-
Notifications
You must be signed in to change notification settings - Fork 1
/
UNET_v1.py
192 lines (176 loc) · 7.48 KB
/
UNET_v1.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
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import AutoTokenizer, RobertaModel
import warnings
warnings.filterwarnings('ignore')
#Frame wise input to transformer block but with splits according to actual word length
tokenizer = AutoTokenizer.from_pretrained("roberta-base")
model = RobertaModel.from_pretrained("roberta-base")
class PositionalEncoding(nn.Module):
def __init__(self, d_model= 256, dropout: float = 0.1, max_len: int = 5000):
super().__init__()
self.dropout = nn.Dropout(p=dropout)
position = torch.arange(max_len).unsqueeze(1)
div_term = torch.exp(torch.arange(0, d_model, 2) * (-math.log(10000.0) / d_model))
self.pe = torch.zeros(max_len, d_model)
self.pe[:, 0::2] = torch.sin(position * div_term)#for even terms sin
self.pe[:, 1::2] = torch.cos(position * div_term)#for odd terms cos
self.b = nn.Linear(in_features=d_model, out_features=d_model)
self.c = nn.Linear(in_features=d_model, out_features=d_model)
def forward(self, t) :
return self.c(self.b(self.pe[:t.size(0)]))
class MHAVEc(nn.Module):
def __init__(self,embVecSize,cross= False):
super().__init__()
# self.state = nn.Linear(C,embVecSize)
self.emb = embVecSize
self.state = nn.LazyLinear(self.emb)
self.cross = cross
def forward(self,InputEmbedding,encoderOutput=None):
query = self.state(InputEmbedding)
key = self.state(InputEmbedding)
value = self.state(InputEmbedding)
if self.cross :
assert encoderOutput != None
key = encoderOutput
value = encoderOutput
multihead_attn = nn.MultiheadAttention(self.emb, 8)
attn_output, _ = multihead_attn(query, key, value)
return attn_output
class Encoder(nn.Module):
def __init__(self,embeddingSize):
super().__init__()
self.mha = MHAVEc(embeddingSize)
self.AddNorm = nn.LayerNorm(embeddingSize)
self.FeedForward = nn.Sequential(
nn.Linear(embeddingSize,80),
nn.ReLU(),
nn.Linear(80,embeddingSize)
)
def forward(self,input):
o1 = self.mha(input)
o2 = self.AddNorm(o1+input)
o3 = self.FeedForward(o2)
o4 = self.AddNorm(o3+o2)
return o4
class Decoder(nn.Module):
def __init__(self ,embeddingSize):
super().__init__()
self.mhaCross = MHAVEc(embeddingSize,cross = True)
self.mha = MHAVEc(embeddingSize)
self.AddNorm = nn.LayerNorm(embeddingSize)
self.FeedForward = nn.Sequential(
nn.Linear(embeddingSize,80),
nn.ReLU(),
nn.Linear(80,embeddingSize)
)
def forward(self,x,encoderOuput):
o1 = self.mha(x)
# print(o1.shape)
# print(x.shape)
o2 = self.AddNorm(o1+x)
# print(o2.shape)
o3 = self.mhaCross(o2,encoderOuput)
# print(o3.shape)
o4 = self.AddNorm(o2+o3)
# print(o4.shape)
o5 = self.FeedForward(o4)
# print(o5.shape)
o6 = self.AddNorm(o5+o4)
# print(o6.shape)
return o6
class TransFrameWise(nn.Module):
""" Frame wise input to transformer block but with splits according to actual word length """
def __init__(self,embVecSize= 256,RobertaWordEmbSize = 768,Encoder=Encoder,Decoder=Decoder,training=True,inference = False):
super().__init__()
self.training = training
self.inference = inference
self.embVecSize = embVecSize
self.dense_1 = nn.LazyLinear(self.embVecSize)
self.dense11 = nn.LazyLinear(self.embVecSize)
self.dense_2 = nn.Linear(self.embVecSize,1)
self.act_fn = nn.SiLU()
self.RobertaWordEmbSize = RobertaWordEmbSize
self.lossf = nn.MSELoss()
self.EncSplitsReq = []
self.dense_3 = nn.Linear(80,1)
self.enc = Encoder(self.embVecSize)
self.dec = Decoder(self.embVecSize)
self.IPpositionalEncoding = PositionalEncoding(self.embVecSize)
def forward(self,TargetSequnce,input):
decoderInput = []
for i in TargetSequnce:
self.EncSplitsReq.append(len(i.split()))
inputs = tokenizer(i, return_tensors="pt")
outputs = model(**inputs)
decoderInput.append(outputs[1])
decIn = torch.stack(decoderInput)
decoderBatchedinput = []
for ind,i in enumerate(decIn):
words = self.EncSplitsReq[ind]
print(words)
outputEmbs = torch.split(i,int(self.RobertaWordEmbSize/words),dim=1)
o = torch.cat(outputEmbs)
_,e = o.shape
if e < self.embVecSize:
o = F.pad(o,(0,self.embVecSize-e))
else:
o = self.dense_1(o)
decoderBatchedinput.append(o)
print(o.shape,'decoder input for all sequences')
bs,C,_ = input.shape
inputs = []
for ind,a in enumerate(input):
#diving input spec to frames
i = a.view(self.EncSplitsReq[ind],C,-1)
_,_,e = i.shape
# i = i.repeat(1,1,int(self.embVecSize/e))
if e < self.embVecSize:
i = F.pad(i,(0,self.embVecSize-e))
else:
i = self.dense11(i)
i += self.dense_2(self.act_fn(self.IPpositionalEncoding(i)))[:, :, None]
b = self.enc(i)
print(b.shape,'encoder output for all squences')
b = self.dense_3(b.permute(0,2,1)).squeeze()
print(b.shape,'encoder hidden state for all squences')
inputs.append(b)
##inputs mei encoder ki hidden states hai framewise and decoderBatchedinput mei decoder to be input word emebdding hai word by word
dec = Decoder(self.embVecSize)
#for training time
for i in range(bs):
words,_ = decoderBatchedinput[i].shape
print(words)
for j in range(words):
if self.training:
TextEMbedding = decoderBatchedinput[i][j].unsqueeze(0)
EncOUT = inputs[i][j].unsqueeze(0)
predicted = dec(TextEMbedding,EncOUT)
print(predicted.shape)
if (j+1 < words):
actual = decoderBatchedinput[i][j+1]
loss = self.lossf(actual,predicted)
print(loss)
# loss.backward()
# #for inference time
if self.inference:
sos = torch.randn((1,256))
EncOUT = inputs[i][j].unsqueeze(0)
pred1 = dec(sos,EncOUT)
subPreds = pred1
print(subPreds.shape)
# for i in range(words-1):
# EncOUT = EncoderHiddenState[i]
# pred1 = dec(sos,EncOUT)
# subPreds = pred1
# print(subPreds.shape)
# for i in range(words-1):
# EncOUT = EncoderHiddenState[i+1] #3,256
# print(subPreds.shape,'-*syb')
# pred = dec(subPreds,EncOUT)
# print(pred.shape,'pred')
# subPreds = torch.cat((subPreds,pred))
BatchTargetSequence = ['my name is ritul and i live here','i study math here','hi hello']
input = torch.randn((3,80,800))