MiZhangWhuer

Hi, @xiaoiker @TDeVries , I have successfully run your code and I found the output of best_dag like this: Best architecture: [3] [3 1] [3 0 0] [3 0 0...

> Yes, the first number in each row are the operation types, which are defined here: > > [enas_pytorch/models/shared_cnn.py](https://github.com/TDeVries/enas_pytorch/blob/a3e54b290787572a0fab920d9f3b3149e09dbc82/models/shared_cnn.py#L73) > > Line 73 in [a3e54b2](/TDeVries/enas_pytorch/commit/a3e54b290787572a0fab920d9f3b3149e09dbc82) > > self.branch_0 = ConvBranch(in_planes,...

> > If I understand correctly, you mean that for these two nodes (index start from 0) in the table : > > [3 0 0] #line 3 (node 2)...

> So I think there might be an error in your upsampling method. You are using scales from 2 to 8 but these scales are not accurate in my opinion....

> In loss.py, detection loss is: > return torch.mean(L_g * y_true_cls * training_mask) + classification_loss > But I think the loss mean should divide only nonzero(y_true_cls * training_mask), so may...

> @MiZhangWhuer mergelayer branch is experimental branch that trying to parallel roirotate layer, it is not stable and may be fail somehow. So don't rely on it. Try to use...

> Training speed very slow, I have only one gpu titanx, have trained for three days. > Train Epoch: 186 [128/960 (13%)] Loss: 1.369987 Detection Loss: 0.646443 Recognition Loss:0.723544 >...

> Hi, looks like problem with data feeding. > > * you can try: use -debug=1 flag to see the training data > > there is piece of bad code...

> Hi, > The function moves all input capsules (poses + agreements) of one receptive field into an additional depth dimension of the array so that afterwards receptive field aggregation...