import gradio as gr
import os
from pathlib import Path
import argparse
import shutil
from train_dreambooth import run_training
from convertosd import convert
from PIL import Image
import torch
css = '''
.instruction{position: absolute; top: 0;right: 0;margin-top: 0px !important}
.arrow{position: absolute;top: 0;right: -8px;margin-top: -8px !important}
#component-4, #component-3, #component-10{min-height: 0}
'''
shutil.unpack_archive("mix.zip", "mix")
model_to_load = "multimodalart/sd-fine-tunable"
maximum_concepts = 3
def swap_text(option):
mandatory_liability = "You must have the right to do so and you are liable for the images you use"
if(option == "object"):
instance_prompt_example = "cttoy"
freeze_for = 50
return [f"You are going to train `object`(s), upload 5-10 images of each object you are planning on training on from different angles/perspectives. {mandatory_liability}:", '''''', f"You should name your concept with a unique made up word that has low chance of the model already knowing it (e.g.: `{instance_prompt_example}` here). Images will be automatically cropped to 512x512.", freeze_for]
elif(option == "person"):
instance_prompt_example = "julcto"
freeze_for = 100
return [f"You are going to train a `person`(s), upload 10-20 images of each person you are planning on training on from different angles/perspectives. {mandatory_liability}:", '''
''', f"You should name the files with a unique word that represent your concept (e.g.: `{instance_prompt_example}` here). Images will be automatically cropped to 512x512.", freeze_for]
elif(option == "style"):
instance_prompt_example = "trsldamrl"
freeze_for = 10
return [f"You are going to train a `style`, upload 10-20 images of the style you are planning on training on. Name the files with the words you would like {mandatory_liability}:", '''
''', f"You should name your files with a unique word that represent your concept (e.g.: `{instance_prompt_example}` here). Images will be automatically cropped to 512x512.", freeze_for]
def count_files(*inputs):
file_counter = 0
for i, input in enumerate(inputs):
if(i < maximum_concepts-1):
if(input):
files = inputs[i+(maximum_concepts*2)]
for j, tile_temp in enumerate(files):
file_counter+= 1
uses_custom = inputs[-1]
type_of_thing = inputs[-4]
if(uses_custom):
Training_Steps = int(inputs[-3])
else:
if(type_of_thing == "person"):
Training_Steps = file_counter*200*2
else:
Training_Steps = file_counter*200
return(gr.update(visible=True, value=f"You are going to train {file_counter} files for {Training_Steps} steps. This should take around {round(Training_Steps/1.5, 2)} seconds, or {round((Training_Steps/1.5)/3600, 2)}. The T4 GPU costs US$0.60 for 1h, so the estimated costs for this training run should be {round(((Training_Steps/1.5)/3600)*0.6, 2)}"))
def train(*inputs):
if os.path.exists("diffusers_model.zip"): os.remove("diffusers_model.zip")
if os.path.exists("model.ckpt"): os.remove("model.ckpt")
file_counter = 0
for i, input in enumerate(inputs):
if(i < maximum_concepts-1):
if(input):
os.makedirs('instance_images',exist_ok=True)
files = inputs[i+(maximum_concepts*2)]
prompt = inputs[i+maximum_concepts]
for j, file_temp in enumerate(files):
file = Image.open(file_temp.name)
width, height = file.size
side_length = min(width, height)
left = (width - side_length)/2
top = (height - side_length)/2
right = (width + side_length)/2
bottom = (height + side_length)/2
image = file.crop((left, top, right, bottom))
image = image.resize((512, 512))
extension = file_temp.name.split(".")[1]
image = image.convert('RGB')
image.save(f'instance_images/{prompt}_({j+1}).jpg', format="JPEG", quality = 100)
file_counter += 1
os.makedirs('output_model',exist_ok=True)
uses_custom = inputs[-1]
type_of_thing = inputs[-4]
if(uses_custom):
Training_Steps = int(inputs[-3])
Train_text_encoder_for = int(inputs[-2])
else:
Training_Steps = file_counter*200
if(type_of_thing == "person"):
class_data_dir = "mix"
Train_text_encoder_for=100
args_txt_encoder = argparse.Namespace(
image_captions_filename = True,
train_text_encoder = True,
pretrained_model_name_or_path=model_to_load,
instance_data_dir="instance_images",
class_data_dir=class_data_dir,
output_dir="output_model",
with_prior_preservation=True,
prior_loss_weight=1.0,
instance_prompt="",
seed=42,
resolution=512,
mixed_precision="fp16",
train_batch_size=1,
gradient_accumulation_steps=1,
gradient_checkpointing=True,
use_8bit_adam=True,
learning_rate=2e-6,
lr_scheduler="polynomial",
lr_warmup_steps=0,
max_train_steps=Training_Steps,
num_class_images=200
)
args_unet = argparse.Namespace(
image_captions_filename = True,
train_only_unet=True,
Session_dir="output_model",
save_starting_step=0,
save_n_steps=0,
pretrained_model_name_or_path=model_to_load,
instance_data_dir="instance_images",
output_dir="output_model",
instance_prompt="",
seed=42,
resolution=512,
mixed_precision="fp16",
train_batch_size=1,
gradient_accumulation_steps=1,
gradient_checkpointing=False,
use_8bit_adam=True,
learning_rate=2e-6,
lr_scheduler="polynomial",
lr_warmup_steps=0,
max_train_steps=Training_Steps
)
run_training(args_txt_encoder)
run_training(args_unet)
elif(type_of_thing == "object" or type_of_thing == "style"):
if(type_of_thing == "object"):
Train_text_encoder_for=30
elif(type_of_thing == "style"):
Train_text_encoder_for=15
class_data_dir = None
stptxt = int((Training_Steps*Train_text_encoder_for)/100)
args_general = argparse.Namespace(
image_captions_filename = True,
train_text_encoder = True,
stop_text_encoder_training = stptxt,
save_n_steps = 0,
pretrained_model_name_or_path = model_to_load,
instance_data_dir="instance_images",
class_data_dir=class_data_dir,
output_dir="output_model",
instance_prompt="",
seed=42,
resolution=512,
mixed_precision="fp16",
train_batch_size=1,
gradient_accumulation_steps=1,
use_8bit_adam=True,
learning_rate=2e-6,
lr_scheduler="polynomial",
lr_warmup_steps = 0,
max_train_steps=Training_Steps,
)
run_training(args_general)
torch.cuda.empty_cache()
#convert("output_model", "model.ckpt")
shutil.rmtree('instance_images')
shutil.make_archive("diffusers_model", 'zip', "output_model")
torch.cuda.empty_cache()
return [gr.update(visible=True, value=["diffusers_model.zip"]), gr.update(visible=True), gr.update(visible=True), gr.update(visible=True)]
def generate(prompt):
from diffusers import StableDiffusionPipeline
pipe = StableDiffusionPipeline.from_pretrained("./output_model", torch_dtype=torch.float16)
pipe = pipe.to("cuda")
image = pipe(prompt).images[0]
return(image)
def push(path):
pass
def convert_to_ckpt():
convert("output_model", "model.ckpt")
return gr.update(visible=True, value=["diffusers_model.zip", "model.ckpt"])
with gr.Blocks(css=css) as demo:
with gr.Box():
if "IS_SHARED_UI" in os.environ:
gr.HTML('''
For it to work, you have to duplicate the Space and run it on your own profile where a (paid) private GPU will be attributed to it during runtime. It will cost you < US$1 to train a model on default settings! 🤑
Now you can attribute a T4 GPU to it (by going to the Settings tab) and run the training below. The GPU will be automatically unassigned after training is over. So you will be billed by the minute between when you activate the GPU and when it finishes training.