11 = nn.Linear(self.hidden_size_1, self.hidden_size_2).type(dtype)
self.l_21 = nn.Linear(self.hidden_size_1, self.hidden_size_2).type(dtype)
self.l_31 = nn.Linear(self.hidden_size_1, self.hidden_size_2).type(dtype)
init_fanin(self.l_01.weight)
init_fanin(self.l_11.weight)
init_fanin(self.l_21.weight)
init_fanin(self.l_31.weight)
self.control_hidden_list.append([self.l_00, self.l_10, self.l_20, self.l_30])
if n_layers == 2:
self.control_hidden_list.append([self.l_01, self.l_11, self.l_21, self.l_31])
self.control_h2o_list = [self.h2o_0, self.h2o_1, self.h2o_2, self.h2o_3]
#self.alpha = []
#for i in range(4):
# self.alpha.append(Alpha(n_layers))
#self.init_controls(self.control_hidden_list, self.control_h2o_list, self.alpha)
#self.h_0 = Variable(torch.zeros(batch_size, hidden_size), requires_grad=True)
#if n_layers == 2:
# self.h_1 = Variable(torch.zeros(batch_size, hidden_size), requires_grad=True)
self.hidden_list = []
self.h2o_list = []
self.phase_list = []
# to initialize grad of control hidden and h2o ... I need to do this stupid thing ...
dummy_x = Variable(torch.zeros(batch_size, input_size), requires_grad=False).type(dtype)
dummy_y = Variable(torch.zeros(batch_size, output_size), requires_grad=False).type(dtype)
dummy_criterion = nn.MSELoss()
if n_layers == 1:
for l, h2o in zip(self.control_hidden_list[0], self.control_h2o_list):
dummy_h = F.relu(l(dummy_x))
dummy_o = h2o(dummy_h)
dummy_loss = dummy_criterion(dummy_o, dummy_y)
dummy_loss.backward()
if n_layers == 2:
for l0, l1, h2o in zip(self.control_hidden_list[0], self.control_hidden_list[1], self.control_h2o_list):
dummy_h0 = F.relu(l0(dummy_x))
dummy_h1 = l1(dummy_h0)
dummy_o = h2o(dummy_h1)
dummy_loss = dummy_criterion(dummy_o, dummy_y)
dummy_loss.backward()
# reset to zero after dummy pass
#self.h_0 = Variable(torch.zeros(batch_size, hidden_size), requires_grad=True)
#if n_layers == 2:
# self.h_1 = Variable(torch.zeros(batch_size, hidden_size), requires_grad=True)
def forward(self,x,phase):
#w = self.weight_from_phase(phase, self.control_hidden_list, self.control_h2o_list)
#hiddens = []
#l0 = nn.Linear(self.input_size, self.hidden_size_1).type(self.dtype)
#hiddens.append(l0)
#h2o = nn.Linear(self.hidden_size_2, self.output_size).type(self.dtype)
#if self.n_layers == 2:
# l1 = nn.Linear(self.hidden_size_1, self.hidden_size_2).type(self.dtype)
# hiddens.append(l1)
#self.set_weight(w, hiddens, h2o)
#self.hidden_list.append(hiddens)
#self.h2o_list.append(h2o)
#self.phase_list.append(phase)
control_hidden_list = self.control_hidden_list
control_h2o_list = self.control_h2o_list
w = torch.from_numpy(spline_w(phase.data.cpu().numpy()))
w0,w1,w2,w3 = compute_multipliers(w,phase.data.cpu().numpy())
w0_h1 = Variable(w0.repeat(1,self.hidden_size_1).type(self.dtype))
w1_h1 = Variable(w1.repeat(1,self.hidden_size_1).type(self.dtype))
w2_h1 = Variable(w2.repeat(1,self.hidden_size_1).type(self.dtype))
w3_h1 = Variable(w3.repeat(1,self.hidden_size_1).type(self.dtype))
if self.n_layers == 2:
w0_h2 = Variable(w0.repeat(1,self.hidden_size_2).type(self.dtype))
w1_h2 = Variable(w1.repeat(1,self.hidden_size_2).type(self.dtype))
w2_h2 = Variable(w2.repeat(1,self.hidden_size_2).type(self.dtype))
w3_h2 = Variable(w3.repeat(1,self.hidden_size_2).type(self.dtype))
w0_o = Variable(w0.repeat(1,self.output_size).type(self.dtype))
w1_o = Variable(w1.repeat(1,self.output_size).type(self.dtype))
w2_o = Variable(w2.repeat(1,self.output_size).type(self.dtype))
w3_o = Variable(w3.repeat(1,self.output_size).type(self.dtype))
h_0 = F.relu(w0_h1*control_hidden_list[0][0](x) + w1_h1*control_hidden_list[0][1](x) + w2_h1*control_hidden_list[0][2](x) + w3_h1*control_hidden_list[0][3](x))
if self.n_layers == 2:
h_1 = F.relu(w0_h2*control_hidden_list[1][0](h_0) + w1_h2*control_hidden_list[1][1](h_0) + w2_h2*control_hidden_list[1][2](h_0) + w3_h2*control_hidden_list[1][3](h_0))
o = w0_o*control_h2o_list[0](h_1) + w1_o*control_h2o_list[1](h_1) + w2_o*control_h2o_list[2](h_1) + w3_o*control_h2o_list[3](h_1)
if self.tanh_flag:
o = F.tanh(o)
else:
o = w0_o*control_h2o_list[0](h_0) + w1_o*control_h2o_list[1](h_0) + w2_o*control_h2o_list[2](h_0) + w3_o*control_h2o_list[3](h_0)
if self.tanh_flag:
o = F.tanh(o)
#h_0 = F.relu(l0(x))
#if self.n_layers == 2:
# h_1 = F.relu(l1(h_0))
# if self.tanh_flag:
# o = F.tanh(h2o(h_1))
# else:
# o = h2o(h_1)
#else:
# if self.tanh_flag:
# o = F.tanh(h2o(h_0))
# else:
# o = h2o(h_0)
return self.scale*o
def reset(self):
#self.h_0 = Variable(torch.zeros(self.batch_size, self.hidden_size), requires_grad=True)
#if self.n_layers == 2:
# self.h_1 = Variable(torch.zeros(self.batch_size, self.hidden_size), requires_grad=True)
self.hidden_list = []
self.h2o_list = []
self.phase_list = []
#self.init_controls(self.control_hidden_list, self.control_h2o_list, self.alpha)
def weight_from_phase(self, phase, control_hidden_list, control_h2o_list):
weight = {}
w = spline_w(phase)
for n in range(len(control_hidden_list)):
for key in control_hidden_list[0][0]._parameters.keys():
weight[key + '_' + str(n)] = control_hidden_list[n][kn(phase, 1)]._parameters[key].data + w*0.5*(control_hidden_list[n][kn(phase, 2)]._parameters[key].data - control_hidden_list[n][kn(phase, 0)]._parameters[key].data) + w*w*(control_hidden_list[n][kn(phase, 0)]._parameters[key].data - 2.5*control_hidden_list[n][kn(phase, 1)]._parameters[key].data + 2*control_hidden_list[n][kn(phase, 2)]._parameters[key].data - 0.5*control_hidden_list[n][kn(phase, 3)]._parameters[key].data) + w*w*w*(1.5*control_hidden_list[n][kn(phase, 1)]._parameters[key].data - 1.5*control_hidden_list[n][kn(phase, 2)]._parameters[key].data + 0.5*control_hidden_list[n][kn(phase, 3)]._parameters[key].data - 0.5*control_hidden_list[n][kn(phase, 0)]._parameters[key].data)
for key in control_h2o_list[0]._parameters.keys():
weight[key] = control_h2o_list[kn(phase, 1)]._parameters[key].data + w*0.5*(control_h2o_list[kn(phase, 2)]._parameters[key].data - control_h2o_list[kn(phase, 0)]._parameters[key].data) + w*w*(control_h2o_list[kn(phase, 0)]._parameters[key].data - 2.5*control_h2o_list[kn(phase, 1)]._parameters[key].data + 2*control_h2o_list[kn(phase, 2)]._parameters[key].data - 0.5*control_h2o_list[kn(phase, 3)]._parameters[key].data) + w*w*w*(1.5*control_h2o_list[kn(phase, 1)]._parameters[key].data - 1.5*control_h2o_list[kn(phase, 2)]._parameters[key].data + 0.5*control_h2o_list[kn(phase, 3)]._parameters[key].data - 0.5*control_h2o_list[kn(phase, 0)]._parameters[key].data)
return weight
def set_weight(self, w, hiddens, h2o):
count = 0
for l in hiddens:
l._parameters['weight'].data = w['weight_' + str(count)]
l._parameters['bias'].data = w['bias_' + str(count)]
count += 1
h2o._parameters['weight'].data = w['weight']
h2o._parameters['bias'].data = w['bias']
#def init_controls(self, list_of_hidden, list_of_h2o, alpha):
# for i in range(len(alpha)):
# for j in range(len(list_of_hidden)):
# l = list_of_hidden[j][i]
# alpha[i]._parameters['weight_' + str(j)] = l._parameters['weight'].data.clone()
# alpha[i]._parameters['bias_' + str(j)] = l._parameters['bias'].data.clone()
# #initialize alpha grads as zero here using shape ...
# alpha[i]._grad['weight_' + str(j)] = torch.zeros(l._parameters['weight'].data.size()).type(self.dtype)
# alpha[i]._grad['bias_' + str(j)] = torch.zeros(l._parameters['bias'].data.size()).type(self.dtype)
# h2o = list_of_h2o[i]
# alpha[i]._parameters['weight'] = h2o._parameters['weight'].data.clone()
# alpha[i]._parameters['bias'] = h2o._parameters['bias'].data.clone()
# alpha[i]._grad['weight'] = torch.zeros(h2o._parameters['weight'].data.size()).type(self.dtype)
# alpha[i]._grad['bias'] = torch.zeros(h2o._parameters['bias'].data.size()).type(self.dtype)
def update_control_gradients(self):
for hiddens, phase in zip(self.hidden_list, self.phase_list):
w = spline_w(phase)
count = 0
for l in hiddens:
for key in l._parameters.keys():
self.control_hidden_list[count][kn(phase,0)]._parameters[key].grad.data += l._parameters[key].grad.data * (-0.5*w + w*w - 0.5*w*w*w)
self.control_hidden_list[count][kn(phase,1)]._parameters[key].grad.data += l._parameters[key].grad.data * (1 - 2.5*w*w + 1.5*w*w*w)
self.control_hidden_list[count][kn(phase,2)]._parameters[key].grad.data += l._parameters[key].grad.data * (0.5*w + 2*w*w - 1.5*w*w*w)
self.control_hidden_list[count][kn(phase,3)]._parameters[key].grad.data += l._parameters[key].grad.data * (-0.5*w*w + 0.5*w*w*w)
count += 1
for h2o, phase in zip(self.h2o_list, self.phase_list):
w = spline_w(phase)
for key in h2o._parameters.keys():
self.control_h2o_list[kn(phase,0)]._parameters[key].grad.data += h2o._parameters[key].grad.data * (-0.5*w + w*w - 0.5*w*w*w)
self.control_h2o_list[kn(phase,1)]._parameters[key].grad.data += h2o._parameters[key].grad.data * (1 - 2.5*w*w + 1.5*w*w*w)
self.control_h2o_list[kn(phase,2)]._parameters[key].grad.data += h2o._parameters[key].grad.data * (0.5*w + 2*w*w - 1.5*w*w*w)
self.control_h2o_list[kn(phase,3)]._parameters[key].grad.data += h2o._parameters[key].grad.data * (-0.5*w*w + 0.5*w*w*w)
#for i in range(len(self.control_hidden_list)):
# for alpha, l in zip(self.alpha, self.control_hidden_list[i]):
# for key in l._parameters.keys():
# l._parameters[key].grad.data += alpha._grad[key + '_' + str(i)]
#for alpha, h2o in zip(self.alpha, self.control_h2o_list):
# for key in h2o._parameters.keys():
# h2o._parameters[key].grad.data += alpha._grad[key]