Adding metrics about torch-ngp HOT 6 CLOSED

Hi,
If I understand correctly, you are trying to use the metrics for Trainer? It is an old piece of code I borrowed from my other projects, so it may be confusing to use.

An example of a metric class can be:

class ROCMeter:
    def __init__(self):
        self.truths = []
        self.preds = []

    def clear(self):
        self.truths = []
        self.preds = []

    def prepare_inputs(self, *inputs):
        outputs = []
        for i, inp in enumerate(inputs):
            if torch.is_tensor(inp):
                inp = inp.detach().cpu().numpy()
            outputs.append(inp)

        return outputs

    def update(self, preds, truths):
        # [B, ], [B, ]
        preds, truths = self.prepare_inputs(preds, truths)
        self.truths.append(truths)
        self.preds.append(preds)


    def measure(self):
        return roc_auc_score(np.concatenate(self.truths, 0), np.concatenate(self.preds, 0))

    def write(self, writer, global_step, prefix=""):
        writer.add_scalar(os.path.join(prefix, "roc"), self.measure(), global_step)

    def report(self):
        return f'ROC = {self.measure():.6f}'

and used like:

trainer = Trainer(..., metrics=[ROCMeter(),], ...)

I will add a PSNR metric recently.

from torch-ngp.

Thank you sooo much! I'll try to implement that. Moreover, can you explain in SDF why are you using the negative sign here? sdfs[self.num_samples // 2:] = -self.sdf_fn(points[self.num_samples // 2:])[:,None].astype(np.float32)

from torch-ngp.

@JoyHighmore I have added the PSNR metric, you can try now!
As for the sdf value, it is just a convention (inner is negative, outer is positive), without the negation it should also work since we extract surface at sdf == 0.

(Besides, I find that the PSNR for lego is still worse compared to the original implementation, and I'll try to figure it out. If you find any problem, welcome to report!)

from torch-ngp.

Sure, there is one more thing that I'd like to discuss. I've noticed that there are some code trying to calculate the sdf by igl.signed_distance. I've also implement both igl and pysdf. It is found that they had quite different results, which not only with the different values but also different indications for inside and outside categories. Do you have any thoughts? What's your consideration for the pysdf? Also, are you trying to have balanced positive and negative sdf since you are using self.num_samples // 2 in this equation?

Best,
Joy

from torch-ngp.

@JoyHighmore That's strange, the SDF value should be unique as long as the mesh is fixed. Inside/Outside doesn't really matter as the only difference is a negation.
In fact the pos/neg points are not balanced, I followed the original implementation to sample 1/8 random points, 3/8 surface points, and 4/8 perturbed surface points.

from torch-ngp.

closed for now.

from torch-ngp.