import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision.datasets
from spikingjelly.activation_based import neuron, surrogate, layer, functional
from torch.cuda import amp
import os
import argparse
from torch.utils.tensorboard import SummaryWriter
import time
import datetime
import sys
[文档]
class PlainNet(nn.Module):
def __init__(self):
super().__init__()
self.fc = nn.Sequential(
layer.Linear(28, 32),
neuron.IFNode(surrogate_function=surrogate.ATan()),
layer.Linear(32, 10),
neuron.IFNode(surrogate_function=surrogate.ATan()),
)
[文档]
def forward(self, x: torch.Tensor):
return self.fc(x).mean(0)
[文档]
class StatefulSynapseNet(nn.Module):
def __init__(self):
super().__init__()
self.fc = nn.Sequential(
layer.Linear(28, 32),
neuron.IFNode(surrogate_function=surrogate.ATan()),
layer.SynapseFilter(tau=2.0, learnable=True),
layer.Linear(32, 10),
neuron.IFNode(surrogate_function=surrogate.ATan()),
)
[文档]
def forward(self, x: torch.Tensor):
return self.fc(x).mean(0)
[文档]
class FeedBackNet(nn.Module):
def __init__(self):
super().__init__()
self.fc = nn.Sequential(
layer.Linear(28, 32),
layer.LinearRecurrentContainer(
neuron.IFNode(surrogate_function=surrogate.ATan(), detach_reset=True),
in_features=32,
out_features=32,
bias=True,
),
layer.Linear(32, 10),
neuron.IFNode(surrogate_function=surrogate.ATan()),
)
[文档]
def forward(self, x: torch.Tensor):
return self.fc(x).mean(0)
[文档]
def main():
# python -m spikingjelly.activation_based.examples.rsnn_sequential_fmnist -device cuda:0 -b 256 -epochs 64 -data-dir /datasets/FashionMNIST/ -amp -cupy -opt adam -lr 0.001 -j 8 -model plain
parser = argparse.ArgumentParser(description="Classify Sequential Fashion-MNIST")
parser.add_argument(
"-model",
default="plain",
type=str,
help='use which model, "plain", "ss" (StatefulSynapseNet) or "fb" (FeedBackNet)',
)
parser.add_argument("-device", default="cuda:0", help="device")
parser.add_argument("-b", default=128, type=int, help="batch size")
parser.add_argument(
"-epochs",
default=64,
type=int,
metavar="N",
help="number of total epochs to run",
)
parser.add_argument(
"-j",
default=4,
type=int,
metavar="N",
help="number of data loading workers (default: 4)",
)
parser.add_argument("-data-dir", type=str, help="root dir of Fashion-MNIST dataset")
parser.add_argument(
"-out-dir",
type=str,
default="./logs",
help="root dir for saving logs and checkpoint",
)
parser.add_argument("-resume", type=str, help="resume from the checkpoint path")
parser.add_argument(
"-amp", action="store_true", help="automatic mixed precision training"
)
parser.add_argument("-cupy", action="store_true", help="use cupy backend")
parser.add_argument("-opt", type=str, help="use which optimizer. SDG or Adam")
parser.add_argument("-momentum", default=0.9, type=float, help="momentum for SGD")
parser.add_argument("-lr", default=0.1, type=float, help="learning rate")
args = parser.parse_args()
print(args)
if args.model == "plain":
net = PlainNet()
elif args.model == "ss":
net = StatefulSynapseNet()
elif args.model == "fb":
net = FeedBackNet()
net.to(args.device)
# `functional.set_step_mode` will not set neurons in LinearRecurrentContainer to use step_mode = 'm'
functional.set_step_mode(net, step_mode="m")
if args.cupy:
# neurons in LinearRecurrentContainer still use step_mode = 's', so, they will still use backend = 'torch'
functional.set_backend(net, backend="cupy")
print(net)
train_set = torchvision.datasets.FashionMNIST(
root=args.data_dir,
train=True,
transform=torchvision.transforms.ToTensor(),
download=True,
)
test_set = torchvision.datasets.FashionMNIST(
root=args.data_dir,
train=False,
transform=torchvision.transforms.ToTensor(),
download=True,
)
train_data_loader = torch.utils.data.DataLoader(
dataset=train_set,
batch_size=args.b,
shuffle=True,
drop_last=True,
num_workers=args.j,
pin_memory=True,
)
test_data_loader = torch.utils.data.DataLoader(
dataset=test_set,
batch_size=args.b,
shuffle=True,
drop_last=False,
num_workers=args.j,
pin_memory=True,
)
scaler = None
if args.amp:
scaler = amp.GradScaler()
start_epoch = 0
max_test_acc = -1
optimizer = None
if args.opt == "sgd":
optimizer = torch.optim.SGD(
net.parameters(), lr=args.lr, momentum=args.momentum
)
elif args.opt == "adam":
optimizer = torch.optim.Adam(net.parameters(), lr=args.lr)
else:
raise NotImplementedError(args.opt)
lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, args.epochs)
if args.resume:
checkpoint = torch.load(args.resume, map_location="cpu")
net.load_state_dict(checkpoint["net"])
optimizer.load_state_dict(checkpoint["optimizer"])
lr_scheduler.load_state_dict(checkpoint["lr_scheduler"])
start_epoch = checkpoint["epoch"] + 1
max_test_acc = checkpoint["max_test_acc"]
out_dir = os.path.join(
args.out_dir, f"{args.model}_b{args.b}_{args.opt}_lr{args.lr}"
)
if args.amp:
out_dir += "_amp"
if args.cupy:
out_dir += "_cupy"
if not os.path.exists(out_dir):
os.makedirs(out_dir)
print(f"Mkdir {out_dir}.")
writer = SummaryWriter(out_dir, purge_step=start_epoch)
with open(os.path.join(out_dir, "args.txt"), "w", encoding="utf-8") as args_txt:
args_txt.write(str(args))
args_txt.write("\n")
args_txt.write(" ".join(sys.argv))
for epoch in range(start_epoch, args.epochs):
start_time = time.time()
net.train()
train_loss = 0
train_acc = 0
train_samples = 0
for img, label in train_data_loader:
optimizer.zero_grad()
img = img.to(args.device)
label = label.to(args.device)
# img.shape = [N, 1, H, W]
img.squeeze_(1) # [N, H, W]
img = img.permute(2, 0, 1) # [W, N, H]
# we regard [W, N, H] as [T, N, H]
label_onehot = F.one_hot(label, 10).float()
if scaler is not None:
with amp.autocast():
out_fr = net(img)
loss = F.mse_loss(out_fr, label_onehot)
scaler.scale(loss).backward()
scaler.step(optimizer)
scaler.update()
else:
out_fr = net(img)
loss = F.mse_loss(out_fr, label_onehot)
loss.backward()
optimizer.step()
train_samples += label.numel()
train_loss += loss.item() * label.numel()
train_acc += (out_fr.argmax(1) == label).float().sum().item()
functional.reset_net(net)
train_time = time.time()
train_speed = train_samples / (train_time - start_time)
train_loss /= train_samples
train_acc /= train_samples
writer.add_scalar("train_loss", train_loss, epoch)
writer.add_scalar("train_acc", train_acc, epoch)
lr_scheduler.step()
net.eval()
test_loss = 0
test_acc = 0
test_samples = 0
with torch.no_grad():
for img, label in test_data_loader:
img = img.to(args.device)
label = label.to(args.device)
# img.shape = [N, 1, H, W]
img.squeeze_(1) # [N, H, W]
img = img.permute(2, 0, 1) # [W, N, H]
# we regard [W, N, H] as [T, N, H]
label_onehot = F.one_hot(label, 10).float()
out_fr = net(img)
loss = F.mse_loss(out_fr, label_onehot)
test_samples += label.numel()
test_loss += loss.item() * label.numel()
test_acc += (out_fr.argmax(1) == label).float().sum().item()
functional.reset_net(net)
test_time = time.time()
test_speed = test_samples / (test_time - train_time)
test_loss /= test_samples
test_acc /= test_samples
writer.add_scalar("test_loss", test_loss, epoch)
writer.add_scalar("test_acc", test_acc, epoch)
save_max = False
if test_acc > max_test_acc:
max_test_acc = test_acc
save_max = True
checkpoint = {
"net": net.state_dict(),
"optimizer": optimizer.state_dict(),
"lr_scheduler": lr_scheduler.state_dict(),
"epoch": epoch,
"max_test_acc": max_test_acc,
}
if save_max:
torch.save(checkpoint, os.path.join(out_dir, "checkpoint_max.pth"))
torch.save(checkpoint, os.path.join(out_dir, "checkpoint_latest.pth"))
print(args)
print(out_dir)
print(
f"epoch = {epoch}, train_loss ={train_loss: .4f}, train_acc ={train_acc: .4f}, test_loss ={test_loss: .4f}, test_acc ={test_acc: .4f}, max_test_acc ={max_test_acc: .4f}"
)
print(
f"train speed ={train_speed: .4f} images/s, test speed ={test_speed: .4f} images/s"
)
print(
f"escape time = {(datetime.datetime.now() + datetime.timedelta(seconds=(time.time() - start_time) * (args.epochs - epoch))).strftime('%Y-%m-%d %H:%M:%S')}\n"
)
if __name__ == "__main__":
main()
