yigitdemirag/srnn-pcm

PyTorch Implementation of Online Training of Spiking Recurrent Neural Networks with Phase-Change Memory Synapses

srnn-pcm's Issues

RESET function update

We noticed a minor bug in the initial version of the 'xbar.reset()` function. This bug is an implementation error, and the paper results are obtained using this old version of the function.

Specifically, the bug causes the device to prevent reinitialization of the programming time variable,tp, to the time of application of the RESET pulse and instead reinitialises tp to 0.

This bug is not so critical because, during e-prop training experiments, we always SET the PCM device just after applying a RESET pulse, which has been updating (hence correcting) the tp variable.

Old version (used in the paper)

    def reset(self, mask, G0=0.1):
        self.Pmem[mask] = 1
        self.tp[mask] = 0 
        self.count[mask] = 0
        self.G[mask] = torch.normal(G0, G0 * 0.1, ((torch.sum(mask),)), device=self.device).clamp(1e-2, self.Gmax)
        self.tracker[mask[:,0,:,:]] = 0 

Corrected version

    def reset(self, tp, mask, G0=0.1):
        ''' Emulates a masked RESET operation on the PCM crossbar.

        Args:
            mask : Mask for selecting PCM device
            G0   : Initial conductance value
        '''
        self.Pmem[mask] = 1
        self.tp[mask] = tp # Please read Issue #1 for more details regarding the change in this line for replication of the results. 
        self.count[mask] = 0
        self.G[mask] = torch.normal(G0, G0 * 0.1, ((torch.sum(mask),)), device=self.device).clamp(1e-2, self.Gmax)
        self.tracker[mask[:,0,:,:]] = 0