Inquiry About PirateNet Code Release Date

[liwenlin664477]

I am writing to inquire about the release date of the PirateNet code.

Thanks!!

[sifanexisted]

Sorry about this. We plan to release it ASAP and will definitely get back to you a more certrain answer next week.

In fact, implementing this arch should be quite easy. As a proof of concept, you can simply define a trainable parameter alpha (initalized by zero) for each block. I believe it should work and be better than plain MLP when the underlying PINN model is deep.

class Bottleneck(nn.Module):
    hidden_dim: int
    output_dim: int
    nonlinearity: float=0.0

    @nn.compact
    def __call__(self, x):
        identity = x

        x = nn.Dense(features=self.hidden_dim)(x)
        x = jnp.tanh(x)

        x = nn.Dense(features=self.hidden_dim)(x)
        x =  jnp.tanh(x)

        x = nn.Dense(features=self.hidden_dim)(x)
        x =  jnp.tanh(x)

        alpha = self.param("alpha", constant(self.nonlinearity), (1,))
        x = alpha * x + (1 - alpha) * identity

        return x

[liwenlin664477]

Sorry about this. We plan to release it ASAP and will definitely get back to you a more certrain answer next week.

In fact, implementing this arch should be quite easy. As a proof of concept, you can simply define a trainable parameter alpha (initalized by zero) for each block. I believe it should work and be better than plain MLP when the underlying PINN model is deep.

class Bottleneck(nn.Module):
    hidden_dim: int
    output_dim: int
    nonlinearity: float=0.0

    @nn.compact
    def __call__(self, x):
        identity = x

        x = nn.Dense(features=self.hidden_dim)(x)
        x = jnp.tanh(x)

        x = nn.Dense(features=self.hidden_dim)(x)
        x =  jnp.tanh(x)

        x = nn.Dense(features=self.hidden_dim)(x)
        x =  jnp.tanh(x)

        alpha = self.param("alpha", constant(self.nonlinearity), (1,))
        x = alpha * x + (1 - alpha) * identity

        return x

Thanks!!