From c531d827725015d43257dc654dc866d5b3717771 Mon Sep 17 00:00:00 2001 From: nikoferro Date: Wed, 22 Feb 2023 23:17:27 +0100 Subject: [PATCH] feat: update tool_transfer_control.py to support cli args --- tool_transfer_control.py | 98 ++++++++++++++++++++++------------------ 1 file changed, 55 insertions(+), 43 deletions(-) diff --git a/tool_transfer_control.py b/tool_transfer_control.py index b84442cc93..29bf69fa32 100644 --- a/tool_transfer_control.py +++ b/tool_transfer_control.py @@ -1,28 +1,12 @@ -path_sd15 = './models/v1-5-pruned.ckpt' -path_sd15_with_control = './models/control_sd15_openpose.pth' -path_input = './models/anything-v3-full.safetensors' -path_output = './models/control_any3_openpose.pth' - - +import argparse import os - - -assert os.path.exists(path_sd15), 'Input path_sd15 does not exists!' -assert os.path.exists(path_sd15_with_control), 'Input path_sd15_with_control does not exists!' -assert os.path.exists(path_input), 'Input path_input does not exists!' -assert os.path.exists(os.path.dirname(path_output)), 'Output folder not exists!' - +import sys import torch from share import * from cldm.model import load_state_dict -sd15_state_dict = load_state_dict(path_sd15) -sd15_with_control_state_dict = load_state_dict(path_sd15_with_control) -input_state_dict = load_state_dict(path_input) - - def get_node_name(name, parent_name): if len(name) <= len(parent_name): return False, '' @@ -31,29 +15,57 @@ def get_node_name(name, parent_name): return False, '' return True, name[len(parent_name):] +def parse_args(): + parser = argparse.ArgumentParser(description='Transfer weights from one model to another') + + parser.add_argument('--path_sd15', required=True, type=str, help='Path to sd15 model') + parser.add_argument('--path_sd15_with_control', required=True, type=str, help='Path to sd15 model with control') + parser.add_argument('--path_input', required=True, type=str, help='Path to input model') + parser.add_argument('--path_output', required=True, type=str, help='Path to output transferred model') + + return parser.parse_args() + + +def main(): + args = parse_args() + + for path in [args.path_sd15, args.path_sd15_with_control, args.path_input]: + if not os.path.exists(path): + print(f"Error: Input path '{path}' does not exist!") + sys.exit(1) + + if not os.path.exists(os.path.dirname(args.path_output)): + print("Error: Output folder does not exist!") + sys.exit(1) + + sd15_state_dict = load_state_dict(args.path_sd15) + sd15_with_control_state_dict = load_state_dict(args.path_sd15_with_control) + input_state_dict = load_state_dict(args.path_input) + + keys = sd15_with_control_state_dict.keys() + + final_state_dict = {} + for key in keys: + is_first_stage, _ = get_node_name(key, 'first_stage_model') + is_cond_stage, _ = get_node_name(key, 'cond_stage_model') + if is_first_stage or is_cond_stage: + final_state_dict[key] = input_state_dict[key] + continue + p = sd15_with_control_state_dict[key] + is_control, node_name = get_node_name(key, 'control_') + if is_control: + sd15_key_name = 'model.diffusion_' + node_name + else: + sd15_key_name = key + if sd15_key_name in input_state_dict: + p_new = p + input_state_dict[sd15_key_name] - sd15_state_dict[sd15_key_name] + else: + p_new = p + final_state_dict[key] = p_new + + torch.save(final_state_dict, args.path_output) + print(f'Transferred model saved at {args.path_output}') + -keys = sd15_with_control_state_dict.keys() - -final_state_dict = {} -for key in keys: - is_first_stage, _ = get_node_name(key, 'first_stage_model') - is_cond_stage, _ = get_node_name(key, 'cond_stage_model') - if is_first_stage or is_cond_stage: - final_state_dict[key] = input_state_dict[key] - continue - p = sd15_with_control_state_dict[key] - is_control, node_name = get_node_name(key, 'control_') - if is_control: - sd15_key_name = 'model.diffusion_' + node_name - else: - sd15_key_name = key - if sd15_key_name in input_state_dict: - p_new = p + input_state_dict[sd15_key_name] - sd15_state_dict[sd15_key_name] - # print(f'Offset clone from [{sd15_key_name}] to [{key}]') - else: - p_new = p - # print(f'Direct clone to [{key}]') - final_state_dict[key] = p_new - -torch.save(final_state_dict, path_output) -print('Transferred model saved at ' + path_output) +if __name__ == '__main__': + main()