| |
| |
| """PyTorch and JAX give the same numbers: FID within 0.01 and LPIPS within 1e-4.""" |
|
|
| from __future__ import annotations |
|
|
| from typing import TYPE_CHECKING |
|
|
| import jax |
| import jax.numpy as jnp |
| import numpy as np |
| import pytest |
| import torch |
|
|
| from conftest import structured_images |
| from sen_bridge import jax as sbj |
| from sen_bridge import torch as sbt |
|
|
| if TYPE_CHECKING: |
| from numpy.typing import NDArray |
|
|
| Pair = tuple[NDArray[np.float32], NDArray[np.float32]] |
|
|
| _FID_BATCH = 12 |
|
|
|
|
| def nhwc(a: NDArray[np.float32]) -> jax.Array: |
| """Convert channel-first images to a channel-last JAX array. |
| |
| Args: |
| a: Images of shape ``[B, 3, H, W]``. |
| |
| Returns: |
| The images as a ``[B, H, W, 3]`` array. |
| """ |
| return jnp.asarray(np.transpose(a, (0, 2, 3, 1))) |
|
|
|
|
| def test_psnr_ssim_parity(pair: Pair) -> None: |
| x, y = pair |
| xt, yt = torch.from_numpy(x), torch.from_numpy(y) |
| for name in ("psnr", "ssim"): |
| got = np.asarray(jax.jit(jax.vmap(getattr(sbj, name)))(nhwc(x), nhwc(y))) |
| assert np.abs(got - getattr(sbt, name)(xt, yt).numpy()).max() < 1e-4 |
|
|
|
|
| @pytest.mark.usefixtures("weights_ready") |
| def test_lpips_parity(pair: Pair) -> None: |
| x, y = pair |
| got = np.asarray(jax.jit(jax.vmap(sbj.lpips))(nhwc(x), nhwc(y))) |
| expected = sbt.lpips(torch.from_numpy(x), torch.from_numpy(y)).numpy() |
| assert np.abs(got - expected).max() < 1e-4 |
|
|
|
|
| @pytest.mark.usefixtures("weights_ready") |
| def test_fid_parity() -> None: |
| real, fake = structured_images(24, seed=3), structured_images(24, seed=4) |
| fid_t, fid_j = sbt.FID(device="cpu"), sbj.FID() |
| for lo in range(0, 24, _FID_BATCH): |
| sl = slice(lo, lo + _FID_BATCH) |
| fid_t.update(torch.from_numpy(real[sl]), real=True) |
| fid_t.update(torch.from_numpy(fake[sl]), real=False) |
| fid_j.update(nhwc(real[sl]), real=True) |
| fid_j.update(nhwc(fake[sl]), real=False) |
| assert abs(fid_t.compute() - fid_j.compute()) < 0.01 |
|
|
|
|
| @pytest.mark.usefixtures("weights_ready") |
| def test_lpips_first_call_inside_jit_does_not_leak_tracers(pair: Pair) -> None: |
| sbj._lpips_params.cache_clear() |
| x, y = nhwc(pair[0]), nhwc(pair[1]) |
| first = jax.jit(sbj.lpips)(x[0], y[0]) |
| batch = jax.jit(jax.vmap(sbj.lpips))( |
| x[:2], y[:2] |
| ) |
| assert np.isfinite(float(first)) |
| assert np.isfinite(np.asarray(batch)).all() |
|
|
|
|
| def test_rounding_agrees_between_backends() -> None: |
| rng = np.random.default_rng(0) |
| ties = (np.arange(256) + 0.5) / 127.5 - 1 |
| values = np.concatenate([rng.uniform(-1.2, 1.2, 100_000), ties]).astype( |
| np.float32 |
| ) |
| levels_t = np.round(sbt.quantise(torch.from_numpy(values)).numpy() * 255) |
| levels_j = np.round(np.asarray(sbj.quantise(jnp.asarray(values))) * 255) |
| assert np.array_equal( |
| levels_t, levels_j |
| ) |
|
|
