Skip to content

Fix MAML graph breaks during Inductor inference #2709

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
msnliu wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Fix MAML graph breaks during Inductor inference #2709

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions torchbenchmark/models/maml/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,6 +23,7 @@ def __init__(self, test, device, batch_size=None, extra_args=[]):
super().__init__(
test=test, device=device, batch_size=batch_size, extra_args=extra_args
)
torch._dynamo.config.trace_autograd_ops = True

# load from disk or synthesize data
use_data_file = False
Expand Down
13 changes: 6 additions & 7 deletions torchbenchmark/models/maml/learner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -145,7 +145,7 @@ def extra_repr(self):

return info

def forward(self, x, vars=None, bn_training=True):
def forward(self, x, vars=None, vars_bn=None, bn_training=True):
"""
This function can be called by finetunning, however, in finetunning, we dont wish to update
running_mean/running_var. Thought weights/bias of bn == updated, it has been separated by fast_weights.
Expand All @@ -157,8 +157,10 @@ def forward(self, x, vars=None, bn_training=True):
:return: x, loss, likelihood, kld
"""

if vars == None:
if vars is None:
vars = self.vars
if vars_bn is None:
vars_bn = self.vars_bn

idx = 0
bn_idx = 0
Expand All @@ -183,10 +185,7 @@ def forward(self, x, vars=None, bn_training=True):
# print('forward:', idx, x.norm().item())
elif name == "bn":
w, b = vars[idx], vars[idx + 1]
running_mean, running_var = (
self.vars_bn[bn_idx],
self.vars_bn[bn_idx + 1],
)
running_mean, running_var = vars_bn[bn_idx], vars_bn[bn_idx + 1]
x = F.batch_norm(
x, running_mean, running_var, weight=w, bias=b, training=bn_training
)
Expand Down Expand Up @@ -219,7 +218,7 @@ def forward(self, x, vars=None, bn_training=True):

# make sure variable == used properly
assert idx == len(vars)
assert bn_idx == len(self.vars_bn)
assert bn_idx == len(vars_bn)

return x

Expand Down
84 changes: 37 additions & 47 deletions torchbenchmark/models/maml/meta.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,3 @@
from copy import deepcopy

import torch
from torch import nn, optim
from torch.nn import functional as F
Expand Down Expand Up @@ -76,60 +74,60 @@ def forward_train(self, x_spt, y_spt, x_qry, y_qry):
0 for _ in range(self.update_step + 1)
] # losses_q[i] is the loss on step i
corrects = [0 for _ in range(self.update_step + 1)]
params = tuple(self.net.parameters())
vars_bn = tuple(self.net.vars_bn)

for i in range(task_num):
# 1. run the i-th task and compute loss for k=0
logits = self.net(x_spt[i], vars=None, bn_training=True)
# 1. compute the gradient for the i-th task
logits = self.net(x_spt[i], params, vars_bn, bn_training=True)
loss = F.cross_entropy(logits, y_spt[i])
grad = torch.autograd.grad(loss, self.net.parameters())
grad = torch.autograd.grad(loss, params)
fast_weights = list(
[p[1] - self.update_lr * p[0] for p in zip(grad, self.net.parameters())]
[p[1] - self.update_lr * p[0] for p in zip(grad, params)]
)

# this is the loss and accuracy before first update
with torch.no_grad():
# [setsz, nway]
logits_q = self.net(x_qry[i], self.net.parameters(), bn_training=True)
logits_q = self.net(x_qry[i], params, vars_bn, bn_training=True)
loss_q = F.cross_entropy(logits_q, y_qry[i])
losses_q[0] += loss_q

pred_q = F.softmax(logits_q, dim=1).argmax(dim=1)
correct = torch.eq(pred_q, y_qry[i]).sum().item()
corrects[0] = corrects[0] + correct
corrects[0] = corrects[0] + torch.eq(pred_q, y_qry[i]).sum()

# this is the loss and accuracy after the first update
with torch.no_grad():
# [setsz, nway]
logits_q = self.net(x_qry[i], fast_weights, bn_training=True)
logits_q = self.net(x_qry[i], fast_weights, vars_bn, bn_training=True)
loss_q = F.cross_entropy(logits_q, y_qry[i])
losses_q[1] += loss_q
# [setsz]
pred_q = F.softmax(logits_q, dim=1).argmax(dim=1)
correct = torch.eq(pred_q, y_qry[i]).sum().item()
corrects[1] = corrects[1] + correct
corrects[1] = corrects[1] + torch.eq(pred_q, y_qry[i]).sum()

for k in range(1, self.update_step):
# 1. run the i-th task and compute loss for k=1~K-1
logits = self.net(x_spt[i], fast_weights, bn_training=True)
# 1. compute grad on theta_pi
logits = self.net(
x_spt[i], fast_weights, vars_bn, bn_training=True
)
loss = F.cross_entropy(logits, y_spt[i])
# 2. compute grad on theta_pi
grad = torch.autograd.grad(loss, fast_weights)
# 3. theta_pi = theta_pi - train_lr * grad
# 2. theta_pi = theta_pi - train_lr * grad
fast_weights = [
p[1] - self.update_lr * p[0] for p in zip(grad, fast_weights)
]

logits_q = self.net(x_qry[i], fast_weights, bn_training=True)
logits_q = self.net(x_qry[i], fast_weights, vars_bn, bn_training=True)
# loss_q will be overwritten and just keep the loss_q on last update step.
loss_q = F.cross_entropy(logits_q, y_qry[i])
losses_q[k + 1] += loss_q

with torch.no_grad():
pred_q = F.softmax(logits_q, dim=1).argmax(dim=1)
correct = (
torch.eq(pred_q, y_qry[i]).sum().item()
) # convert to numpy
corrects[k + 1] = corrects[k + 1] + correct
corrects[k + 1] = (
corrects[k + 1] + torch.eq(pred_q, y_qry[i]).sum()
)

# end of all tasks
# sum over all losses on query set across all tasks
Expand All @@ -143,7 +141,7 @@ def forward_train(self, x_spt, y_spt, x_qry, y_qry):
# print(torch.norm(p).item())
self.meta_optim.step()

accs = torch.tensor(corrects) / (querysz * task_num)
accs = torch.stack(corrects).to(torch.float32) / (querysz * task_num)

return accs

Expand All @@ -160,61 +158,53 @@ def finetunning(self, x_spt, y_spt, x_qry, y_qry):

corrects = [0 for _ in range(self.update_step_test + 1)]

# in order to not ruin the state of running_mean/variance and bn_weight/bias
# we finetunning on the copied model instead of self.net
net = deepcopy(self.net)
# Keep the adapted task's BatchNorm state separate from the base model.
net = self.net
params = tuple(net.parameters())
vars_bn = tuple(buffer.detach().clone() for buffer in net.vars_bn)

# 1. run the i-th task and compute loss for k=0
logits = net(x_spt)
# 1. compute the gradient for the task
logits = net(x_spt, params, vars_bn, bn_training=True)
loss = F.cross_entropy(logits, y_spt)
grad = torch.autograd.grad(loss, net.parameters())
grad = torch.autograd.grad(loss, params)
fast_weights = list(
map(lambda p: p[1] - self.update_lr * p[0], zip(grad, net.parameters()))
map(lambda p: p[1] - self.update_lr * p[0], zip(grad, params))
)

# this is the loss and accuracy before first update
with torch.no_grad():
# [setsz, nway]
logits_q = net(x_qry, net.parameters(), bn_training=True)
logits_q = net(x_qry, params, vars_bn, bn_training=True)
# [setsz]
pred_q = F.softmax(logits_q, dim=1).argmax(dim=1)
# scalar
correct = torch.eq(pred_q, y_qry).sum().item()
corrects[0] = corrects[0] + correct
corrects[0] = corrects[0] + torch.eq(pred_q, y_qry).sum()

# this is the loss and accuracy after the first update
with torch.no_grad():
# [setsz, nway]
logits_q = net(x_qry, fast_weights, bn_training=True)
logits_q = net(x_qry, fast_weights, vars_bn, bn_training=True)
# [setsz]
pred_q = F.softmax(logits_q, dim=1).argmax(dim=1)
# scalar
correct = torch.eq(pred_q, y_qry).sum().item()
corrects[1] = corrects[1] + correct
corrects[1] = corrects[1] + torch.eq(pred_q, y_qry).sum()

for k in range(1, self.update_step_test):
# 1. run the i-th task and compute loss for k=1~K-1
logits = net(x_spt, fast_weights, bn_training=True)
# 1. compute grad on theta_pi
logits = net(x_spt, fast_weights, vars_bn, bn_training=True)
loss = F.cross_entropy(logits, y_spt)
# 2. compute grad on theta_pi
grad = torch.autograd.grad(loss, fast_weights)
# 3. theta_pi = theta_pi - train_lr * grad
fast_weights = list(
map(lambda p: p[1] - self.update_lr * p[0], zip(grad, fast_weights))
)

logits_q = net(x_qry, fast_weights, bn_training=True)
logits_q = net(x_qry, fast_weights, vars_bn, bn_training=True)
# loss_q will be overwritten and just keep the loss_q on last update step.
loss_q = F.cross_entropy(logits_q, y_qry)

with torch.no_grad():
pred_q = F.softmax(logits_q, dim=1).argmax(dim=1)
correct = torch.eq(pred_q, y_qry).sum().item() # convert to numpy
corrects[k + 1] = corrects[k + 1] + correct

del net
corrects[k + 1] = corrects[k + 1] + torch.eq(pred_q, y_qry).sum()

accs = torch.tensor(corrects) / querysz
accs = torch.stack(corrects).to(torch.float32) / querysz

return accs

Expand Down