stable-diffusion-trainer 0.1.9

Stable Diffusion Trainer
Documentation 
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

//! The Trainer module contains the training configuration and the training process.

use std::{path::PathBuf, process::Command};

pub mod training;
pub mod output;
pub mod optimizer;
pub mod scheduler;
pub mod parameters;

pub use training::*;
pub use output::*;
pub use optimizer::*;
pub use scheduler::*;
pub use parameters::*;

use crate::{environment::Environment, model_file_format::ModelFileFormat, precision::FloatPrecision};

/// The Trainer structure.
pub struct Trainer {
    /// The environment to use for the training process.
    pub environment: Environment,
    /// The number of times to repeat the training images.
    pub training_images_repeat: usize,
    /// The number of times to repeat the regularization images.
    pub regularization_images_repeat: usize,
    /// The maximum resolution of the images to use for the training process.
    pub resolution: (usize, usize),
    /// The format to save the model as.
    pub save_model_as: ModelFileFormat,
    /// The module to use for the network.
    pub network_module: String,
    /// The learning rate for the text encoder.
    pub text_encoder_lr: f32,
    /// The learning rate for the unet.
    pub unet_lr: f32,
    /// The number of cycles for the learning rate scheduler.
    pub lr_scheduler_num_cycles: usize,
    /// The learning rate for the training process.
    pub learning_rate: f32,
    /// The number of warmup steps for the learning rate.
    pub lr_warmup_steps: usize,
    /// The batch size for the training process.
    pub train_batch_size: usize,
    /// The maximum number of training steps.
    pub max_train_steps: usize,
    /// The frequency to 
    pub save_every_n_epochs: usize,
    /// The precision to use for mixed precision training.
    pub mixed_precision: FloatPrecision,
    /// The precision to use for saving the model.
    pub save_precision: FloatPrecision,
    /// The maximum gradient norm.
    pub max_grad_norm: f32,
    /// The maximum number of data loader workers.
    pub max_data_loader_n_workers: usize,
    /// The number of steps for the bucket resolution.
    pub bucket_reso_steps: usize,
    /// The noise offset.
    pub noise_offset: f32,
}

impl Default for Trainer {
    fn default() -> Self {
        Trainer {
            environment: Default::default(),
            training_images_repeat: 40,
            regularization_images_repeat: 1,
            resolution: (1024,1024),
            save_model_as: ModelFileFormat::Safetensors,
            network_module: "networks.lora".to_string(),
            text_encoder_lr: 5e-05,
            unet_lr: 0.0001,
            lr_scheduler_num_cycles: 1,
            learning_rate: 0.0001,
            lr_warmup_steps: 48,
            train_batch_size: 1,
            max_train_steps: 480,
            save_every_n_epochs: 1,
            mixed_precision: FloatPrecision::F16,
            save_precision: FloatPrecision::F16,
            max_grad_norm: 1.0,
            max_data_loader_n_workers: 0,
            bucket_reso_steps: 64,
            noise_offset: 0.0,
        }
    }
}

impl Trainer {
    /// Create a new Trainer.
    pub fn new() -> Self {
        Default::default()
    }

    /// Set the environment for the training process.
    pub fn with_environment(mut self, environment: Environment) -> Self {
        self.environment = environment;
        self
    }

    fn training_dir() -> PathBuf {
        if let Some(path) = std::env::var_os("TRAINING_DIR") {
            PathBuf::from(path)
        } else {
            std::env::temp_dir().join(uuid::Uuid::new_v4().to_string())
        }
    }

    /// Start the training process.
    pub fn start(&mut self, parameters: &Parameters) {
        let training_dir = Self::training_dir();
        self.prepare(parameters, &training_dir);
        self.activate();
        self.caption(parameters, &training_dir);
        self.train(parameters, &training_dir);
        self.deactivate();
    }

    fn image_dir(training_dir: &PathBuf) -> PathBuf {
        training_dir.join("img")
    }

    fn reg_dir(training_dir: &PathBuf) -> PathBuf {
        training_dir.join("reg")
    }

    fn subject_dir(&self, parameters: &Parameters, training_dir: &PathBuf) -> PathBuf {
        Self::image_dir(training_dir).join(format!("{}_{} {}", self.training_images_repeat, parameters.prompt.instance, parameters.prompt.class))
    }

    fn activate(&mut self) {
        self.environment.activate();
    }

    fn deactivate(&mut self) {
        self.environment.deactivate();
    }

    fn prepare(&self, parameters: &Parameters, training_dir: &PathBuf) {
        let image_dir = self.subject_dir(parameters, training_dir);
        let class_dir = Self::reg_dir(training_dir).join(format!("{}_{}", self.regularization_images_repeat, parameters.prompt.class));
        std::fs::create_dir_all(training_dir.join("log")).unwrap();
        std::fs::create_dir_all(training_dir.join("model")).unwrap();
        std::fs::create_dir_all(&image_dir).unwrap();
        std::fs::create_dir_all(&class_dir).unwrap();
        println!("{}", parameters.dataset.training.path().display());
        for file in parameters.dataset.training.path().read_dir().unwrap() {
            let file = file.unwrap().path();
            let file_name = file.file_name().unwrap();
            std::fs::copy(&file, image_dir.join(file_name)).unwrap();
        }

        if let Some(regularization) = &parameters.dataset.regularization {
            for file in regularization.path().read_dir().unwrap() {
                let file = file.unwrap().path();
                let file_name = file.file_name().unwrap();
                std::fs::copy(&file, class_dir.join(file_name)).unwrap();
            }
        }
    }

    fn caption(&self, parameters: &Parameters, training_dir: &PathBuf) {
        let image_dir = self.subject_dir(parameters, training_dir);
        let python_executable = self.environment.python_executable_path();
        Command::new(python_executable)
        .arg(self.environment.kohya_ss().join("finetune").join("make_captions.py"))
        .args(["--batch_size", "1"])
        .args(["--num_beams", "1"])
        .args(["--top_p", "0.9"])
        .args(["--max_length", "75"])
        .args(["--min_length", "5"])
        .arg("--beam_search")
        .args(["--caption_extension", ".txt"])
        .arg(&image_dir)
        .args(["--caption_weights", "https://storage.googleapis.com/sfr-vision-language-research/BLIP/models/model_large_caption.pth"])
        .status()
        .expect("Failed to execute command");
        for txt in image_dir.read_dir().unwrap() {
            let txt = txt.unwrap().path();
            if txt.extension().unwrap() == "txt" {
                let content = format!("{} {} {}", parameters.prompt.instance, parameters.prompt.class, std::fs::read_to_string(&txt).unwrap());
                std::fs::write(txt, content).expect("Failed to update txt file");
            }
        }
    }

    fn train(&self, parameters: &Parameters, training_dir: &PathBuf) {
        Command::new("accelerate")
        .arg("launch")
        .arg("--num_cpu_threads_per_process=2")
        .arg(self.environment.kohya_ss().join("sdxl_train_network.py"))
        .args(["--train_data_dir", &Self::image_dir(training_dir).display().to_string()])
        .args(["--reg_data_dir", &Self::reg_dir(training_dir).display().to_string()])
        .args(["--output_dir", &parameters.output.directory.display().to_string()])
        .args(["--output_name", &parameters.output.name])
        .args(["--pretrained_model_name_or_path", &parameters.training.pretrained_model])
        .args(["--resolution", &format!("{},{}", self.resolution.0, self.resolution.1)])
        .args(["--save_model_as", &self.save_model_as.to_string()])
        .args(["--network_alpha", &parameters.network.alpha.to_string()])
        .args(["--network_module", &self.network_module])
        .args(["--network_dim", &parameters.network.dimension.to_string()])
        .args(["--text_encoder_lr", &self.text_encoder_lr.to_string()])
        .args(["--unet_lr", &self.unet_lr.to_string()])
        .args(["--lr_scheduler_num_cycles", &self.lr_scheduler_num_cycles.to_string()])
        .arg("--no_half_vae")
        .args(["--learning_rate", &self.learning_rate.to_string()])
        .args(["--lr_scheduler", &parameters.training.learning_rate.scheduler.to_string()])
        // .args(["--lr_warmup_steps", &self.lr_warmup_steps.to_string()])
        .args(["--train_batch_size", &self.train_batch_size.to_string()])
        // .args(["--max_train_steps", &self.max_train_steps.to_string()])
        .args(["--save_every_n_epochs", &self.save_every_n_epochs.to_string()])
        .args(["--mixed_precision", &self.mixed_precision.to_string()])
        .args(["--save_precision", &self.save_precision.to_string()])
        .args(["--optimizer_type", &parameters.training.optimizer.to_string()])
        .args(["--max_grad_norm", &self.max_grad_norm.to_string()])
        .args(["--max_data_loader_n_workers", &self.max_data_loader_n_workers.to_string()])
        
        // Move it to Adafactor
        // .args(["--optimizer_args", "scale_parameter=False", "relative_step=False", "warmup_init=False"])
        
        .arg("--xformers")
        .arg("--enable_bucket")
        .args(["--min_bucket_reso", "256"])
        .args(["--max_bucket_reso", "2048"])
        .args(["--bucket_reso_steps", &self.bucket_reso_steps.to_string()])
        .arg("--bucket_no_upscale")
        .args(["--noise_offset", &self.noise_offset.to_string()])
        .status()
        .expect("Failed to execute command");
    }
}