flaxsr 0.0.7

Super Resolution tools with Jax/Flax

FlaxSR

Super Resolution models with Jax/Flax

Currently, Flax is using CUDA/CuDNN from wheel, but TensorFlow is using local CUDA/CuDNN, which is causing conflicts. We will fix it as soon as possible.<\b>

HOW TO USE

Install

pip install flaxsr

Usage

You can easily load model/losses and train model using custom train_states.

• Train example

import flaxsr
import jax
import jax.numpy as jnp
import numpy as np
import optax

model_kwargs = {
    'n_filters': 64, 'n_blocks': 8, 'scale': 4
}
model = flaxsr.get("models", "vdsr", **model_kwargs)  # This equals flaxsr.models.VDSR(**model_kwargs)
losses = [
    flaxsr.losses.L1Loss(reduce='sum'),
    flaxsr.get('losses', 'vgg', feats_from=(6, 8, 14,), before_act=False, reduce='mean')
]
loss_weights = (.1, 1.)
loss_wrapper = flaxsr.losses.LossWrapper(losses, loss_weights)
params = model.init(jax.random.PRNGKey(0), jnp.ones((1, 8, 8, 3), dtype=jnp.float32))
tx = optax.adam(1e-3)

state = flaxsr.training.TrainState.create(
    apply_fn=model.apply, params=params, tx=tx, losses=loss_wrapper
)

hr = jnp.ones((1, 32, 32, 3), dtype=jnp.float32)
lr = jnp.ones((1, 8, 8, 3), dtype=jnp.float32)
batch = (lr, hr)

state_new, loss = flaxsr.training.discriminative_train_step(state, batch)

assert state_new.step == 1
np.not_equal(state_new.params['params']['Conv_0']['kernel'], state.params['params']['Conv_0']['kernel'])

flaxsr.get keywords

• models

  • SRCNN: srcnn
  • FSRCNN: fsrcnn
  • ESPCN: espcn
  • VDSR: vdsr
  • EDSR: edsr
  • MDSR: mdsr
  • SRResNet: srresnet
  • SRGAN: srgan
  • NCNet: ncnet

• losses

  • L1Loss: l1
  • L2Loss: l2
  • CharbonnierLoss: charbonnier
  • VGGLoss: vgg
  • MinmaxDriscriminatorLoss: minmax_discriminator
  • MinmaxGeneratorLoss: minmax_generator
  • LeastSquareDiscriminatorLoss: least_square_discriminator
  • LeastSquareGeneratorLoss: least_square_generator
  • RelativisticDiscriminatorLoss: relativistic_discriminator
  • RelativisticGeneratorLoss: relativistic_generator
  • TotalVariationLoss: tv
  • FrequencyReconstructionLoss: freq_recon
  • EdgeLoss: edge

• layers

  • DropPath: droppath
  • DropPathFast: droppath_fast
  • PixelShuffle: pixelshuffle
  • NearestConv: nearestconv

• train_step

  • discriminative_train_step: discriminative