Saved intermediate values of the graph are freed when you call .backward() or autograd.grad()

[Salman_Abbasi]

Hi,
I am trying to run for loops inside each training iteration(part of my code, is attached below). What I would like to achieve is that for each iith and ith loop i have fresh tensors with no grads, as at end of the loop I save the output to vector, and before the next iteration(epoch), these are backpropagated,.

First of all I do not know how to make the tensor and all intermediary tensors in each of the for loops to be gradient free, after each iith loop ends. So I know there is a problem there. Second if I run this code, it runs through the first epoch and I presume it goes on through the second, but did not backpropagate the second time. I get the error “Trying to backward through the graph a second time (or directly access saved tensors after they have already been freed). Saved intermediate values of the graph are freed when you call .backward() or autograd.grad().”

Please note that I have not added the model here, as the model function works ok, if I do not have these two for loops inside the epoch(iterations). For this loop I also used retain_graph=True, it give me in_place variable error. But I do not want to use this option, and is more interested in what is happening. I appreciate assistance.

for epoch in range(epochs):
  for ii in range(seis_forward.shape[2]):# for all shots
      for i in range(tmmax_all_times.shape[2]):
          time_step=tmmax_all_times[:,:,i]
          time_step=torch.flatten(time_step)
          time_step=nxx.reshape(len(time_step),1)
      
          net_Input_tensor = torch.cat((vp_new,nxx,nzz,time_step),dim=1)
      
          x=net_Input_tensor
          u,u_xx,u_zz,u_tt,vel_pred=solve_pde_pinn_fwi(x, dnn_Linear.train(),vel_in)
      

          #loss PDE
          u=u.reshape(self.v0.T.shape)
          u_xx=u_xx.reshape(self.v0.T.shape)
          u_zz=u_zz.reshape(self.v0.T.shape)
          u_tt=u_tt.reshape(self.v0.T.shape)
  
  
  
          u_train[:,:,i,ii]=u
          u_xx_train[:,:,i,ii]=u_xx
          u_zz_train[:,:,i,ii]=u_zz
          u_tt_train[:,:,i,ii]=u_tt
          
          seis_pred[:,:,ii]=u[receivers_depth,first_rcvr_point-1:last_rcvr_point]

          vpp=vel_pred.reshape(self.v0.shape).T
          wave_analytic_from_nn=u_tt-((vpp)**2)*(u_xx+u_zz) 

          wave_analytic_from_nn_all[:,:,i,ii]= wave_analytic_from_nn
  

  PDE_Loss=loss(wave_analytic_from_nn_all,u_nn_zeros)
  seismic_loss=loss(seis_pred,seis_forward)
  
 Total_loss=seismic_loss + PDE_Loss
      
 
  optimizer.zero_grad()     
  Total_loss.backward()

[Salman_Abbasi]

In case someone is interested. The following helped me:

Detach, no_grad and requires_grad

Hello It’s general question but currently I’m looking at tutorial: http://pytorch.org/tutorials/intermediate/reinforcement_q_learning.html why in optimize_model function - we have detach() method and in select_action we use with torch.no_grad(): on the other hand doc: http://pytorch.org/docs/stable/notes/autograd.html mentions only requires_grad of course I understand we don’t want to compute gradients here - but i don’t fully understand the difference between all those 3 methods… als…

I had to detach and preferably empty all tensors used inside the for loops. I am not adding the code here. But for the for loops inside epoch loop, I have detached and zeroed tensors, before they are being computed and saved