Problems with binary classifier & metrics

Hi,

I am new to PyTorch and am trying to get started with Ignite. I want to train a simple binary classifier:

class NeuralNet(nn.Module):
  def __init__(self, input_size, hidden_size, num_classes):
    super(NeuralNet, self).__init__()
    self.fc1 = nn.Linear(input_size, hidden_size)
    self.relu = nn.ReLU()
    self.fc2 = nn.Linear(hidden_size, num_classes)
    self.out_act = nn.Sigmoid()

  def forward(self, x):
    out = self.fc1(x)
    out = self.relu(out)
    out = self.fc2(out)
    out = self.out_act(out)
    return out

First I had (plus some events):

trainer = create_supervised_trainer(model, optimizer, nn.BCELoss(), device=device)

evaluator = create_supervised_evaluator(
    model, metrics={'precision': Precision(),
                    'recall': Recall(),
                    'accuracy': BinaryAccuracy()}, device=device)

This failed with:

~/.local/share/virtualenvs/bot2-Z_RSwUyv/lib/python3.6/site-packages/ignite/metrics/precision.py in update(self, output)
     29         num_classes = y_pred.size(1)
     30         indices = torch.max(y_pred, 1)[1]
---> 31         correct = torch.eq(indices, y)
     32         pred_onehot = to_onehot(indices, num_classes)
     33         all_positives = pred_onehot.sum(dim=0)

RuntimeError: Expected object of type torch.LongTensor but found type torch.FloatTensor for argument #2 'other'

Then I found the output_transform parameter and changed it to:

def output_transform(output):
  y_pred, y = output
  return y_pred.gt(0.5).long(), y.long()

evaluator = create_supervised_evaluator(
    model, metrics={'precision': Precision(output_transform=output_transform),
                    'recall': Recall(output_transform=output_transform),
                    'accuracy': BinaryAccuracy(output_transform=output_transform)}, device=device)

Even if this worked, I do not think it was a good solution. In my opinion, it should be possible to do the transformation in just one place instead of having to pass it to each metric. Is this somehow possible? Anyway, this fails with:

~/.local/share/virtualenvs/bot2-Z_RSwUyv/lib/python3.6/site-packages/ignite/metrics/precision.py in update(self, output)
     35             true_positives = torch.zeros_like(all_positives)
     36         else:
---> 37             correct_onehot = to_onehot(indices[correct], num_classes)
     38             true_positives = correct_onehot.sum(dim=0)
     39         if self._all_positives is None:

IndexError: too many indices for tensor of dimension 1

Am I on the right track?

[vfdev-5]

@anmolsjoshi please go ahead. I think nobody has yet taken this one.

[anmolsjoshi]

Is someone working on this? Can I help out?