Visualizations for CNN trained for timeseries classification

[Yannis Kalfas]

Would it be possible to handle networks trained for timeseries classification? Example CNN below:

class Cnn1d(nn.Module):

    def __init__(self, outputs=2):
        super(Cnn1d, self).__init__()
        self.outputs = outputs
        self.conv1 = nn.Conv1d(1, 16, 3)
        self.bn1 = nn.BatchNorm1d(16)
        self.pool1 = nn.MaxPool1d(2)

        self.conv2 = nn.Conv1d(16, 32, 3)
        self.bn2 = nn.BatchNorm1d(32)
        self.pool2 = nn.MaxPool1d(2)

        self.conv3 = nn.Conv1d(32, 64, 3)
        self.bn3 = nn.BatchNorm1d(64)
        self.pool3 = nn.MaxPool1d(2)

        self.conv4 = nn.Conv1d(64, 128, 3)
        self.bn4 = nn.BatchNorm1d(128)
        self.pool4 = nn.MaxPool1d(2) 

        self.dropout = nn.Dropout()
        self.avgPool = nn.AvgPool1d(127)
        self.fc1 = nn.Linear(256, self.outputs)

    def forward(self, x):

        x = self.conv1(x)
        x = F.relu(self.bn1(x))
        x = self.pool1(x)

        x = self.conv2(x)
        x = F.relu(self.bn2(x))
        x = self.pool2(x)

        x = self.conv3(x)
        x = F.relu(self.bn3(x))
        x = self.pool3(x)

        x = self.conv4(x)
        x = F.relu(self.bn4(x))
        x = self.pool4(x)

        x = self.dropout(x)
        x = self.avgPool(x)
        x = x.view(x.shape[0], -1)
        x = self.fc1(x)

        return x

[Kazuto Nakashima]

If you make grad-cam of class 1 at pool4:

def get_fmaps(module, inputs, outputs):
    global fmaps
    fmaps = outputs.detach()


def get_grads(module, inputs, outputs):
    global grads
    grads = outputs[0].detach()


model = Cnn1d().eval()
model.pool4.register_forward_hook(get_fmaps)
model.pool4.register_backward_hook(get_grads)

signals = torch.randn(1, 1, 4096)  # batch, channel, length
target_class = 1  # or 0
logit = model(signals)
mask = torch.zeros_like(logit)
mask[:, target_class] = 1.0
logit.backward(gradient=mask)

weights = grads.mean(dim=2, keepdim=True)
gradcam = (fmaps * weights).sum(dim=1, keepdim=True).relu()

[Yannis Kalfas]

@kazuto1011 thanks for the response 👍

How would this work with 1 Cnn1d output unit (e.g. sigmoid where classes are mapped in the range [0,1]) instead of 2 outputs (softmax w/ 2 units)? Maybe I'm missing some theoretical detail on GradCam.

[Kazuto Nakashima]

You can get the class-wise gradients as follows.
For a positive class:

logit = model(signals) # before sigmoid
mask = torch.ones_like(logit)
logit.backward(gradient=mask)

For a negative class:

logit = model(signals)
mask = torch.ones_like(logit)
(-logit).backward(gradient=mask)