File size: 5,836 Bytes
e12111a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
"""Recycling strategy for diffusion training."""
from typing import Callable

import chex
import jax
import jax.numpy as jnp
from jax import lax
from omegaconf import DictConfig

from imgx.datasets.constant import IMAGE, LABEL
from imgx.datasets.dataset_info import DatasetInfo
from imgx.diffusion.time_sampler import TimeSampler
from imgx.task.diffusion_segmentation.diffusion import DiffusionSegmentation
from imgx.task.diffusion_segmentation.diffusion_step import get_loss_logits_metrics
from imgx.task.diffusion_segmentation.train_state import TrainState


def get_recycling_loss_step(
    train_state: TrainState,
    dataset_info: DatasetInfo,
    loss_config: DictConfig,
    diffusion_model: DiffusionSegmentation,
    time_sampler: TimeSampler,
    prev_step: str,
    reverse_step: bool,
) -> Callable[
    [chex.ArrayTree, chex.ArrayTree, jax.Array],
    tuple[jnp.ndarray, tuple[jnp.ndarray, chex.ArrayTree, jnp.ndarray, jnp.ndarray]],
]:
    """Return loss_step for recycling diffusion.

    Args:
        train_state: train state.
        dataset_info: dataset info to transform label to mask.
        loss_config: have weights of diff losses.
        diffusion_model: segmentation diffusion model.
        time_sampler: time sampler for training.
        prev_step: max or next.
            max means the previous step is num_timesteps - 1
            next means the previous step is min(t+1, num_timesteps - 1)
        reverse_step: reverse the previous step or not.

    Returns:
        loss_step: loss step function.
    """

    def loss_step(
        params: chex.ArrayTree,
        batch: dict[str, jnp.ndarray],
        key: jax.Array,
    ) -> tuple[jnp.ndarray, tuple[jnp.ndarray, chex.ArrayTree, jnp.ndarray, jnp.ndarray]]:
        """Apply forward and calculate loss."""
        key_dropout_re, key_dropout, key_t, key_noise_re, key_noise = jax.random.split(
            key=key, num=5
        )
        image, label = batch[IMAGE], batch[LABEL]
        mask_true = dataset_info.label_to_mask(label, axis=-1, dtype=image.dtype)
        x_start = diffusion_model.mask_to_x(mask=mask_true)
        batch_size = image.shape[0]

        # t, t_index, probs_t
        # t_re, t_index_re
        # empirically, sample t from [0, num_timesteps - 1) is better than
        # sampling from [0, num_timesteps), for most max / next options.
        if prev_step == "max":
            if reverse_step:
                raise ValueError("reverse_step should be False when prev_step is max.")
            t, t_index, probs_t = time_sampler.sample(
                key=key_t,
                batch_size=batch_size,
                t_index_min=0,  # inclusive
                t_index_max=diffusion_model.num_timesteps - 1,  # exclusive
                loss_count_hist=train_state.loss_count_hist,
                loss_sq_hist=train_state.loss_sq_hist,
            )
            t_index_re = jnp.full(
                shape=(batch_size,),
                fill_value=diffusion_model.num_timesteps - 1,
                dtype=jnp.int32,
            )
            t_re = time_sampler.t_index_to_t(t_index=t_index_re)
        elif prev_step == "next":
            t, t_index, probs_t = time_sampler.sample(
                key=key_t,
                batch_size=batch_size,
                t_index_min=0,  # inclusive
                t_index_max=diffusion_model.num_timesteps - 1,  # exclusive
                loss_count_hist=train_state.loss_count_hist,
                loss_sq_hist=train_state.loss_sq_hist,
            )
            t_index_re = t_index + 1
            t_re = time_sampler.t_index_to_t(t_index=t_index_re)
        else:
            raise ValueError(f"prev_step {prev_step} not recognised.")

        # recycling to get predicted x_start
        noise_re = diffusion_model.sample_noise(
            key=key_noise_re, shape=x_start.shape, dtype=x_start.dtype
        )
        x_t_re = diffusion_model.q_sample(x_start=x_start, noise=noise_re, t_index=t_index_re)
        mask_t_re = diffusion_model.x_to_mask(x_t_re)
        model_out_re = train_state.apply_fn(
            {"params": params},
            True,  # is_train
            image,
            mask_t_re,
            t_re,
            rngs={"dropout": key_dropout_re},
        )
        x_start_pred_re = diffusion_model.predict_xstart_from_model_out_xt(
            model_out=model_out_re,
            x_t=x_t_re,
            t_index=t_index_re,
        )

        # x_t
        if reverse_step and (prev_step == "next"):
            noise = noise_re  # with interpolation, it's the same noise
            x_t = diffusion_model.predict_xprev_from_xstart_xt(
                x_start=x_start_pred_re,
                x_t=x_t_re,
                t_index=t_index_re,
            )
        else:
            noise = diffusion_model.sample_noise(
                key=key_noise, shape=x_start_pred_re.shape, dtype=x_start_pred_re.dtype
            )
            x_t = diffusion_model.q_sample(x_start=x_start_pred_re, noise=noise, t_index=t_index)
        mask_t = diffusion_model.x_to_mask(x_t)
        mask_t = lax.stop_gradient(mask_t)

        # forward
        model_out = train_state.apply_fn(
            {"params": params},
            True,  # is_train
            image,
            mask_t,
            t,
            rngs={"dropout": key_dropout},
        )

        # loss
        loss, loss_batch, logits, metrics = get_loss_logits_metrics(
            batch=batch,
            x_start=x_start,
            x_t=x_t,
            t_index=t_index,
            probs_t=probs_t,
            noise=noise,
            model_out=model_out,
            dataset_info=dataset_info,
            loss_config=loss_config,
            diffusion_model=diffusion_model,
        )

        return loss, (logits, metrics, loss_batch, t_index)

    return loss_step