OneTrainer/modules/dataLoader/ErnieBaseDataLoader.py at e221eae9650d7a596dc20b07f8d96a863654517c · Nerogar/OneTrainer · GitHub

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
import os

from modules.dataLoader.BaseDataLoader import BaseDataLoader
from modules.dataLoader.mixin.DataLoaderText2ImageMixin import DataLoaderText2ImageMixin
from modules.model.ErnieModel import HIDDEN_STATES_LAYER, PROMPT_MAX_LENGTH, ErnieModel
from modules.modelSetup.BaseErnieSetup import BaseErnieSetup
from modules.util import factory
from modules.util.config.TrainConfig import TrainConfig
from modules.util.enum.ModelType import ModelType
from modules.util.TrainProgress import TrainProgress

from mgds.pipelineModules.DecodeTokens import DecodeTokens
from mgds.pipelineModules.DecodeVAE import DecodeVAE
from mgds.pipelineModules.EncodeMistralText import EncodeMistralText
from mgds.pipelineModules.EncodeVAE import EncodeVAE
from mgds.pipelineModules.PadMaskedTokens import PadMaskedTokens
from mgds.pipelineModules.PruneMaskedTokens import PruneMaskedTokens
from mgds.pipelineModules.RescaleImageChannels import RescaleImageChannels
from mgds.pipelineModules.SampleVAEDistribution import SampleVAEDistribution
from mgds.pipelineModules.SaveImage import SaveImage
from mgds.pipelineModules.SaveText import SaveText
from mgds.pipelineModules.ScaleImage import ScaleImage
from mgds.pipelineModules.Tokenize import Tokenize


@factory.register(BaseDataLoader, ModelType.ERNIE)
class ErnieBaseDataLoader(
    BaseDataLoader,
    DataLoaderText2ImageMixin,
):
    def _preparation_modules(self, config: TrainConfig, model: ErnieModel):
        rescale_image = RescaleImageChannels(image_in_name='image', image_out_name='image', in_range_min=0, in_range_max=1, out_range_min=-1, out_range_max=1)
        encode_image = EncodeVAE(in_name='image', out_name='latent_image_distribution', vae=model.vae, autocast_contexts=[model.autocast_context], dtype=model.train_dtype.torch_dtype())
        image_sample = SampleVAEDistribution(in_name='latent_image_distribution', out_name='latent_image', mode='mean')
        downscale_mask = ScaleImage(in_name='mask', out_name='latent_mask', factor=0.125)
        tokenize_prompt = Tokenize(in_name='prompt', tokens_out_name='tokens', mask_out_name='tokens_mask', tokenizer=model.tokenizer, max_token_length=PROMPT_MAX_LENGTH)
        encode_prompt = EncodeMistralText(tokens_name='tokens', tokens_attention_mask_in_name='tokens_mask', hidden_state_out_name='text_encoder_hidden_state', tokens_attention_mask_out_name='tokens_mask',
                                          text_encoder=model.text_encoder, hidden_state_output_index=HIDDEN_STATES_LAYER, autocast_contexts=[model.autocast_context], dtype=model.train_dtype.torch_dtype())
        prune_masked_tokens = PruneMaskedTokens(tokens_name='tokens', tokens_mask_name='tokens_mask', hidden_state_name='text_encoder_hidden_state')

        modules = [rescale_image, encode_image, image_sample]
        if config.masked_training or config.model_type.has_mask_input():
            modules.append(downscale_mask)
        modules += [tokenize_prompt, encode_prompt]

        if config.text_caching:
            modules.append(prune_masked_tokens)

        return modules

    def _cache_modules(self, config: TrainConfig, model: ErnieModel, model_setup: BaseErnieSetup):
        image_split_names = ['latent_image', 'original_resolution', 'crop_offset']

        if config.masked_training or config.model_type.has_mask_input():
            image_split_names.append('latent_mask')

        image_aggregate_names = ['crop_resolution', 'image_path']

        text_split_names = []

        sort_names = image_aggregate_names + image_split_names + [
            'prompt', 'tokens', 'tokens_mask', 'text_encoder_hidden_state',
            'concept'
        ]

        text_split_names += ['tokens', 'tokens_mask', 'text_encoder_hidden_state']

        return self._cache_modules_from_names(
            model, model_setup,
            image_split_names=image_split_names,
            image_aggregate_names=image_aggregate_names,
            text_split_names=text_split_names,
            sort_names=sort_names,
            config=config,
            text_caching=config.text_caching,
        )

    def _output_modules(self, config: TrainConfig, model: ErnieModel, model_setup: BaseErnieSetup):
        pad_masked_tokens = PadMaskedTokens(tokens_name='tokens', tokens_mask_name='tokens_mask', hidden_state_name='text_encoder_hidden_state', max_length=PROMPT_MAX_LENGTH)

        output_names = [
            'image_path', 'latent_image',
            'prompt',
            'tokens',
            'tokens_mask',
            'original_resolution', 'crop_resolution', 'crop_offset',
        ]

        if config.masked_training or config.model_type.has_mask_input():
            output_names.append('latent_mask')

        output_names.append('text_encoder_hidden_state')

        output_module_list = self._output_modules_from_out_names(
            model, model_setup,
            output_names=output_names,
            config=config,
            use_conditioning_image=False,
            vae=model.vae,
            autocast_context=[model.autocast_context],
            train_dtype=model.train_dtype,
        )

        if config.text_caching:
            output_module_list = [pad_masked_tokens] + output_module_list

        return output_module_list

    def _debug_modules(self, config: TrainConfig, model: ErnieModel):
        debug_dir = os.path.join(config.debug_dir, "dataloader")

        def before_save_fun():
            model.vae_to(self.train_device)

        decode_image = DecodeVAE(in_name='latent_image', out_name='decoded_image', vae=model.vae, autocast_contexts=[model.autocast_context], dtype=model.train_dtype.torch_dtype())
        upscale_mask = ScaleImage(in_name='latent_mask', out_name='decoded_mask', factor=8)
        decode_prompt = DecodeTokens(in_name='tokens', out_name='decoded_prompt', tokenizer=model.tokenizer)
        save_image = SaveImage(image_in_name='decoded_image', original_path_in_name='image_path', path=debug_dir, in_range_min=-1, in_range_max=1, before_save_fun=before_save_fun)
        save_mask = SaveImage(image_in_name='decoded_mask', original_path_in_name='image_path', path=debug_dir, in_range_min=0, in_range_max=1, before_save_fun=before_save_fun)
        save_prompt = SaveText(text_in_name='decoded_prompt', original_path_in_name='image_path', path=debug_dir, before_save_fun=before_save_fun)

        modules = []

        modules.append(decode_image)
        modules.append(save_image)

        if config.masked_training or config.model_type.has_mask_input():
            modules.append(upscale_mask)
            modules.append(save_mask)

        modules.append(decode_prompt)
        modules.append(save_prompt)

        return modules

    def _create_dataset(
            self,
            config: TrainConfig,
            model: ErnieModel,
            model_setup: BaseErnieSetup,
            train_progress: TrainProgress,
            is_validation: bool = False,
    ):
        return DataLoaderText2ImageMixin._create_dataset(self,
            config, model, model_setup, train_progress, is_validation,
            aspect_bucketing_quantization=64,
        )