English model architecture

english/shobdo_english.py

@@ -0,0 +1,39 @@
+
+import torch.nn as nn
+import torch.nn.functional as F
+
+class ConvBnRelu(nn.Module):
+    def __init__(self,i,o,k=3,s=1,p=1):
+        super().__init__()
+        self.b=nn.Sequential(nn.Conv2d(i,o,k,s,p,bias=False),nn.BatchNorm2d(o),nn.ReLU(inplace=True))
+    def forward(self,x): return self.b(x)
+
+class LightCNN(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.b1=nn.Sequential(ConvBnRelu(1,32),ConvBnRelu(32,32),nn.MaxPool2d(2,2))
+        self.b2=nn.Sequential(ConvBnRelu(32,64),ConvBnRelu(64,64),nn.MaxPool2d(2,2))
+        self.b3=nn.Sequential(ConvBnRelu(64,128),ConvBnRelu(128,128),nn.MaxPool2d((2,1)))
+        self.b4=nn.Sequential(ConvBnRelu(128,256),ConvBnRelu(256,256),nn.MaxPool2d((2,1)))
+        self.b5=nn.Sequential(ConvBnRelu(256,256),ConvBnRelu(256,256))
+        self.pool=nn.AdaptiveAvgPool2d((1,None))
+    def forward(self,x):
+        for b in [self.b1,self.b2,self.b3,self.b4,self.b5]: x=b(x)
+        return self.pool(x).squeeze(2)
+
+class BiLSTM(nn.Module):
+    def __init__(self,i,h,o):
+        super().__init__()
+        self.rnn=nn.LSTM(i,h,bidirectional=True,batch_first=True)
+        self.fc=nn.Linear(h*2,o)
+    def forward(self,x): o,_=self.rnn(x); return self.fc(o)
+
+class Model(nn.Module):
+    def __init__(self,input_channel,output_channel,hidden_size,num_class):
+        super().__init__()
+        self.cnn=LightCNN()
+        self.rnn=nn.Sequential(BiLSTM(256,hidden_size,hidden_size),BiLSTM(hidden_size,hidden_size,num_class))
+    def forward(self,x):
+        f=self.cnn(x).permute(0,2,1)
+        o=self.rnn(f).permute(1,0,2)
+        return F.log_softmax(o,dim=2)