Upload main.py
Browse files
main.py
ADDED
@@ -0,0 +1,244 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy as np
|
2 |
+
from numba import njit
|
3 |
+
from tqdm import tqdm
|
4 |
+
import math
|
5 |
+
import random
|
6 |
+
from matplotlib import pyplot as plt
|
7 |
+
import pickle
|
8 |
+
|
9 |
+
|
10 |
+
# whitelist = "ёйцукенгшщзхъфывапролджэячсмитьбю "
|
11 |
+
|
12 |
+
def text_to_arr(text: str):
|
13 |
+
return np.array([ord(x) for x in text.lower()])
|
14 |
+
|
15 |
+
@njit
|
16 |
+
def longest_common_substring(s1, s2):
|
17 |
+
current_match_start = -1
|
18 |
+
current_match_end = -1
|
19 |
+
|
20 |
+
best_match_start = current_match_start
|
21 |
+
best_match_end = current_match_end
|
22 |
+
|
23 |
+
min_len = min(len(s1), len(s2))
|
24 |
+
for i in range(min_len):
|
25 |
+
if s1[i] == s2[i]:
|
26 |
+
current_match_start = current_match_end = i
|
27 |
+
j = 0
|
28 |
+
while s1[i+j] == s2[i+j] and i+j < min_len:
|
29 |
+
j += 1
|
30 |
+
current_match_end = current_match_start + j
|
31 |
+
|
32 |
+
if current_match_end - current_match_start > best_match_end - best_match_start:
|
33 |
+
best_match_start = current_match_start
|
34 |
+
best_match_end = current_match_end
|
35 |
+
|
36 |
+
return s1[best_match_start:best_match_end]
|
37 |
+
|
38 |
+
def not_found_in(q, data):
|
39 |
+
for l in data:
|
40 |
+
count = 0
|
41 |
+
lq = len(q)-1
|
42 |
+
for v in l:
|
43 |
+
if v == q[count]:
|
44 |
+
count += 1
|
45 |
+
else:
|
46 |
+
count = 0
|
47 |
+
if count == lq:
|
48 |
+
return False
|
49 |
+
return True
|
50 |
+
|
51 |
+
class Layer:
|
52 |
+
def __init__(self, mem_len: int = 100, max_size: int = 6):
|
53 |
+
self.mem_len = mem_len
|
54 |
+
self.common_strings = []
|
55 |
+
self.previously_seen = []
|
56 |
+
self.max_size = max_size+1
|
57 |
+
def __call__(self, input_arr, training: bool = True):
|
58 |
+
o = []
|
59 |
+
li = len(input_arr)
|
60 |
+
for i in range(li):
|
61 |
+
for y, cs in enumerate(self.common_strings):
|
62 |
+
if (i+cs.shape[0]) <= li and (input_arr[i:i+cs.shape[0]] == cs).all():
|
63 |
+
o.append(y)
|
64 |
+
if training:
|
65 |
+
cl = 0
|
66 |
+
n = None
|
67 |
+
for i, line in enumerate(self.previously_seen):
|
68 |
+
t = longest_common_substring(input_arr, line)
|
69 |
+
l = len(t)
|
70 |
+
if l > cl and l < self.max_size:
|
71 |
+
cl = l
|
72 |
+
n = i
|
73 |
+
r = t
|
74 |
+
if self.previously_seen != []:
|
75 |
+
if n is not None and len(r) > 1:
|
76 |
+
self.previously_seen.pop(n)
|
77 |
+
if not_found_in(r, self.common_strings):
|
78 |
+
self.common_strings.append(r)
|
79 |
+
self.previously_seen = self.previously_seen[-self.mem_len:]
|
80 |
+
self.previously_seen.append(input_arr)
|
81 |
+
return o
|
82 |
+
|
83 |
+
def comparefilter(f1, f2):
|
84 |
+
o = 0
|
85 |
+
hss = 0.5
|
86 |
+
for k in f1:
|
87 |
+
if k in f2 and k in f1:
|
88 |
+
o += np.sum((f2[k] > hss)==(f1[k] > hss))
|
89 |
+
return (o >= len(f1)*hss)
|
90 |
+
|
91 |
+
class StrConv:
|
92 |
+
def __init__(self, filters: int, size: int = 4):
|
93 |
+
self.filter_amount = filters
|
94 |
+
self.filters = [{} for _ in range(filters)] # [{43: [3 2 0 3]},]
|
95 |
+
self.bias = np.zeros((self.filter_amount,))
|
96 |
+
self.size = 3
|
97 |
+
def regularize(self):
|
98 |
+
for n, f in enumerate(self.filters):
|
99 |
+
for f2 in self.filters[:n]:
|
100 |
+
if random.randint(0, 100) < 10 and comparefilter(f, f2):
|
101 |
+
self.filters[n] = {}
|
102 |
+
def __call__(self, input_arr, training: bool = True, debug=False):
|
103 |
+
if len(input_arr) <= self.size:
|
104 |
+
return []
|
105 |
+
o = np.zeros((input_arr.shape[0]-self.size, self.filter_amount))
|
106 |
+
for i in range(input_arr.shape[0]-self.size):
|
107 |
+
for n, c in enumerate(input_arr[i:i+self.size]):
|
108 |
+
for fn, f in enumerate(self.filters):
|
109 |
+
if c in f:
|
110 |
+
o[i, fn] += f[c][n]
|
111 |
+
o += self.bias
|
112 |
+
m = np.max(np.abs(o))
|
113 |
+
if m != 0: o /= m
|
114 |
+
if debug:
|
115 |
+
plt.imshow(o)
|
116 |
+
plt.show()
|
117 |
+
if training:
|
118 |
+
for i in range(input_arr.shape[0]-self.size):
|
119 |
+
for n, c in enumerate(input_arr[i:i+self.size]):
|
120 |
+
for fn, f in enumerate(self.filters):
|
121 |
+
if c in f:
|
122 |
+
# s = np.sum(f[c])
|
123 |
+
# if s > 1000:
|
124 |
+
# f[c] = (f[c]/(s/(self.size*1000))).astype(np.int64)
|
125 |
+
self.filters[fn][c][n] = o[i, fn]*0.1+f[c][n]*0.9
|
126 |
+
else:
|
127 |
+
f[c] = np.random.uniform(0, 1, (self.size))
|
128 |
+
f[c][n] = o[i, fn]
|
129 |
+
# for t in range(self.size, input_arr.shape[0]):
|
130 |
+
# for f in range(self.filter_amount):
|
131 |
+
# self.filters[f] = o[t-self.size, f]
|
132 |
+
"""
|
133 |
+
s = 0
|
134 |
+
for a in self.filters:
|
135 |
+
for b in a:
|
136 |
+
s += np.sum(b)
|
137 |
+
if s > 100:
|
138 |
+
s /= self.filter_amount
|
139 |
+
for a in self.filters:
|
140 |
+
for b in a:
|
141 |
+
a[b] = (a[b]/s).astype(dtype=np.int64)
|
142 |
+
"""
|
143 |
+
self.bias -= np.sum(o, axis=0)# / o.shape[0]
|
144 |
+
# print(o)
|
145 |
+
maxed = np.zeros((o.shape[0],)) # could have different outputs, not only max of o, like o>(self.size//2) or o without processing
|
146 |
+
for i in range(maxed.shape[0]):
|
147 |
+
maxed[i] = np.argmax(o[i])
|
148 |
+
return maxed
|
149 |
+
|
150 |
+
with open("dataset.txt", "r") as f:
|
151 |
+
lines = f.read().rstrip("\n").split("\n")[:40000]
|
152 |
+
|
153 |
+
w = {}
|
154 |
+
w2 = {}
|
155 |
+
|
156 |
+
c = 0
|
157 |
+
|
158 |
+
#layer = Layer(mem_len=1000, max_size=4)
|
159 |
+
#layer2 = Layer(mem_len=1000, max_size=6)
|
160 |
+
|
161 |
+
with open("l1_large.pckl", "rb") as f: layer = pickle.load(f)
|
162 |
+
with open("l2_large.pckl", "rb") as f: layer2 = pickle.load(f)
|
163 |
+
with open("w1_large.pckl", "rb") as f: w = pickle.load(f)
|
164 |
+
with open("w2_large.pckl", "rb") as f: w2 = pickle.load(f)
|
165 |
+
"""
|
166 |
+
for n, text in tqdm(enumerate(lines[:-1])):
|
167 |
+
if text.strip() != "" and lines[n+1].strip() != "" and text != lines[n+1]:
|
168 |
+
t = layer(text_to_arr(text), training=True)
|
169 |
+
t = layer(text_to_arr(text), training=False)
|
170 |
+
c += 1
|
171 |
+
# if c == 10:
|
172 |
+
# c = 0
|
173 |
+
# layer.regularize()
|
174 |
+
# layer2.regularize()
|
175 |
+
if len(t) != 0:
|
176 |
+
t2 = layer2(np.array(t), training=True)
|
177 |
+
t2 = layer2(np.array(t), training=False)
|
178 |
+
for a in t2:
|
179 |
+
if a in w2:
|
180 |
+
w2[a].append(n+1)
|
181 |
+
else:
|
182 |
+
w2[a] = [n+1,]
|
183 |
+
for a in t:
|
184 |
+
if a in w:
|
185 |
+
w[a].append(n+1)
|
186 |
+
else:
|
187 |
+
w[a] = [n+1,]
|
188 |
+
|
189 |
+
for n, text in tqdm(enumerate(lines[:200])):
|
190 |
+
if text.strip() != "" and lines[n+1].strip() != "" and text != lines[n+1]:
|
191 |
+
t = layer(text_to_arr(text), training=True)
|
192 |
+
t = layer(text_to_arr(text), training=False)
|
193 |
+
c += 1
|
194 |
+
# if c == 10:
|
195 |
+
# c = 0
|
196 |
+
# layer.regularize()
|
197 |
+
# layer2.regularize()
|
198 |
+
if len(t) != 0:
|
199 |
+
t2 = layer2(np.array(t), training=True)
|
200 |
+
t2 = layer2(np.array(t), training=False)
|
201 |
+
for a in t2:
|
202 |
+
if a in w2:
|
203 |
+
w2[a].append(n+1)
|
204 |
+
else:
|
205 |
+
w2[a] = [n+1,]
|
206 |
+
for a in t:
|
207 |
+
if a in w:
|
208 |
+
w[a].append(n+1)
|
209 |
+
else:
|
210 |
+
w[a] = [n+1,]
|
211 |
+
|
212 |
+
with open("l1_large.pckl", "wb") as f: pickle.dump(layer, f)
|
213 |
+
with open("l2_large.pckl", "wb") as f: pickle.dump(layer2, f)
|
214 |
+
with open("w1_large.pckl", "wb") as f: pickle.dump(w, f)
|
215 |
+
with open("w2_large.pckl", "wb") as f: pickle.dump(w2, f)
|
216 |
+
"""
|
217 |
+
# print(layer.filters)
|
218 |
+
|
219 |
+
#for arr in layer.common_strings:
|
220 |
+
# print(''.join([chr(a) for a in arr]))
|
221 |
+
|
222 |
+
print(len(lines), "responses available")
|
223 |
+
|
224 |
+
import threeletterai
|
225 |
+
|
226 |
+
while True:
|
227 |
+
msg = input("Message: ")
|
228 |
+
if len(msg) < 4:
|
229 |
+
print(threeletterai.getresp(msg))
|
230 |
+
continue
|
231 |
+
processed = layer(text_to_arr(msg), training=False)
|
232 |
+
processed = np.array(processed)
|
233 |
+
processed2 = layer2(processed, training=False)
|
234 |
+
# print(processed)
|
235 |
+
# print(processed2)
|
236 |
+
o = np.zeros(len(lines), dtype=np.int16)
|
237 |
+
for a in processed:
|
238 |
+
if a in w:
|
239 |
+
o[w[a]] += 1
|
240 |
+
for a in processed2:
|
241 |
+
if a in w2:
|
242 |
+
o[w2[a]] += 1
|
243 |
+
print(lines[np.argmax(o)], f" {np.max(o)} sure")
|
244 |
+
|