File size: 5,259 Bytes
cfeea40
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) Microsoft Corporation.
# Licensed under the MIT License.

from typing import Callable, Generic, TypeVar

import torch

from mattergen.diffusion.corruption.multi_corruption import MultiCorruption, apply
from mattergen.diffusion.data.batched_data import BatchedData
from mattergen.diffusion.losses import Loss
from mattergen.diffusion.model_target import ModelTarget
from mattergen.diffusion.model_utils import convert_model_out_to_score
from mattergen.diffusion.score_models.base import ScoreModel
from mattergen.diffusion.timestep_samplers import TimestepSampler, UniformTimestepSampler

T = TypeVar("T", bound=BatchedData)
BatchTransform = Callable[[T], T]


def identity(x: T) -> T:
    return x


class DiffusionModule(torch.nn.Module, Generic[T]):
    """Denoising diffusion model for a multi-part state"""

    def __init__(
        self,
        model: ScoreModel[T],
        corruption: MultiCorruption[T],
        loss_fn: Loss,
        pre_corruption_fn: BatchTransform | None = None,
        timestep_sampler: TimestepSampler | None = None,
    ) -> None:
        super().__init__()
        self.model = model
        self.corruption = corruption
        self.loss_fn = loss_fn
        self.pre_corruption_fn = pre_corruption_fn or identity
        self.model_targets = {k: ModelTarget(v) for k, v in loss_fn.model_targets.items()}

        self.timestep_sampler = timestep_sampler or UniformTimestepSampler(
            min_t=1e-5,
            max_t=corruption.T,
        )

        # Check corruption for nn.Modules and register them here.
        self._register_corruption_modules()

    def _register_corruption_modules(self):
        """
        Register corruptions that are instances of `torch.nn.Module`s for proper device, parameter,
        etc handling.
        """
        assert isinstance(self.corruption, MultiCorruption)
        for idx, (key, _corruption) in enumerate(self.corruption._corruptions.items()):
            if isinstance(_corruption, torch.nn.Module):
                self.register_module(f"MultiCorruption:{idx}:{key}", _corruption)

    def calc_loss(
        self, batch: T, node_is_unmasked: torch.LongTensor | None = None
    ) -> tuple[torch.Tensor, dict[str, torch.Tensor]]:
        """
        Calculate loss and metrics given a batch of clean data which may include
        context/conditioning fields. Add noise, predict score using score model, then calculate
        loss.

        Args:
            batch: batch of training data
            node_is_unmasked: mask that has a value 1 for nodes that are included in the loss, and
                a value of 0 for nodes that should be ignored. If None, all nodes are included.

        Returns:
            loss: the loss for the batch
            metrics: a dictionary of metrics for the batch
        """
        batch = self.pre_corruption_fn(batch)

        noisy_batch, t = self._corrupt_batch(batch)

        score_model_output = self.model(noisy_batch, t)
        loss, metrics = self.loss_fn(
            multi_corruption=self.corruption,
            batch=batch,
            noisy_batch=noisy_batch,
            score_model_output=score_model_output,
            t=t,
            node_is_unmasked=node_is_unmasked,
        )
        assert loss.numel() == 1

        return loss, metrics

    def _corrupt_batch(
        self,
        batch: T,
    ) -> tuple[T, torch.Tensor]:
        """
        Corrupt a batch of data for use in a training step:
        - sample a different timestep for each sample in the batch
        - add noise according to the corruption process

        Args:
            batch: Batch of clean states

        Returns:
            noisy_batch: batch of noisy samples
            t: the timestep used for each sample in the batch

        """
        # Sample timesteps
        t = self.sample_timesteps(batch)

        # Add noise to data
        noisy_batch = self.corruption.sample_marginal(batch, t)

        return noisy_batch, t

    def score_fn(self, x: T, t: torch.Tensor) -> T:
        """Calculate the score of a batch of data at a given timestep

        Args:
            x: batch of data
            t: timestep

        Returns:
            score: score of the batch of data at the given timestep
        """
        model_out: T = self.model(x, t)
        fns = {k: convert_model_out_to_score for k in self.corruption.sdes.keys()}

        scores = apply(
            fns=fns,
            model_out=model_out,
            broadcast=dict(t=t, batch=x),
            sde=self.corruption.sdes,
            model_target=self.model_targets,
            batch_idx=self.corruption._get_batch_indices(x),
        )

        return model_out.replace(**scores)

    def sample_timesteps(self, batch: T) -> torch.Tensor:
        """Sample the timesteps, which will be used to determine how much noise
        to add to data.

        Args:
           batch: batch of data to be corrupted

        Returns: sampled timesteps
        """
        return self.timestep_sampler(
            batch_size=batch.get_batch_size(),
            device=self._get_device(batch),
        )

    def _get_device(self, batch: T) -> torch.device:
        return next(batch[k].device for k in self.corruption.sdes.keys())