-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathlstm_test.py
210 lines (175 loc) · 6.83 KB
/
lstm_test.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
""" imports """
import dtorch.nn as nn
import dtorch.optim as optim
import dtorch.jtensors as jt
import dtorch.loss as ls
import torch as t
import pandas as pd
import numpy as np
import sys
import matplotlib.pyplot as plt
sys.setrecursionlimit(15000)
sequence_length = 4
def load_dataset(filename : str):
dataset = pd.read_csv(filename)
data = np.array(list(dataset.date))
return jt.JTensors(data - data[-1], require_grads=True), jt.JTensors((np.array(list(dataset.high)) + np.array(list(dataset.low))) / 2, require_grads=True)
x_train, y_train = load_dataset(sys.argv[1])
y_train_s = jt.JTensors(((y_train.numpy() / y_train.numpy().max())) - 0.7, require_grads=True)
x_train.update(np.lib.stride_tricks.as_strided((y_train_s() / y_train_s().max()), shape=(len(y_train_s) - 2, sequence_length), strides=(y_train_s().strides[0], y_train_s().strides[0])))
y_train.update(x_train[1:])
x_train.update(x_train[:-1])
# a tester
# w x y z => [[w, x, y], [x, y, z]]
x_train_tensor = x_train.reshape(553, sequence_length, 1)
y_train_tensor = y_train.reshape(553, sequence_length, 1)
x_torch = t.tensor(x_train.numpy(), dtype=t.float32, requires_grad=True).unsqueeze(-1)
y_torch = t.tensor(y_train.numpy(), dtype=t.float32, requires_grad=True).unsqueeze(-1)
class MyBalls(nn.Module):
def __init__(self):
super().__init__()
self.lstm = nn.RNN(1, 1, 2, batch_first=True, bias=True)
def forward(self, x):
res, h = self.lstm(x)
return res, h
class MyTorchBalls(t.nn.Module):
def __init__(self):
super(MyTorchBalls, self).__init__()
self.rnn = t.nn.RNN(1, 1, 2, batch_first=True, bias=True)
def forward(self, x):
return self.rnn(x)
torch_model = MyTorchBalls()
model = MyBalls()
torch_model.rnn.weight_hh_l0 = t.nn.Parameter(t.tensor([[0.5]], requires_grad=True))
torch_model.rnn.weight_ih_l0 = t.nn.Parameter(t.tensor([[0.5]], requires_grad=True))
torch_model.rnn.bias_hh_l0 = t.nn.Parameter(t.tensor([[0.5]], requires_grad=True))
torch_model.rnn.bias_ih_l0 = t.nn.Parameter(t.tensor([[0.5]], requires_grad=True))
torch_model.rnn.weight_hh_l1 = t.nn.Parameter(t.tensor([[0.5]], requires_grad=True))
torch_model.rnn.weight_ih_l1 = t.nn.Parameter(t.tensor([[0.5]], requires_grad=True))
torch_model.rnn.bias_hh_l1 = t.nn.Parameter(t.tensor([[0.5]], requires_grad=True))
torch_model.rnn.bias_ih_l1 = t.nn.Parameter(t.tensor([[0.5]], requires_grad=True))
model.lstm.weight_hh_l1 = nn.Parameter(jt.JTensors([[0.5]]))
model.lstm.weight_ih_l1 = nn.Parameter(jt.JTensors([[0.5]]))
model.lstm.bias_hh_l1 = nn.Parameter(jt.JTensors([[0.5]]))
model.lstm.bias_ih_l1 = nn.Parameter(jt.JTensors([[0.5]]))
model.lstm.weight_hh_l0 = nn.Parameter(jt.JTensors([[0.5]]), 'weight_hh_l0')
model.lstm.weight_ih_l0 = nn.Parameter(jt.JTensors([[0.5]]), 'weight_ih_l0')
model.lstm.bias_hh_l0 = nn.Parameter(jt.JTensors([[0.5]]))
model.lstm.bias_ih_l0 = nn.Parameter(jt.JTensors([[0.5]]))
# i = input('load model ?')
# if i == 'y':
# torch_model.load_state_dict(t.load('model.pt'))
# torch_model.eval()
#
# m, _ = torch_model(x_torch)
#
# valid = 0
# tot = 0
# for i in range(len(m)):
# o = m[i].item() - m[i - 1].item()
# u = y_torch[i][-1].item() - y_torch[i - 1][-1].item()
# if o < 0 and u < 0 or o > 0 and u > 0 or o == 0 and u == 0:
# valid += 1
# tot += 1
#
# print(valid / tot)
#
# plt.scatter(np.arange(0, len(x_train())), y_train_s()[:-3], label='Training data')
# plt.plot(np.arange(0, len(x_train())), m.detach().numpy(), color='red', label='Predicted')
# #plt.plot(x_train(), y_pred(), color='red', label='Predicted')
# plt.xlabel('x')
# plt.ylabel('y')
# plt.title('Non-linear Model Predicting Sine Function')
# plt.legend()
# plt.show()
#print(model.lstm._all_weights)
optimizer = optim.SGD(model.parameters(), lr=0.01)
optimizer_torch = t.optim.SGD(torch_model.parameters(), lr=0.01)
#x = jt.JTensors([[1], [40], [2], [8]]) / 40
#x_torch = t.tensor([[[1], [40], [2], [8]]], dtype=t.float32) / 40
#x_train_tensor = jt.JTensors([[[2], [40]], [[1], [30]]]) / 40
#x_torch = t.tensor([[[2], [40]], [[1], [30]]], dtype=t.float32) / 40
#print(t.tanh(t.matmul(t.tensor([[[1], [3]]], dtype=t.float32), model.lstm.weight_ih_l0.t()) + model.lstm.bias_ih_l0 + t.matmul(t.tensor([[0]], dtype=t.float32), model.lstm.weight_hh_l0.t()) + model.lstm.bias_hh_l0))
loss = ls.MSELoss()
loss_torch = t.nn.MSELoss()
#y_train_tensor = jt.JTensors([[[40], [8]], [[30], [9]]]) / 40
#y_torch = t.tensor([[[40], [8]], [[30], [9]]], dtype=t.float32) / 40
a= 0
#print("y:", y_train_tensor)
epochs = 1
for i in range(epochs):
optimizer.zero_grad()
m, prout = model(x_train_tensor)
#print("m", m)
optimizer_torch.zero_grad()
m_torch, prout_torch = torch_model.forward(x_torch)
#print("m_torch", m_torch)
#print("prout:", prout)
a = loss(m, y_train_tensor)
print(f"loss at {i}:", a[0])
a.backward()
print("autograd:")
print(model.lstm.weight_hh_l0.get().grad)
print(model.lstm.weight_ih_l0.get().grad)
optimizer.step()
u_torch = loss_torch(m_torch, y_torch)
u_torch.backward()
print("torch:")
print(torch_model.rnn.weight_hh_l0.grad)
print(torch_model.rnn.weight_ih_l0.grad)
optimizer_torch.step()
print(f"loss_torch at {i}:", u_torch.item())
model.eval()
torch_model.eval()
m, prout = model.forward(x_train_tensor)
m_torch, prout_torch = torch_model.forward(x_torch)
#t.save(torch_model.state_dict(), "torch_model.pt")
#print("prout:", prout)
#print("prout_torch:", prout_torch)
print("m:", m)
print("m_torch:", m_torch)
# a b c =>
#m = ()
#valid = 0
#tot = 0
#for i in range(len(m_torch)):
# #print("===")
# #print(m_torch[i])
# #print(y_train_tensor[i])
# tot += 1
# valid += (m_torch[i][-1].item() > x_torch[i][-1]) == (y_train_tensor[i][-1][0] > x_torch[i][-1])
#
#print("valid:", valid)
#print("tot:", tot)
#print("accuracy:", valid / tot)
#print(a)
#plt.scatter(np.arange(0, len(x_train())), y_train_s()[:-3], label='Training data')
#print(y_train_s()[-3:])
#print([x[2] for x in m.numpy()][:3])
#print(m.numpy()[-3:])
#plt.plot(np.arange(0, len(x_train())), [x[2] for x in m.numpy()], color='red', label='Predicted')
##plt.plot(x_train(), y_pred(), color='red', label='Predicted')
#plt.xlabel('x')
#plt.ylabel('y')
#plt.title('Non-linear Model Predicting Sine Function')
#plt.legend()
#plt.show()
#i = input("save model? (y/n): ")
#if i == "y":
# t.save(torch_model.state_dict(), "./model.pt")
# print("model saved")
#else:
# print("model not saved")
# save the model
#i = input("save model? (y/n): ")
#if i == "y":
# model.save("./model.jt")
# print("model saved")
#else:
# print("model not saved")
#print("autograd:")
#print(model.lstm.weight_hh_l0)
#print(model.lstm.weight_ih_l0)
#print("torch:")
#print(torch_model.lstm.weight_hh_l0)
#print(torch_model.lstm.weight_ih_l0)