imirzadeh / stable-continual-learning Goto Github PK

In the scripts replicate_experiment_2.sh here, to run the split-CIFAR experiments, the command is

python -m stable_sgd.main --dataset cifar100 --tasks 20 --epochs-per-task 1 --lr 0.1 --gamma 0.8 --hiddens 256 --batch-size 10 --dropout 0.5 --seed 1234

which only provides results around 2x% average accuracy

I found in a closed issue here that the setting should be

python -m stable_sgd.main --dataset cifar100 --tasks 20 --epochs-per-task 1 --lr 0.15 --gamma 0.85 --batch-size 10 --dropout 0.1 --seed 1234

Maybe the scripts need some updates @imirzadeh?

I think that something is missed in the Resnet18 implementation:

  class BasicBlock (nn.Module):
	expansion = 1

	def __init__(self, in_planes, planes, stride=1, config={}):
		super(BasicBlock, self).__init__()
		self.conv1 = conv3x3(in_planes, planes, stride)
		self.conv2 = conv3x3(planes, planes)

		self.shortcut = nn.Sequential()
		if stride != 1 or in_planes != self.expansion * planes:
			self.shortcut = nn.Sequential(
				nn.Conv2d(in_planes, self.expansion * planes, kernel_size=1,
						  stride=stride, bias=False),
			)
		self.IC1 = nn.Sequential(
			nn.BatchNorm2d(planes),
			nn.Dropout(p=config['dropout'])
			)

		self.IC2 = nn.Sequential(
			nn.BatchNorm2d(planes),
			nn.Dropout(p=config['dropout'])
			)

	def forward(self, x):
		out = self.conv1(x)
		out = relu(out)
		out = self.IC1(out)

		out += self.shortcut(x)
		out = relu(out)
		out = self.IC2(out)
		return out

Two attributes are defined in the class BasicBlock. They represent the classical convolution operations that are used for Resnet-X architectures:

self.conv1 = conv3x3(in_planes, planes, stride)
self.conv2 = conv3x3(planes, planes)

The problem is that in the forward pass only the first one (i.e. self.conv1) and the two BatchNormalization Layers are used to compute the output. Furthermore, when the model is load in the gpu both conv1 and conv2 are moved into it and the second one is unused in the forward. So i think that the code of the forward pass should be:

def forward(self, x):
          out = self.conv1(x)
          out = relu(out)
          out = self.IC1(out)
          out = self.conv2(out)
          out = self.IC2(out)
          
          out += self.shortcut(x)
          out = relu(out)
          return out

`
Fixing this problem, i am not able to reproduce the results of the paper. On cifar 100 using the hyperparameters provided in this closed issue :

--dataset cifar100 --tasks 20 --epochs-per-task 1 --lr 0.15 --gamma 0.85 --batch-size 10 --dropout 0.1 --seed 1234

average accuracy = 51.339999999999996, forget = 0.11200000000000002

If use instead the hyperparameters provided in replicate_experiment_2.sh

--dataset cifar100 --tasks 20 --epochs-per-task 1 --lr 0.1 --gamma 0.8 --hiddens 256 --batch-size 10 --dropout 0.5 --seed 1234

average accuracy = 44.550000000000004, forget = 0.05421052631578946

These results differ much from the results reported in the paper:

average_accuracy=59.9 and forgetting=0.08

Could you provide the correct hyperparameters?

Hi,

Thanks for sharing the code. However I'm not able to reproduce the paper results in experiment 2 particularly. I downloaded the code and ran bash replicate_experiment_2.sh as per the instructions. It seems like Stable SGD always returns Avg accuracy less than or approximately equal to the Naive SGD for all three datasets. I also tried varying hyper-parameters such as gamma, lr and dropout etc. but the Stable SGD results did not come near to ones given in the Paper.

Is there really a bug in replicate_experiment_2.sh or it is my misunderstanding? I was wondering if you could provide any hints for reproducing the results.

Thanks in advance.

Hello,

Thanks for sharing the code.
However, I've had trouble reproducing the results, starting with experiment 1.
I've set up a fresh Python 3.7 environment and installed dependencies with bash setup_and_install.sh. Then I've run bash replicate_experiment_1.sh and here are the results I've got:

• rotated mnist, naive: 87.0, stable: 86.8 (average accuracy, averaged over 3 runs). According to the paper, stable should be > 90 (Figure 3c)
• permuted mnist, naive: 77.7, stable: 61.7. According to the paper, stable should be > 90.

Any hints for reproducing the results would be very helpful.

Hello,

Thanks for sharing the code.
However, I've had trouble reproducing the results of the naive method on permuted mnist, starting with experiment 1.
Here are my results:
ACC: 82.26, Forgetting: 0.18. Actually, as shown in Table 2 in the original manuscript, the naive version should achieve 44.4% in ACC and 0.53% in forgetting.
Could you share more details about the naive method in Table 2?
Thanks very much!