I am trying to implement momentum in my implementation of SGD with momentum. From my understanding this update look like this:
parameters -= (lr * (p.grad*0.1 + p_delta_prev*0.9))
My question is how I should store my previous deltas from every update
Here is what I have in my update function:
#we now want to do the update with momentum
#momentum takes derivative, multiplies it by 0.1, then takes the previous update,
#multiplies it by 0.9 and we add the two together
#alpha = 0.1, beta = 0.9; p-=grad*0.1 + p*0.9
def update(x,y,lr):
wd = 1e-5
y_hat = model(x)
# weight decay
w2 = 0.
for p in model.parameters(): w2 += (p**2).sum()
# add to regular loss
loss = loss_func(y_hat, y) + w2*wd
loss.backward()
with torch.no_grad():
for p in model.parameters():
#p.grad is the slope of the line of that parameter
#current_p-previous_p to get difference
p_update = (lr * (p.grad*0.1 + p*0.9))
p.sub_(p_update)
p.grad.zero_()
return loss.item()
Here the p*0.9 should be replace by the p_delta_prev. But how should I store these deltas for every parameter? If I save them to a tensor wouldn't I would be effectively copying the weight deltas to memory making my model two times the size. What would be a good way to accomplish this? I do not want to use an inbuilt function that does that activation for me. I did look into the pytorch sgd.py and it looks like the store the states.
I have updated the code:
#we now want to do the update with momentum
#momentum takes derivative, multiplys it by 0.1, then takes the previous update,
#multiplies it by 0.9 and we add the two together
#alpha = 0.1, beta = 0.9; p-=grad*0.1 + p*0.9
p_delta = {}
def update(x,y,lr):
wd = 1e-5
y_hat = model(x)
# weight decay
w2 = 0.
for p in model.parameters(): w2 += (p**2).sum()
# add to regular loss
loss = loss_func(y_hat, y) + w2*wd
loss.backward()
with torch.no_grad():
i = 0
for p in model.parameters():
#p.grad is the slope of the line of that parameter
if i not in p_delta:#check if key exists
p_delta[i] = torch.zeros_like(p)
p_update = (lr *p.grad) + (p_delta[i]*0.9)
p_delta[i] = p_update.clone()
p.sub_(p_update)
p.grad.zero_()
print((p_delta[i]))
i+=1
return loss.item()
I think the code in the Excel spreadsheet is incorrect. Jeremy seems to show: lr* ((p.grad*0.1) + (p_delta[i]*0.9)) but many tutorials seem to show: (lr *p.grad) + (p_delta[i]*0.9) If we implement Jeremy’s code the loss actually is slower than vanilla GD. The part of the video is here: https://youtu.be/CJKnDu2dxOE?t=6581
