q_sample can only generate noise image.

[lvyufeng]

I have reproduced the ddpm with MindSpore, all Modules and initialized methods are the same, but the q_sample method can not generate an ordinary image but always returns a noise. I have check the flax implementation in denoising-diffusion-flax and found the code snippet:

    for t in reversed(jnp.arange(timesteps)):
        rng, *step_rng = jax.random.split(rng, num=jax.local_device_count() + 1)
        step_rng = jnp.asarray(step_rng)
        x, x0 = p_sample_step(state, step_rng, x, jax_utils.replicate(t), x0)
        list_x0.append(x0)
    # normalize to [0,1]
    img = unnormalize_to_zero_to_one(jnp.asarray(x0))

but found the pytorch code:

        for t in tqdm(reversed(range(0, self.num_timesteps)), desc = 'sampling loop time step', total = self.num_timesteps):
            self_cond = x_start if self.self_condition else None
            img, x_start = self.p_sample(img, t, self_cond)

        img = unnormalize_to_zero_to_one(img)

it seems the final image use x_start not img.

The other problem is that I have tried x_start or img, both of them can only return noise image.

[lhaippp]

did you produce the colored noise such as:

I think I have also failed to produce a normal image

[lvyufeng]

did you produce the colored noise such as: I think I have also failed to produce a normal image

exactly same.

[lhaippp]

issue62
as mentioned, I solved the problem, hope it help

[Ludrasaum]

It might be useful to set EMA to False

[AdamWojtczak]

@Ludrasaum
Hi, How to set EMA to False?