From b29390b84f9ae7c6bb6eb2dc0211057fb0aaf939 Mon Sep 17 00:00:00 2001
From: Hammouda Elbez <hammouda.elbez@univ-lille.fr>
Date: Mon, 17 Jul 2023 11:41:43 +0200
Subject: [PATCH] CIFAR10 Norse sim updated
---
Norse/CIFAR-10/CIFAR10_conv_bench.py | 37 ++++++++++++++++------------
1 file changed, 21 insertions(+), 16 deletions(-)
diff --git a/Norse/CIFAR-10/CIFAR10_conv_bench.py b/Norse/CIFAR-10/CIFAR10_conv_bench.py
index b75a869..0f3ffbc 100644
--- a/Norse/CIFAR-10/CIFAR10_conv_bench.py
+++ b/Norse/CIFAR-10/CIFAR10_conv_bench.py
@@ -84,21 +84,20 @@ for maxTh, Alpha, reinforcement, compression, layerwise in np.array(list(itertoo
def __init__(self, num_channels=3, feature_size=32, method="super", alpha=100):
super(ConvNet, self).__init__()
- self.features = int(((feature_size - 4) / 2 - 4) / 2)
- self.conv1_out_channels = 32
- self.conv2_out_channels = 128
+ self.conv1_out_channels = 128
+ self.conv2_out_channels = 256
self.fc1_out_channels = 1024
self.out_channels = 10
self.conv1 = torch.nn.Conv2d(num_channels, self.conv1_out_channels, 5, 1, bias=False)
self.conv2 = torch.nn.Conv2d(self.conv1_out_channels, self.conv2_out_channels, 5, 1, bias=False)
- self.fc1 = torch.nn.Linear(self.features**2 * self.conv2_out_channels, self.fc1_out_channels, bias=False)
- self.lif0 = LIFCell(p=LIFParameters(method=method, alpha=alpha, v_th=0.25))
- self.lif1 = LIFCell(p=LIFParameters(method=method, alpha=alpha, v_th=0.25))
- self.lif2 = LIFCell(p=LIFParameters(method=method, alpha=alpha, v_th=0.25))
- self.out = LILinearCell(self.fc1_out_channels, self.out_channels)
+ self.fc1 = torch.nn.Linear(256*5*5, self.fc1_out_channels, bias=False)
+ self.fc2 = torch.nn.Linear(self.fc1_out_channels, 1000, bias=False)
+ self.lif0 = LIFCell(p=LIFParameters(method=method, alpha=alpha, v_th=0.3))
+ self.lif1 = LIFCell(p=LIFParameters(method=method, alpha=alpha, v_th=0.3))
+ self.lif2 = LIFCell(p=LIFParameters(method=method, alpha=alpha, v_th=0.3))
+ self.lif3 = LIFCell(p=LIFParameters(method=method, alpha=alpha, v_th=0.3))
+ self.out = LILinearCell(1000, self.out_channels,p=LIFParameters(method=method, alpha=alpha, v_th=0.3))
- #LILinearCell(self.fc1_out_channels, self.out_channels)
-
def forward(self, x):
seq_length = x.shape[0]
batch_size = x.shape[1]
@@ -111,16 +110,22 @@ for maxTh, Alpha, reinforcement, compression, layerwise in np.array(list(itertoo
)
for ts in range(seq_length):
- z = self.conv1(x[ts, :])
+ z = torch.nn.functional.relu(self.conv1(x[ts, :]))
+
z, s0 = self.lif0(z, s0)
z = torch.nn.functional.max_pool2d(z, 2, 2)
- z = self.out_channels * self.conv2(z)
+ z = 3 * torch.nn.functional.relu(self.conv2(z))
+
z, s1 = self.lif1(z, s1)
z = torch.nn.functional.max_pool2d(z, 2, 2)
- z = z.view(-1, self.features**2 * self.conv2_out_channels)
+
+ z = z.view(-1, 256*5*5)
z = self.fc1(z)
z, s2 = self.lif2(z, s2)
- v, so = self.out(torch.nn.functional.relu(z), so)
+ z = self.fc2(z)
+ z, s3 = self.lif3(z, s3)
+ v, so = self.out(z, so)
+
voltages[ts, :, :] = v
return voltages
@@ -171,7 +176,7 @@ for maxTh, Alpha, reinforcement, compression, layerwise in np.array(list(itertoo
torch.autograd.set_detect_anomaly(True)
T = 35
- LR = 0.001
+ LR = 3e-5
EPOCHS = 100 # Increase this for improved accuracy
if torch.cuda.is_available():
@@ -180,7 +185,7 @@ for maxTh, Alpha, reinforcement, compression, layerwise in np.array(list(itertoo
DEVICE = torch.device("cpu")
model = Model(
- encoder=encode.SpikeLatencyLIFEncoder(T), snn=ConvNet(alpha=80), decoder=decode).to(DEVICE)
+ encoder=encode.SpikeLatencyLIFEncoder(T,p=LIFParameters(v_th=0.3)), snn=ConvNet(alpha=80), decoder=decode).to(DEVICE)
optimizer = torch.optim.Adam(model.parameters(), lr=LR)
--
GitLab