summaryrefslogtreecommitdiffstats
path: root/training/util.py
blob: 60d64f086570d85e551347ea749fc51847104105 (plain) (blame)
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
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
from pathlib import Path
import json
import copy
import itertools
from typing import Iterable, Optional
from contextlib import contextmanager

import torch
from PIL import Image


def save_args(basepath: Path, args, extra={}):
    info = {"args": vars(args)}
    info["args"].update(extra)
    with open(basepath.joinpath("args.json"), "w") as f:
        json.dump(info, f, indent=4)


def make_grid(images, rows, cols):
    w, h = images[0].size
    grid = Image.new('RGB', size=(cols*w, rows*h))
    for i, image in enumerate(images):
        grid.paste(image, box=(i % cols*w, i//cols*h))
    return grid


class AverageMeter:
    def __init__(self, name=None):
        self.name = name
        self.reset()

    def reset(self):
        self.sum = self.count = self.avg = 0

    def update(self, val, n=1):
        self.sum += val * n
        self.count += n
        self.avg = self.sum / self.count


class CheckpointerBase:
    def __init__(
        self,
        datamodule,
        output_dir: Path,
        sample_image_size: int,
        sample_batches: int,
        sample_batch_size: int,
        seed: Optional[int] = None
    ):
        self.datamodule = datamodule
        self.output_dir = output_dir
        self.sample_image_size = sample_image_size
        self.seed = seed if seed is not None else torch.random.seed()
        self.sample_batches = sample_batches
        self.sample_batch_size = sample_batch_size

    @torch.inference_mode()
    def save_samples(self, pipeline, step, num_inference_steps, guidance_scale=7.5, eta=0.0):
        samples_path = Path(self.output_dir).joinpath("samples")

        train_data = self.datamodule.train_dataloader
        val_data = self.datamodule.val_dataloader

        generator = torch.Generator(device=pipeline.device).manual_seed(self.seed)

        grid_cols = min(self.sample_batch_size, 4)
        grid_rows = (self.sample_batches * self.sample_batch_size) // grid_cols

        for pool, data, gen in [("stable", val_data, generator), ("val", val_data, None), ("train", train_data, None)]:
            all_samples = []
            file_path = samples_path.joinpath(pool, f"step_{step}.jpg")
            file_path.parent.mkdir(parents=True, exist_ok=True)

            batches = list(itertools.islice(itertools.cycle(data), self.sample_batch_size * self.sample_batches))
            prompt_ids = [
                prompt
                for batch in batches
                for prompt in batch["prompt_ids"]
            ]
            nprompt_ids = [
                prompt
                for batch in batches
                for prompt in batch["nprompt_ids"]
            ]

            for i in range(self.sample_batches):
                start = i * self.sample_batch_size
                end = (i + 1) * self.sample_batch_size
                prompt = prompt_ids[start:end]
                nprompt = nprompt_ids[start:end]

                samples = pipeline(
                    prompt=prompt,
                    negative_prompt=nprompt,
                    height=self.sample_image_size,
                    width=self.sample_image_size,
                    generator=gen,
                    guidance_scale=guidance_scale,
                    eta=eta,
                    num_inference_steps=num_inference_steps,
                    output_type='pil'
                ).images

                all_samples += samples

                del samples

            image_grid = make_grid(all_samples, grid_rows, grid_cols)
            image_grid.save(file_path, quality=85)

            del all_samples
            del image_grid

        del generator


# Adapted from torch-ema https://github.com/fadel/pytorch_ema/blob/master/torch_ema/ema.py#L14
class EMAModel:
    """
    Exponential Moving Average of models weights
    """

    def __init__(
        self,
        parameters: Iterable[torch.nn.Parameter],
        update_after_step=0,
        inv_gamma=1.0,
        power=2 / 3,
        min_value=0.0,
        max_value=0.9999,
    ):
        """
        @crowsonkb's notes on EMA Warmup:
            If gamma=1 and power=1, implements a simple average. gamma=1, power=2/3 are good values for models you plan
            to train for a million or more steps (reaches decay factor 0.999 at 31.6K steps, 0.9999 at 1M steps),
            gamma=1, power=3/4 for models you plan to train for less (reaches decay factor 0.999 at 10K steps, 0.9999
            at 215.4k steps).
        Args:
            inv_gamma (float): Inverse multiplicative factor of EMA warmup. Default: 1.
            power (float): Exponential factor of EMA warmup. Default: 2/3.
            min_value (float): The minimum EMA decay rate. Default: 0.
        """
        parameters = list(parameters)
        self.shadow_params = [p.clone().detach() for p in parameters]

        self.collected_params = None

        self.update_after_step = update_after_step
        self.inv_gamma = inv_gamma
        self.power = power
        self.min_value = min_value
        self.max_value = max_value

        self.decay = 0.0
        self.optimization_step = 0

    def get_decay(self, optimization_step):
        """
        Compute the decay factor for the exponential moving average.
        """
        step = max(0, optimization_step - self.update_after_step - 1)
        value = 1 - (1 + step / self.inv_gamma) ** -self.power

        if step <= 0:
            return 0.0

        return max(self.min_value, min(value, self.max_value))

    @torch.no_grad()
    def step(self, parameters):
        parameters = list(parameters)

        self.optimization_step += 1

        # Compute the decay factor for the exponential moving average.
        self.decay = self.get_decay(self.optimization_step)

        for s_param, param in zip(self.shadow_params, parameters):
            if param.requires_grad:
                s_param.mul_(self.decay)
                s_param.add_(param.data, alpha=1 - self.decay)
            else:
                s_param.copy_(param)

        torch.cuda.empty_cache()

    def copy_to(self, parameters: Iterable[torch.nn.Parameter]) -> None:
        """
        Copy current averaged parameters into given collection of parameters.
        Args:
            parameters: Iterable of `torch.nn.Parameter`; the parameters to be
                updated with the stored moving averages. If `None`, the
                parameters with which this `ExponentialMovingAverage` was
                initialized will be used.
        """
        parameters = list(parameters)
        for s_param, param in zip(self.shadow_params, parameters):
            param.data.copy_(s_param.data)

    def to(self, device=None, dtype=None) -> None:
        r"""Move internal buffers of the ExponentialMovingAverage to `device`.
        Args:
            device: like `device` argument to `torch.Tensor.to`
        """
        # .to() on the tensors handles None correctly
        self.shadow_params = [
            p.to(device=device, dtype=dtype) if p.is_floating_point() else p.to(device=device)
            for p in self.shadow_params
        ]

    def state_dict(self) -> dict:
        r"""
        Returns the state of the ExponentialMovingAverage as a dict.
        This method is used by accelerate during checkpointing to save the ema state dict.
        """
        # Following PyTorch conventions, references to tensors are returned:
        # "returns a reference to the state and not its copy!" -
        # https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict
        return {
            "decay": self.decay,
            "optimization_step": self.optimization_step,
            "shadow_params": self.shadow_params,
            "collected_params": self.collected_params,
        }

    def load_state_dict(self, state_dict: dict) -> None:
        r"""
        Loads the ExponentialMovingAverage state.
        This method is used by accelerate during checkpointing to save the ema state dict.
        Args:
            state_dict (dict): EMA state. Should be an object returned
                from a call to :meth:`state_dict`.
        """
        # deepcopy, to be consistent with module API
        state_dict = copy.deepcopy(state_dict)

        self.decay = state_dict["decay"]
        if self.decay < 0.0 or self.decay > 1.0:
            raise ValueError("Decay must be between 0 and 1")

        self.optimization_step = state_dict["optimization_step"]
        if not isinstance(self.optimization_step, int):
            raise ValueError("Invalid optimization_step")

        self.shadow_params = state_dict["shadow_params"]
        if not isinstance(self.shadow_params, list):
            raise ValueError("shadow_params must be a list")
        if not all(isinstance(p, torch.Tensor) for p in self.shadow_params):
            raise ValueError("shadow_params must all be Tensors")

        self.collected_params = state_dict["collected_params"]
        if self.collected_params is not None:
            if not isinstance(self.collected_params, list):
                raise ValueError("collected_params must be a list")
            if not all(isinstance(p, torch.Tensor) for p in self.collected_params):
                raise ValueError("collected_params must all be Tensors")
            if len(self.collected_params) != len(self.shadow_params):
                raise ValueError("collected_params and shadow_params must have the same length")

    @contextmanager
    def apply_temporary(self, parameters: Iterable[torch.nn.Parameter]):
        try:
            parameters = list(parameters)
            original_params = [p.clone() for p in parameters]
            self.copy_to(parameters)
            yield
        finally:
            for s_param, param in zip(original_params, parameters):
                param.data.copy_(s_param.data)