Which backprop method is correct for RNN?

[csarofeen]

Accumulating loss incrementally with timestep as in the tutorial, and sending all tiemsteps to RNN seem to produce the same output/hidden/loss but loss.backwards is calculating different parameter gradients. Is there a correct and incorrect method to do this? Which is right?

import torch
from torch import nn
from torch.autograd import Variable

torch.backends.cudnn.enabled=False

model  = nn.LSTM(5, 5, 1).cuda()
model2 = nn.LSTM(5, 5, 1).cuda()

for i in range(len(model2.all_weights)):
    for j in range(len(model2.all_weights[i])):
        model2.all_weights[i][j].data.copy_(model.all_weights[i][j].data)

crit = nn.MSELoss().cuda()
crit2 = nn.MSELoss().cuda()

input = Variable(torch.randn(2,1,5).cuda())
target = Variable(torch.ones(2,1,5).cuda(), requires_grad=False)
hidden = [ Variable(torch.randn(1,1,5).cuda().fill_(0.0)),
            Variable(torch.randn(1,1,5).cuda().fill_(0.0))]

output, hidden = model(input, hidden)
loss = crit(output, target)
loss.backward(retain_variables=True)

hidden2 = [ Variable(torch.randn(1,1,5).cuda().fill_(0.0)),
            Variable(torch.randn(1,1,5).cuda().fill_(0.0))]

loss2 = 0
for i in range(input.size(0)):
    output2, hidden2 = model(input[i].view(1,1,-1), hidden2)
    loss2 += crit2(output2[0], target[i])

loss2 = loss2/2
loss2.backward(retain_variables=True)

diff = 0
max_w = 0
for i in range(len(model2.all_weights)):
    for j in range(len(model2.all_weights[i])):
        diff = max(diff, (model2.all_weights[i][j].grad - model.all_weights[i][j].grad).abs().max().data[0])
        
        max_w = max(model2.all_weights[i][j].grad.max().data[0], max_w)
        max_w = max(model.all_weights[i][j].grad.max().data[0], max_w)

dh = (hidden[0]-hidden2[0]).abs().max().data[0]
dc = (hidden[1]-hidden2[1]).abs().max().data[0]
do = (output[1]-output2).abs().max().data[0]
dl = (loss-loss2).abs().max().data[0]

print("Diff in output : " + str(do))
print("Diff in hidden states : "+str(dh) +", "+str(dc))
print("Diff in loss : " + str(dl))

print("Max weight grad found : " +str(max_w))
print("Diff in weight grad : " + str(diff))

[apaszke]

You've used the same model twice. You probably meant to use model2 in the for loop. Once you change it the diff in weight grad is 0.

[csarofeen]

Sorry about that. Thanks!