The Training Time Problem（Solved)

[bohanfeng]

[QingbiaoLi]

Hello Bohan,

Thanks for your email.

According to the training time, you can try the following method.

1. Use SSD to store the data,
2. Use convert the N in to batch size B to speed up a bit of the training progress.

    def forward(self, inputTensor):
        B = inputTensor.shape[0] # batch size
        # N = inputTensor.shape[1]
        # C =
        (B,N,C,W,H) = inputTensor.shape
        # print(inputTensor.shape)
        # print(B,N,C,W,H)
        # B x G x N
        input_currentAgent = inputTensor.reshape(B*N,C,W,H).to(self.config.device)
        featureMap = self.ConvLayers(input_currentAgent).to(self.config.device)
        featureMapFlatten = featureMap.view(featureMap.size(0), -1).to(self.config.device)
        compressfeature = self.compressMLP(featureMapFlatten).to(self.config.device)
        extractFeatureMap_old = compressfeature.reshape(B,N,self.numFeatures2Share).to(self.config.device)
        extractFeatureMap = extractFeatureMap_old.permute([0,2,1]).to(self.config.device)
        # DCP
        for l in range(self.L):
            # \\ Graph filtering stage:
            # There is a 3*l below here, because we have three elements per
            # layer: graph filter, nonlinearity and pooling, so after each layer
            # we're actually adding elements to the (sequential) list.
            self.GFL[2 * l].addGSO(self.S) # add GSO for GraphFilter
        # B x F x N - > B x G x N,
        sharedFeature = self.GFL(extractFeatureMap)
        (_, num_G, _) = sharedFeature.shape
        sharedFeature_permute =sharedFeature.permute([0,2,1]).to(self.config.device)
        sharedFeature_stack = sharedFeature_permute.reshape(B*N,num_G)
        action_predict = self.actionsMLP(sharedFeature_stack)
        return action_predict 

[bohanfeng]

Thanks very much!