lromul / argus Goto Github PK

Can I use custom metrics?
Also if not, do you have any suggestions on metrics for Multiclass Segmentation problems?

I am training a model on 4 different GPUs. During the training, I use the MonitorCheckpoint callback for saving the model on disk. Due to long training time, my script stops after training. The test is performed by hand, reinitializing a python kernel and loading the trained model.

The loading is performed with model = argus.load_model() and, by printing the model variable, I can see the 4 devices I used in training. However, I would like to move the model to a single GPU for the tensor operations I need to perform must be on the same device. Indeed I got:
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:2 and cuda:0!

Is there any way to re-assign the cuda device? Using model.set_device(0) does not work, I can see all the previous GPU running.

Thanks in advance, and congratulations for the package!

When I install argus, there are some module cannot be found. However, these modules can be found after installing pytorch-argus. Are there any difference between argus and pytorch-argus?

Hey, I am currently facing a problem using argus when trying to use multiple model (resnet then pspnet).
It's fine when I'm doing it conventionally but got error TypeError: forward() missing 1 required positional argument when using argus.

the model that I meant is something like this :

class PSPNet(nn.Module):
    def __init__(self,in_ch=1,nb_classes=6,bins=(1, 2, 3, 6)):
        super(PSPNet,self).__init__()
        fea_dim = 128
        self.prev_model= resnet(in_ch,nb_classes) #feature map extract
        self.ppm = PPM(fea_dim, int(fea_dim/len(bins)), bins)
        self.cls = nn.Sequential(
            nn.Conv2d(fea_dim*2, 64, kernel_size=3, padding=1, bias=False),
            nn.BatchNorm2d(64),
            nn.ReLU(inplace=True),
            nn.Conv2d(64, nb_classes, kernel_size=1)
        )
    def forward(self,x):
        x = self.prev_model(x)
        x = self.ppm(x)
        x = self.cls(x)
        return x

class resnet(nn.Module):
       blabla

Any recommendation?
thanks.

Currently the default num_epochs is 1 and there is no support for the infinite num_epochs (using -1 for example).

What about this feature for the "I do not know how many epochs I need, I just want to provide the EarlyStopping callback or stop the training with KeyboardInterrupt".

Checking that either num_epochs > 0 or EarlyStopping_instance in callbacks is a valid option.