f32c32/models/discriminator.py

"""Discriminator from StyleGAN. https://github.com/google-research/maskgit/blob/main/maskgit/nets/discriminator.py"""

import functools
import math
from typing import Any, Tuple
import flax.linen as nn
from flax.linen.initializers import xavier_uniform
import jax
from jax import lax
import jax.numpy as jnp
import ml_collections

default_kernel_init = xavier_uniform()

def _conv_dimension_numbers(input_shape):
    """Computes the dimension numbers based on the input shape."""
    ndim = len(input_shape)
    lhs_spec = (0, ndim - 1) + tuple(range(1, ndim - 1))
    rhs_spec = (ndim - 1, ndim - 2) + tuple(range(0, ndim - 2))
    out_spec = lhs_spec
    return lax.ConvDimensionNumbers(lhs_spec, rhs_spec, out_spec)


class BlurPool2D(nn.Module):
    """A layer to do channel-wise blurring + subsampling on 2D inputs.

    Reference:
    Zhang et al. Making Convolutional Networks Shift-Invariant Again.
    https://arxiv.org/pdf/1904.11486.pdf.
    """
    filter_size: int = 4
    strides: Tuple[int, int] = (2, 2)
    padding: str = 'SAME'

    def setup(self):
        if self.filter_size == 3:
            self.filter = [1., 2., 1.]
        elif self.filter_size == 4:
            self.filter = [1., 3., 3., 1.]
        elif self.filter_size == 5:
            self.filter = [1., 4., 6., 4., 1.]
        elif self.filter_size == 6:
            self.filter = [1., 5., 10., 10., 5., 1.]
        elif self.filter_size == 7:
            self.filter = [1., 6., 15., 20., 15., 6., 1.]
        else:
            raise ValueError('Only filter_size of 3, 4, 5, 6 or 7 is supported.')

        self.filter = jnp.array(self.filter, dtype=jnp.float32)
        self.filter = self.filter[:, None] * self.filter[None, :]
        with jax.default_matmul_precision('float32'):
            self.filter /= jnp.sum(self.filter)
        self.filter = jnp.reshape(
            self.filter, [self.filter.shape[0], self.filter.shape[1], 1, 1])

    @nn.compact
    def __call__(self, inputs):
        channel_num = inputs.shape[-1]
        dimension_numbers = _conv_dimension_numbers(inputs.shape)
        depthwise_filter = jnp.tile(self.filter, [1, 1, 1, channel_num])
        with jax.default_matmul_precision('float32'):
            outputs = lax.conv_general_dilated(inputs, depthwise_filter, self.strides,
                self.padding, feature_group_count=channel_num, dimension_numbers=dimension_numbers)
        return outputs

class ResBlock(nn.Module):
    """StyleGAN ResBlock for D.

    https://github.com/rosinality/stylegan2-pytorch/blob/master/model.py#L618
    """
    filters: int
    activation_fn: Any

    @nn.compact
    def __call__(self, x):
        input_dim = x.shape[-1]
        residual = x
        x = nn.Conv(input_dim, (3, 3), kernel_init=default_kernel_init)(x)
        x = self.activation_fn(x)
        x = BlurPool2D(filter_size=4)(x)
        residual = BlurPool2D(filter_size=4)(residual)
        residual = nn.Conv(self.filters, (1, 1), use_bias=False, kernel_init=default_kernel_init)(residual)
        x = nn.Conv(self.filters, (3, 3), kernel_init=default_kernel_init)(x)
        x = self.activation_fn(x)
        out = (residual + x) / math.sqrt(2)
        return out


class Discriminator(nn.Module):
    """StyleGAN Discriminator."""
    config: ml_collections.ConfigDict

    def setup(self):
        self.input_size = self.config.image_size
        self.activation_fn = functools.partial(jax.nn.leaky_relu, negative_slope=0.2)
        self.channel_multiplier = 1

    @nn.compact
    def __call__(self, x):
        filters = {
            4: 512,
            8: 512,
            16: 512,
            32: 512,
            64: 256 * self.channel_multiplier,
            128: 128 * self.channel_multiplier,
            256: 64 * self.channel_multiplier,
            512: 32 * self.channel_multiplier,
            1024: 16 * self.channel_multiplier,
        }
        x = nn.Conv(filters[self.input_size], (3, 3), kernel_init=default_kernel_init)(x)
        x = self.activation_fn(x)
        log_size = int(math.log2(self.input_size))
        for i in range(log_size, 2, -1):
            x = ResBlock(filters[2**(i - 1)], self.activation_fn)(x)
            print("Disc shape", x.shape)
        x = nn.Conv(filters[4], (3, 3), kernel_init=default_kernel_init)(x)
        x = self.activation_fn(x)
        x = x.reshape((x.shape[0], -1))
        x = nn.Dense(filters[4], kernel_init=default_kernel_init)(x)
        x = self.activation_fn(x)
        x = nn.Dense(1, kernel_init=default_kernel_init)(x)
        return x