File size: 5,894 Bytes
34393ef
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
import sys
import torch
from torch import nn
import numpy as np
from .CNN_models import Conv_AE,Conv_VAE,cal_conv_shape
from .Self_attention import self_attention

class Baseline(nn.Module):
    def __init__(self,channel_ls,padding_ls,diliat_ls,latent_dim,kernel_size,dropout_ls,num_label,loss_fn,pad_to):
        """
        Conv - Dense Framework DL Regressor to predict ribosome load (rl) from 5' UTR sequence
        Arguments:
        ...channel_ls,padding_ls,diliat_ls,latent_dim,kernel_size,num_label : parameter to define the Conv layers
        ...loss_fn : regression loss function `MSELoss` or `MyHingeLoss`
        """
        super(Baseline,self).__init__()
        #         ==<<|  properties  |>>==
        self.kernel_size = kernel_size
        self.loss_fn = loss_fn
        self.channel_ls = channel_ls
        self.padding_ls = padding_ls
        self.diliat_ls =  diliat_ls
        self.L_in = pad_to
        self.loss_dict_keys = ['Total','MAE','R.M.S.E']
        # the basic element block of CNN
        
        #         ==<<|  Conv layers  |>>==
        # 3 layers in the 5'UTR paper
        self.encoder = nn.ModuleList(
            [self.Conv_block(channel_ls[i],channel_ls[i+1],padding_ls[i],diliat_ls[i],dropout_rate=dropout_ls[i]) for i in range(len(channel_ls)-1)]
        )        
        
        # compute the output shape of conv layer
        self.out_len = int(self.compute_out_dim(kernel_size,L_in=self.L_in))
        self.out_dim = self.out_len * channel_ls[-1]
        
        #         ==<<|  Dense layers  |>>==
        self.fc_out = nn.Sequential( 
            nn.Linear(self.out_dim,40),
            nn.Dropout(0.2),
            nn.ReLU(),
            
            nn.Linear(40,num_label)
        )
        
        self.define_loss()  
        self.acc_hinge = MyHingeLoss(None)
        
        self.apply(self.weight_initialize)
        self.rl_posi = None       # single output
        
    def weight_initialize(self, model):
        if type(model) in [nn.Linear]:
        	nn.init.xavier_uniform_(model.weight)
            nn.init.zeros_(model.bias)
        elif type(model) in [nn.LSTM, nn.RNN, nn.GRU]:
            nn.init.orthogonal_(model.weight_hh_l0)
            nn.init.xavier_uniform_(model.weight_ih_l0)
            nn.init.zeros_(model.bias_hh_l0)
            nn.init.zeros_(model.bias_ih_l0)
        elif isinstance(model, nn.Conv1d):
            nn.init.kaiming_normal_(model.weight, nonlinearity='leaky_relu',)
        elif isinstance(model, nn.BatchNorm1d):
            nn.init.constant_(model.weight, 1)
            nn.init.constant_(model.bias, 0)
    
    def Conv_block(self,inChan,outChan,padding,diliation,stride=1,dropout_rate=0): 
        """
        Building Block of stack conv
        """
        net = nn.Sequential(
                    nn.Conv1d(inChan,outChan,self.kernel_size,stride=1,padding=padding,dilation=diliation),
                    # nn.BatchNorm1d(outChan),
                    nn.ReLU(),
                    nn.Dropout(dropout_rate))
        return net
    
    def define_loss(self):
        if self.loss_fn in ['mse','MSE']:
            self.regression_loss = nn.MSELoss(reduction='mean')
        else:
            self.regression_loss = MyHingeLoss(reduction='mean')
        
    
    def encode(self,X):
        
        X = X.transpose(1,2)  # to B*4*100
        Z = X
        for model in self.encoder:
            Z = model(Z)
        return Z
    
    def forward(self,X):
        """
        Conv -> flatten -> Dense
        """
        batch_size = X.shape[0]
        
        z = self.encode(X)
        z_flat = z.view(batch_size,-1)
        out = self.fc_out(z_flat)       # no activation for the last layer
        
        return out
        
        
    def compute_acc(self,out,X,Y,popen):
        """
        for this regression task, accuracy  is the percentage that prediction error < epsilon (Lambda) 
        """
        
        epsilon = popen.epsilon
        
        batch_size = Y.shape[0]
        rl_pred = out[self.rl_posi] if type(out) == tuple else out
        if rl_pred.shape != Y.shape:
            if len(rl_pred.shape) == 2:
                rl_pred = rl_pred.squeeze()
            if len(Y.shape) == 2:
                Y = Y.squeeze()
        with torch.no_grad(): 
            loss = self.acc_hinge(rl_pred,Y,epsilon=0.3)
            n_inrange = (loss==0).squeeze().sum().item()
        
        return {"Acc":n_inrange / batch_size}
       
        
    def compute_loss(self,out,X,Y,popen):
        """
        it's termed `chimela_loss` to keep compatability with MTL MODELS (the same dataset was used)
        `chiemela_loss` requires three input : 
        ...out:
        ...Y:
        ...Lambda : which is the epsilon of hinge loss if possible
        """
        self.Lambda = popen.chimerla_weight
        batch_size = Y.shape[0]
        rl_pred = out[self.rl_posi] if type(out) == tuple else out
        
        #  Hinge or MSE
        if self.loss_fn in ['mse','MSE']:
            loss = self.regression_loss(rl_pred,Y)
        else:
            loss = self.regression_loss(rl_pred,Y,self.Lambda).squeeze()
        with torch.no_grad():
            MAE = torch.abs(rl_pred-Y).mean()                                  # Mean Absolute Error
            RMSE = torch.sqrt( torch.sum((rl_pred-Y)**2) / batch_size)         # Root Mean Square Error
        return {"Total":loss,"MAE":MAE,"R.M.S.E":RMSE}
    
    def compute_out_dim(self,kernel_size,L_in = 100,num_layer=None):
        """
        manually compute the final length of convolved sequence
        """
        loop_times = num_layer if type(num_layer) == int else len(self.channel_ls)-1
        for i in range(loop_times):
            L_out = cal_conv_shape(L_in,kernel_size,stride=1,padding=self.padding_ls[i],diliation=self.diliat_ls[i])
            L_in = L_out
        return L_out