automatic/cli/modules/train-lora.py

#!/bin/env python

"""
Extract approximating LoRA by SVD from two SD models
Based on: <https://github.com/kohya-ss/sd-scripts/blob/main/networks/train_network.py>

Train LoRA with custom preprocessing, tagging and bucketing

Disabled/broken:
- `accelerate` with *dynamo* enabled
- `xformers` due to *faketensors* requirement
- `mem_eff_attn` due to *forwardfunc* mismatch
- 'use_8bit_adam` due to *bitsandbyttes* CUDA errors
"""

import os
import gc
import sys
import json
import argparse
import tempfile
import torch
import transformers
from pathlib import Path
from util import log, Map, get_memory
import process
import multiinterrogate
import lora_latents

sys.path.append(os.path.join(os.path.dirname(__file__), '..', '..', 'modules', 'lora'))
from train_network import train


options = Map({
    "v2": False,
    "v_parameterization": False,
    "pretrained_model_name_or_path": "",
    "train_data_dir": "",
    "shuffle_caption": False,
    "caption_extension": ".txt",
    "caption_extention": None,
    "keep_tokens": None,
    "color_aug": False,
    "flip_aug": False,
    "face_crop_aug_range": None,
    "random_crop": False,
    "debug_dataset": False,
    "resolution": "512,512",
    "cache_latents": True,
    "enable_bucket": False,
    "min_bucket_reso": 256,
    "max_bucket_reso": 1024,
    "bucket_reso_steps": 64,
    "bucket_no_upscale": False,
    "reg_data_dir": None,
    "in_json": "",
    "dataset_repeats": 1,
    "output_dir": "",
    "output_name": "",
    "save_precision": "fp16",
    "save_every_n_epochs": None,
    "save_n_epoch_ratio": None,
    "save_last_n_epochs": None,
    "save_last_n_epochs_state": None,
    "save_state": False,
    "resume": None,
    "train_batch_size": 1,
    "max_token_length": None,
    "use_8bit_adam": False,
    "mem_eff_attn": False,
    "xformers": False,
    "vae": None,
    "learning_rate": 1e-04,
    "max_train_steps": 8000,
    "max_train_epochs": None,
    "max_data_loader_n_workers": 8,
    "persistent_data_loader_workers": False,
    "seed": 42,
    "gradient_checkpointing": False,
    "gradient_accumulation_steps": 1,
    "mixed_precision": "fp16",
    "full_fp16": False,
    "clip_skip": None,
    "logging_dir": None,
    "log_prefix": None,
    "lr_scheduler": "cosine",
    "lr_warmup_steps": 0,
    "prior_loss_weight": 1.0,
    "no_metadata": False,
    "save_model_as": "ckpt",
    "unet_lr": 0.001,
    "text_encoder_lr": 5e-05,
    "lr_scheduler_num_cycles": 1,
    "lr_scheduler_power": 1,
    "network_weights": None,
    "network_module": "networks.lora",
    "network_dim": 16,
    "network_alpha": 1.0,
    "network_args": None,
    "network_train_unet_only": False,
    "network_train_text_encoder_only": False,
    "training_comment": "mood-magic",
    "caption_dropout_rate": 0.0,
    "caption_dropout_every_n_epochs": None,
    "caption_tag_dropout_rate": 0.0,
})


def mem_stats():
    gc.collect()
    if torch.cuda.is_available():
        with torch.no_grad():
            torch.cuda.empty_cache()
        with torch.cuda.device('cuda'):
            torch.cuda.empty_cache()
            torch.cuda.ipc_collect()
    mem = get_memory()
    log.info({ 'memory': { 'ram': mem.ram, 'gpu': mem.gpu } })


if __name__ == '__main__':
    parser = argparse.ArgumentParser(description = 'train lora')
    parser.add_argument('--model', type=str, default=None, required=True, help='original model to use a base for training')
    parser.add_argument('--input', type=str, default=None, required=True, help='input folder with training images')
    parser.add_argument('--dir', type=str, default=None, required=True, help='folder containing lora checkpoints')
    parser.add_argument('--name', type=str, default=None, required=True, help='lora name')
    parser.add_argument('--interim', type=int, default=0, help = 'save interim checkpoints after n epoch')
    parser.add_argument('--noprocess', default = False, action='store_true', help = 'skip processing and use existing input data')
    parser.add_argument('--nocaptions', default = False, action='store_true', help = 'skip creating captions and tags')
    parser.add_argument('--nolatents', default = False, action='store_true', help = 'skip generating vae latents')
    parser.add_argument('--gradient', type=int, default=1, required=False, help='gradient accumulation steps, default: %(default)s')
    parser.add_argument('--steps', type=int, default=5000, required=False, help='training steps, default: %(default)s')
    parser.add_argument('--dim', type=int, default=128, required=False, help='network dimension, default: %(default)s')
    parser.add_argument('--lr', type=float, default=1e-04, required=False, help='model learning rate, default: %(default)s')
    parser.add_argument('--unetlr', type=float, default=1e-04, required=False, help='unet learning rate, default: %(default)s')
    parser.add_argument('--textlr', type=float, default=5e-05, required=False, help='text encoder learning rate, default: %(default)s')
    args = parser.parse_args()
    if not os.path.exists(args.model) or not os.path.isfile(args.model):
        log.error({ 'lora cannot find model': args.model })
        exit(1)
    options.pretrained_model_name_or_path = args.model
    if not os.path.exists(args.input) or not os.path.isdir(args.input):
        log.error({ 'lora cannot find training dir': args.input })
        exit(1)
    if not os.path.exists(args.dir) or not os.path.isdir(args.dir):
        log.error({ 'lora cannot find training dir': args.dir })
        exit(1)
    options.output_dir = args.dir
    options.output_name = args.name
    options.max_train_steps = args.steps
    options.network_dim = args.dim
    options.gradient_accumulation_steps = args.gradient
    options.save_every_n_epochs = args.interim if args.interim > 0 else None
    options.learning_rate = args.lr
    options.unet_lr = args.unetlr
    options.text_encoder_lr = args.textlr
    log.info({ 'train lora args': vars(options) })
    transformers.logging.set_verbosity_error()
    mem_stats()

    if args.noprocess:
        dir = args.input
        options.train_data_dir = dir
        options.in_json = None
    else:
        dir = os.path.join(tempfile.gettempdir(), args.name, '10_processed')
        Path(dir).mkdir(parents=True, exist_ok=True)

        # preprocess
        for root, _sub_dirs, folder in os.walk(args.input):
            files = [os.path.join(root, f) for f in folder]
        for f in files:
            res, metadata = process.process_file(f = f, dst = dir, preview = False, offline = True)
        process.unload_models()
        options.train_data_dir = os.path.join(tempfile.gettempdir(), args.name)
        mem_stats()

    if not args.nocaptions:
        # interrogate
        for root, _sub_dirs, folder in os.walk(dir):
            files = [os.path.join(root, f) for f in folder]
        metadata = multiinterrogate.interrogate_files(Map({ 'input': dir, 'json': '', 'tag': args.name }), files)
        json_file = os.path.join(dir, args.name + '.json')
        with open(json_file, "w") as outfile:
            outfile.write(json.dumps(metadata, indent=2))
        multiinterrogate.unload_model()
        mem_stats()
        options.in_json = json_file

        log.info({ 'processed': res, 'inputs': len(files), 'metadata': json_file })

    if not args.nolatents:
        # create latents
        lora_latents.create_vae_latents(Map({ 'input': dir, 'json': json_file }))
        lora_latents.unload_vae()
        mem_stats()

    train(options)
    mem_stats()