-
Notifications
You must be signed in to change notification settings - Fork 0
/
mca_assignment3.py
159 lines (145 loc) · 4.83 KB
/
mca_assignment3.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
import torch
from torch.autograd import Variable
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from nltk.corpus import abc
import nltk
from sklearn.manifold import TSNE
import seaborn as sns
import matplotlib.pyplot as plt
import pandas as pd
window_context = 2
class CBOWModel(nn.Module):
def __init__(self, len_vocab,len_embed):
super(CBOWModel, self).__init__()
self.embed = nn.Embedding(len_vocab, len_embed)
self.layer = nn.Linear(len_embed, len_vocab)
def forward(self,x):
temp = torch.mean(self.embed(x),dim=0)
temp = temp.view((1, -1))
temp = self.layer(temp)
temp = F.log_softmax(temp)
return temp
class SkipGramModel(nn.Module):
def __init__(self, len_vocab, len_embed):
super(SkipGramModel, self).__init__()
self.embed = nn.Embedding(len_vocab, len_embed)
self.layer = nn.Linear(len_embed,len_vocab)
def forward(self, center, context):
temp_center = self.embed(center)
temp_center = temp_center.view((1, -1))
temp_context = self.embed(context)
temp_context = temp_context.view((1, -1))
temp_context = torch.t(temp_context)
#s = torch.mm(temp_center,temp_context)
log_probs = F.logsigmoid(temp_center)
return log_probs
def predict(self,word):
return self.embed(word)
def plot_tsne_skip(skip_data,skip_model,ser):
vectors = []
words = []
for in_w,out_w in skip_data:
if in_w in words:
continue
in_w_var = Variable(torch.LongTensor([w2i[in_w]]))
out_w_var = Variable(torch.LongTensor([w2i[out_w]]))
log_probs = skip_model.predict(in_w_var, out_w_var)
vectors.append(log_probs.view((-1)).detach().numpy())
words.append(in_w)
#print(vectors[0])
vectors_2 = TSNE(n_components=2).fit_transform(vectors)
plt.figure()
sns.set(rc={'figure.figsize':(15,15)})
#print(vectors_2)
#print(words)
s = sns.scatterplot(vectors_2[:,0], vectors_2[:,1],palette = 'Blues')
label_point(vectors_2[:,0],vectors_2[:,1],words,s)
plt.savefig(str(ser)+'.png')
plt.close()
def label_point(x, y, val, ax):
a = pd.concat({'x': pd.Series(x), 'y': pd.Series(y), 'val': pd.Series(val)}, axis=1)
for i, point in a.iterrows():
ax.text(point['x']+.02, point['y'], str(point['val']))
def skipgram_dataset(text):
dataset = []
temp = 2
start = temp
end = len(text)-temp
for i in range(start, end):
dataset.append((text[i], text[i-temp]))
dataset.append((text[i], text[i-(temp-1)]))
dataset.append((text[i], text[i+(temp-1)]))
dataset.append((text[i], text[i+temp]))
return dataset
def cbow_dataset(text):
dataset = []
temp = 2
start = temp
end = len(text)-temp
for i in range(start, end):
context = [text[i - temp]]
context.append(text[i - (temp-1)])
context.append(text[i + (temp-1)])
context.append(text[i + temp])
target = text[i]
dataset.append((context, target))
return dataset
def train_cbow(model):
optimizer = optim.SGD(model.parameters(), lr=lr)
for epoch in range(epochs):
total = .0
for x,y in cbow_train:
model.zero_grad()
index = [w2i[w] for w in x]
tensor = Variable(torch.LongTensor(index))
log_probs = model(tensor)
compare = Variable(torch.LongTensor([w2i[y]]))
loss = loss_fn(log_probs, compare)
loss.backward()
optimizer.step()
total += loss.data.item()
print(epoch,total_loss)
return model
def train_skipgram(model):
optimizer = optim.SGD(model.parameters(), lr=lr)
for epoch in range(epochs):
total= .0
for x, y in skipgram_train:
temp_x = Variable(torch.LongTensor([w2i[x]]))
temp_y = Variable(torch.LongTensor([w2i[y]]))
model.zero_grad()
log_probs = model(temp_x, temp_y)
compare = Variable(torch.Tensor([1]))
loss = loss_fn(log_probs[0],compare)
loss.backward()
optimizer.step()
total += loss.data.item()
print(epoch,total_loss)
plot_tsne_skip(skipgram_train[:1000],model,epoch)
return model
nltk.download('abc')
text = list(abc.words())
vocab = set(text)
vocab_size = len(vocab)
embd_size = 50
lr = 0.1
epochs = 50
hidden_size = 100
pt = 0
for word in vocab:
w2i[word] = pt
i2w[pt] = word
pt+=1
subset = text[:5000]
cbow_train = cbow_dataset(subset)
skipgram_train = skipgram_dataset(subset)
loss_fn = nn.NLLLoss()
cbow = CBOWModel(vocab_size, embd_size)
cbow_model = train_cbow(cbow)
torch.save(cbow.state_dict(), 'cbow.pth')
loss_fn = nn.MSELoss()
skip_gram = SkipGramModel(vocab_size, embd_size)
sg_model= train_skipgram(skip_gram)
torch.save(skip_gram.state_dict(), 'skipgram.pth')