Part 1: Fit a Neural Field to a 2D Image
For the whole part 1 we have num_epochs=3000 and batch_size=10000. The architecture of the neural net follows the one given in the project spec.
Fox With Original Parameters
{"num_layers": 3, "hidden_dim": 256, "num_frequencies": 10, "learning_rate": 1e-2}
Decrease Fox Hidden Dimension to 64
{"num_layers": 3, "hidden_dim": 64, "num_frequencies": 10, "learning_rate": 1e-2}
We can see that the results perform significantly worse than the original parameters. This make sense
since we decrease the amount of information contained in the layers by decreasing the hidden dimensions.
Decrease Fox num_layers to 1
{"num_layers": 1, "hidden_dim": 256, "num_frequencies": 10, "learning_rate": 1e-2}
We can see that this performs slightly worse than the original parameter. This make sense
since we are having less layers for the net to learn. However, it isn't as bad
as decreasing the hidden dimensions. This make sense since decreasing num_layers just mean that the neural
net couldn't learn as deep of a relationship between pixels. However, all the informations are still
retained as the original.
Custuom Image (Campanile)
{"num_layers": 3, "hidden_dim": 256, "num_frequencies": 10, "learning_rate": 1e-2}
This Cusuom Image is larger than the fox image, which should theoritaically make it harder to reconstruct.
Using the original parameters we can see that this reconstruction is indeed worse than the fox image.
