【pytorch】GPU メモリのアロケーションの謎

同じFunctionを複数回呼ぶことによってGPUメモリの増加量が予想してたのと違かったからメモしておく

1回call

    def forward(self, left, right):

        left_feature     = self.feature_extraction(left)
        right_feature  = self.feature_extraction(right)
 
        #matching
        cost_left = self.create_costvolume(left_feature,right_feature)#batch,channel,disparity,height,width
        pred_left = self.estimate_disparity(cost_left,left.size()[2],left.size()[3])
        
        return pred_left,pred_left

2回call

``` def forward(self, left, right):

    left_feature     = self.feature_extraction(left)
    right_feature  = self.feature_extraction(right)

    #matching
    cost_left = self.create_costvolume(left_feature,right_feature)#batch,channel,disparity,height,width
    pred_left = self.estimate_disparity(cost_left,left.size()[2],left.size()[3])
   
    pred_left = self.estimate_disparity(cost_left,left.size()[2],left.size()[3])

    return pred_left,pred_left

![image.png](https://qiita-image-store.s3.ap-northeast-1.amazonaws.com/0/482094/f29fc937-dde7-b57c-c3d6-51358eb1ccee.png)




<h2>3回call</h2>

def forward(self, left, right):

    left_feature     = self.feature_extraction(left)
    right_feature  = self.feature_extraction(right)

    #matching
    cost_left = self.create_costvolume(left_feature,right_feature)#batch,channel,disparity,height,width
    pred_left = self.estimate_disparity(cost_left,left.size()[2],left.size()[3])
    
    pred_left = self.estimate_disparity(cost_left,left.size()[2],left.size()[3])
    pred_left = self.estimate_disparity(cost_left,left.size()[2],left.size()[3])
   

    return pred_left,pred_left

![image.png](https://qiita-image-store.s3.ap-northeast-1.amazonaws.com/0/482094/63616c27-56f8-2543-cc60-f4d924f52621.png)



<h2>4回call</h2>

def forward(self, left, right):

    left_feature     = self.feature_extraction(left)
    right_feature  = self.feature_extraction(right)

    #matching
    cost_left = self.create_costvolume(left_feature,right_feature)#batch,channel,disparity,height,width
    pred_left = self.estimate_disparity(cost_left,left.size()[2],left.size()[3])
    
    pred_left = self.estimate_disparity(cost_left,left.size()[2],left.size()[3])
    pred_left = self.estimate_disparity(cost_left,left.size()[2],left.size()[3])
    pred_left = self.estimate_disparity(cost_left,left.size()[2],left.size()[3])       

    return pred_left,pred_left

![image.png](https://qiita-image-store.s3.ap-northeast-1.amazonaws.com/0/482094/de5ec79d-d2f7-682d-b4ba-fea7705c1881.png)


<h1>結論</h1>
2回callするまではメモリは増えるが3回目以降増えなくなった理由について、わかる人いたら教えてください