File size: 3,218 Bytes
6d613c2 | 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 | import imp
import os
base = imp.load_source("base", os.path.join(os.path.dirname(__file__), "base.py"))
def get_config(name):
return globals()[name]()
def _get_config(base_model="kontext", n_gpus=1, gradient_step_per_epoch=1, reward_fn={}, name=""):
config = base.get_config()
config.base_model = base_model
config.dataset = "../edit-r1-dataset"
config.pretrained.model = "black-forest-labs/FLUX.1-Kontext-dev"
config.sample.num_steps = 6
config.sample.eval_num_steps = 15
config.sample.guidance_scale = 2.5
config.resolution = 512
config.train.beta = 0.0001
config.sample.noise_level = 0.7
bsz = 3
config.sample.num_image_per_prompt = 12
config.sample.ban_std_thres = 0.05
config.sample.ban_mean_thres = 0.9
config.sample.ban_prompt = False
num_groups = 24
while True:
if bsz < 1:
assert False, "Cannot find a proper batch size."
if (
num_groups * config.sample.num_image_per_prompt % (n_gpus * bsz) == 0
and bsz * n_gpus % config.sample.num_image_per_prompt == 0
):
n_batch_per_epoch = num_groups * config.sample.num_image_per_prompt // (n_gpus * bsz)
if n_batch_per_epoch % gradient_step_per_epoch == 0:
config.sample.train_batch_size = bsz
config.sample.num_batches_per_epoch = n_batch_per_epoch
config.train.batch_size = config.sample.train_batch_size
config.train.gradient_accumulation_steps = (
config.sample.num_batches_per_epoch // gradient_step_per_epoch
)
break
bsz -= 1
# special design, the test set has a total of 1018/2212/2048 for ocr/geneval/pickscore, to make gpu_num*bs*n as close as possible to it, because when the number of samples cannot be divided evenly by the number of cards, multi-card will fill the last batch to ensure each card has the same number of samples, affecting gradient synchronization.
config.sample.test_batch_size = bsz
if n_gpus > 32:
config.sample.test_batch_size = config.sample.test_batch_size // 2
config.prompt_fn = "geneval"
config.run_name = f"nft_{base_model}_{name}"
config.save_dir = f"logs/nft/{base_model}/{name}"
config.reward_fn = reward_fn
config.decay_type = 1
config.beta = 1.0
config.train.adv_mode = "all"
# config.sample.guidance_scale = 1.0
config.sample.deterministic = True
config.sample.solver = "dpm2"
return config
def kontext_mllm_reward():
reward_fn = {
"mllm_score_continue": 1.0,
}
config = _get_config(
base_model="kontext",
n_gpus=24,
gradient_step_per_epoch=1,
reward_fn=reward_fn,
name="mllm_score_continue",
)
return config
def kontext_mllm_reward_ban_prompt():
reward_fn = {
"mllm_score_continue": 1.0,
}
config = _get_config(
base_model="kontext",
n_gpus=24,
gradient_step_per_epoch=1,
reward_fn=reward_fn,
name="mllm_score_continue_ban_prompt",
)
config.sample.ban_prompt = True
config.sample.ban_std_thres = 0.05
return config |