From 2aa485b5afb13fd6aab79777e4dfc488591b2f1c Mon Sep 17 00:00:00 2001 From: Kohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com> Date: Mon, 9 Oct 2023 22:52:09 +0800 Subject: add lora bundle system --- extensions-builtin/Lora/network.py | 1 + extensions-builtin/Lora/networks.py | 48 +++++++++++++++++++++++++++++++++++++ 2 files changed, 49 insertions(+) (limited to 'extensions-builtin') diff --git a/extensions-builtin/Lora/network.py b/extensions-builtin/Lora/network.py index d8e8dfb7..6021fd8d 100644 --- a/extensions-builtin/Lora/network.py +++ b/extensions-builtin/Lora/network.py @@ -93,6 +93,7 @@ class Network: # LoraModule self.unet_multiplier = 1.0 self.dyn_dim = None self.modules = {} + self.bundle_embeddings = {} self.mtime = None self.mentioned_name = None diff --git a/extensions-builtin/Lora/networks.py b/extensions-builtin/Lora/networks.py index 315682b3..652b8ebe 100644 --- a/extensions-builtin/Lora/networks.py +++ b/extensions-builtin/Lora/networks.py @@ -15,6 +15,7 @@ import torch from typing import Union from modules import shared, devices, sd_models, errors, scripts, sd_hijack +from modules.textual_inversion.textual_inversion import Embedding module_types = [ network_lora.ModuleTypeLora(), @@ -149,9 +150,15 @@ def load_network(name, network_on_disk): is_sd2 = 'model_transformer_resblocks' in shared.sd_model.network_layer_mapping matched_networks = {} + bundle_embeddings = {} for key_network, weight in sd.items(): key_network_without_network_parts, network_part = key_network.split(".", 1) + if key_network_without_network_parts == "bundle_emb": + emb_name, vec_name = network_part.split(".", 1) + emb_dict = bundle_embeddings.get(emb_name, {}) + emb_dict[vec_name] = weight + bundle_embeddings[emb_name] = emb_dict key = convert_diffusers_name_to_compvis(key_network_without_network_parts, is_sd2) sd_module = shared.sd_model.network_layer_mapping.get(key, None) @@ -195,6 +202,8 @@ def load_network(name, network_on_disk): net.modules[key] = net_module + net.bundle_embeddings = bundle_embeddings + if keys_failed_to_match: logging.debug(f"Network {network_on_disk.filename} didn't match keys: {keys_failed_to_match}") @@ -210,11 +219,14 @@ def purge_networks_from_memory(): def load_networks(names, te_multipliers=None, unet_multipliers=None, dyn_dims=None): + emb_db = sd_hijack.model_hijack.embedding_db already_loaded = {} for net in loaded_networks: if net.name in names: already_loaded[net.name] = net + for emb_name in net.bundle_embeddings: + emb_db.register_embedding_by_name(None, shared.sd_model, emb_name) loaded_networks.clear() @@ -257,6 +269,41 @@ def load_networks(names, te_multipliers=None, unet_multipliers=None, dyn_dims=No net.dyn_dim = dyn_dims[i] if dyn_dims else 1.0 loaded_networks.append(net) + for emb_name, data in net.bundle_embeddings.items(): + # textual inversion embeddings + if 'string_to_param' in data: + param_dict = data['string_to_param'] + param_dict = getattr(param_dict, '_parameters', param_dict) # fix for torch 1.12.1 loading saved file from torch 1.11 + assert len(param_dict) == 1, 'embedding file has multiple terms in it' + emb = next(iter(param_dict.items()))[1] + vec = emb.detach().to(devices.device, dtype=torch.float32) + shape = vec.shape[-1] + vectors = vec.shape[0] + elif type(data) == dict and 'clip_g' in data and 'clip_l' in data: # SDXL embedding + vec = {k: v.detach().to(devices.device, dtype=torch.float32) for k, v in data.items()} + shape = data['clip_g'].shape[-1] + data['clip_l'].shape[-1] + vectors = data['clip_g'].shape[0] + elif type(data) == dict and type(next(iter(data.values()))) == torch.Tensor: # diffuser concepts + assert len(data.keys()) == 1, 'embedding file has multiple terms in it' + + emb = next(iter(data.values())) + if len(emb.shape) == 1: + emb = emb.unsqueeze(0) + vec = emb.detach().to(devices.device, dtype=torch.float32) + shape = vec.shape[-1] + vectors = vec.shape[0] + else: + raise Exception(f"Couldn't identify {emb_name} in lora: {name} as neither textual inversion embedding nor diffuser concept.") + + embedding = Embedding(vec, emb_name) + embedding.vectors = vectors + embedding.shape = shape + + if emb_db.expected_shape == -1 or emb_db.expected_shape == embedding.shape: + emb_db.register_embedding(embedding, shared.sd_model) + else: + emb_db.skipped_embeddings[name] = embedding + if failed_to_load_networks: sd_hijack.model_hijack.comments.append("Networks not found: " + ", ".join(failed_to_load_networks)) @@ -565,6 +612,7 @@ extra_network_lora = None available_networks = {} available_network_aliases = {} loaded_networks = [] +loaded_bundle_embeddings = {} networks_in_memory = {} available_network_hash_lookup = {} forbidden_network_aliases = {} -- cgit v1.2.1 From 3d8b1af6beb9015f6b3573661d8ed00275f6129f Mon Sep 17 00:00:00 2001 From: Kohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com> Date: Tue, 10 Oct 2023 12:09:33 +0800 Subject: Support string_to_param nested dict format: bundle_emb.EMBNAME.string_to_param.KEYNAME --- extensions-builtin/Lora/networks.py | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) (limited to 'extensions-builtin') diff --git a/extensions-builtin/Lora/networks.py b/extensions-builtin/Lora/networks.py index 652b8ebe..ab3517d8 100644 --- a/extensions-builtin/Lora/networks.py +++ b/extensions-builtin/Lora/networks.py @@ -157,7 +157,11 @@ def load_network(name, network_on_disk): if key_network_without_network_parts == "bundle_emb": emb_name, vec_name = network_part.split(".", 1) emb_dict = bundle_embeddings.get(emb_name, {}) - emb_dict[vec_name] = weight + if vec_name.split('.')[0] == 'string_to_param': + _, k2 = vec_name.split('.', 1) + emb_dict['string_to_param'] = {k2: weight} + else: + emb_dict[vec_name] = weight bundle_embeddings[emb_name] = emb_dict key = convert_diffusers_name_to_compvis(key_network_without_network_parts, is_sd2) @@ -301,6 +305,7 @@ def load_networks(names, te_multipliers=None, unet_multipliers=None, dyn_dims=No if emb_db.expected_shape == -1 or emb_db.expected_shape == embedding.shape: emb_db.register_embedding(embedding, shared.sd_model) + print(f'registered bundle embedding: {embedding.name}') else: emb_db.skipped_embeddings[name] = embedding -- cgit v1.2.1 From 2282eb8dd5905e8ed71231a0b8fc77599d10c12f Mon Sep 17 00:00:00 2001 From: Kohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com> Date: Tue, 10 Oct 2023 12:11:00 +0800 Subject: Remove dev debug print --- extensions-builtin/Lora/networks.py | 1 - 1 file changed, 1 deletion(-) (limited to 'extensions-builtin') diff --git a/extensions-builtin/Lora/networks.py b/extensions-builtin/Lora/networks.py index ab3517d8..465e24c8 100644 --- a/extensions-builtin/Lora/networks.py +++ b/extensions-builtin/Lora/networks.py @@ -305,7 +305,6 @@ def load_networks(names, te_multipliers=None, unet_multipliers=None, dyn_dims=No if emb_db.expected_shape == -1 or emb_db.expected_shape == embedding.shape: emb_db.register_embedding(embedding, shared.sd_model) - print(f'registered bundle embedding: {embedding.name}') else: emb_db.skipped_embeddings[name] = embedding -- cgit v1.2.1 From 81e94de3185d42dba4e7bb72cf836f683f28b03f Mon Sep 17 00:00:00 2001 From: Kohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com> Date: Tue, 10 Oct 2023 14:44:20 +0800 Subject: Add warning when meet emb name conflicting Choose standalone embedding (in /embeddings folder) first --- extensions-builtin/Lora/lora_logger.py | 33 ++++++++++++++ extensions-builtin/Lora/networks.py | 80 ++++++++++++++++++++-------------- 2 files changed, 81 insertions(+), 32 deletions(-) create mode 100644 extensions-builtin/Lora/lora_logger.py (limited to 'extensions-builtin') diff --git a/extensions-builtin/Lora/lora_logger.py b/extensions-builtin/Lora/lora_logger.py new file mode 100644 index 00000000..d50e90f0 --- /dev/null +++ b/extensions-builtin/Lora/lora_logger.py @@ -0,0 +1,33 @@ +import sys +import copy +import logging + + +class ColoredFormatter(logging.Formatter): + COLORS = { + "DEBUG": "\033[0;36m", # CYAN + "INFO": "\033[0;32m", # GREEN + "WARNING": "\033[0;33m", # YELLOW + "ERROR": "\033[0;31m", # RED + "CRITICAL": "\033[0;37;41m", # WHITE ON RED + "RESET": "\033[0m", # RESET COLOR + } + + def format(self, record): + colored_record = copy.copy(record) + levelname = colored_record.levelname + seq = self.COLORS.get(levelname, self.COLORS["RESET"]) + colored_record.levelname = f"{seq}{levelname}{self.COLORS['RESET']}" + return super().format(colored_record) + + +logger = logging.getLogger("lora") +logger.propagate = False + + +if not logger.handlers: + handler = logging.StreamHandler(sys.stdout) + handler.setFormatter( + ColoredFormatter("[%(name)s]-%(levelname)s: %(message)s") + ) + logger.addHandler(handler) \ No newline at end of file diff --git a/extensions-builtin/Lora/networks.py b/extensions-builtin/Lora/networks.py index 465e24c8..12f70576 100644 --- a/extensions-builtin/Lora/networks.py +++ b/extensions-builtin/Lora/networks.py @@ -17,6 +17,8 @@ from typing import Union from modules import shared, devices, sd_models, errors, scripts, sd_hijack from modules.textual_inversion.textual_inversion import Embedding +from lora_logger import logger + module_types = [ network_lora.ModuleTypeLora(), network_hada.ModuleTypeHada(), @@ -206,7 +208,40 @@ def load_network(name, network_on_disk): net.modules[key] = net_module - net.bundle_embeddings = bundle_embeddings + embeddings = {} + for emb_name, data in bundle_embeddings.items(): + # textual inversion embeddings + if 'string_to_param' in data: + param_dict = data['string_to_param'] + param_dict = getattr(param_dict, '_parameters', param_dict) # fix for torch 1.12.1 loading saved file from torch 1.11 + assert len(param_dict) == 1, 'embedding file has multiple terms in it' + emb = next(iter(param_dict.items()))[1] + vec = emb.detach().to(devices.device, dtype=torch.float32) + shape = vec.shape[-1] + vectors = vec.shape[0] + elif type(data) == dict and 'clip_g' in data and 'clip_l' in data: # SDXL embedding + vec = {k: v.detach().to(devices.device, dtype=torch.float32) for k, v in data.items()} + shape = data['clip_g'].shape[-1] + data['clip_l'].shape[-1] + vectors = data['clip_g'].shape[0] + elif type(data) == dict and type(next(iter(data.values()))) == torch.Tensor: # diffuser concepts + assert len(data.keys()) == 1, 'embedding file has multiple terms in it' + + emb = next(iter(data.values())) + if len(emb.shape) == 1: + emb = emb.unsqueeze(0) + vec = emb.detach().to(devices.device, dtype=torch.float32) + shape = vec.shape[-1] + vectors = vec.shape[0] + else: + raise Exception(f"Couldn't identify {emb_name} in lora: {name} as neither textual inversion embedding nor diffuser concept.") + + embedding = Embedding(vec, emb_name) + embedding.vectors = vectors + embedding.shape = shape + embedding.loaded = None + embeddings[emb_name] = embedding + + net.bundle_embeddings = embeddings if keys_failed_to_match: logging.debug(f"Network {network_on_disk.filename} didn't match keys: {keys_failed_to_match}") @@ -229,8 +264,9 @@ def load_networks(names, te_multipliers=None, unet_multipliers=None, dyn_dims=No for net in loaded_networks: if net.name in names: already_loaded[net.name] = net - for emb_name in net.bundle_embeddings: - emb_db.register_embedding_by_name(None, shared.sd_model, emb_name) + for emb_name, embedding in net.bundle_embeddings.items(): + if embedding.loaded: + emb_db.register_embedding_by_name(None, shared.sd_model, emb_name) loaded_networks.clear() @@ -273,37 +309,17 @@ def load_networks(names, te_multipliers=None, unet_multipliers=None, dyn_dims=No net.dyn_dim = dyn_dims[i] if dyn_dims else 1.0 loaded_networks.append(net) - for emb_name, data in net.bundle_embeddings.items(): - # textual inversion embeddings - if 'string_to_param' in data: - param_dict = data['string_to_param'] - param_dict = getattr(param_dict, '_parameters', param_dict) # fix for torch 1.12.1 loading saved file from torch 1.11 - assert len(param_dict) == 1, 'embedding file has multiple terms in it' - emb = next(iter(param_dict.items()))[1] - vec = emb.detach().to(devices.device, dtype=torch.float32) - shape = vec.shape[-1] - vectors = vec.shape[0] - elif type(data) == dict and 'clip_g' in data and 'clip_l' in data: # SDXL embedding - vec = {k: v.detach().to(devices.device, dtype=torch.float32) for k, v in data.items()} - shape = data['clip_g'].shape[-1] + data['clip_l'].shape[-1] - vectors = data['clip_g'].shape[0] - elif type(data) == dict and type(next(iter(data.values()))) == torch.Tensor: # diffuser concepts - assert len(data.keys()) == 1, 'embedding file has multiple terms in it' - - emb = next(iter(data.values())) - if len(emb.shape) == 1: - emb = emb.unsqueeze(0) - vec = emb.detach().to(devices.device, dtype=torch.float32) - shape = vec.shape[-1] - vectors = vec.shape[0] - else: - raise Exception(f"Couldn't identify {emb_name} in lora: {name} as neither textual inversion embedding nor diffuser concept.") - - embedding = Embedding(vec, emb_name) - embedding.vectors = vectors - embedding.shape = shape + for emb_name, embedding in net.bundle_embeddings.items(): + if embedding.loaded is None and emb_name in emb_db.word_embeddings: + logger.warning( + f'Skip bundle embedding: "{emb_name}"' + ' as it was already loaded from embeddings folder' + ) + continue + embedding.loaded = False if emb_db.expected_shape == -1 or emb_db.expected_shape == embedding.shape: + embedding.loaded = True emb_db.register_embedding(embedding, shared.sd_model) else: emb_db.skipped_embeddings[name] = embedding -- cgit v1.2.1 From 891ccb767c3815db48a124677d1cd0f204018ad4 Mon Sep 17 00:00:00 2001 From: Kohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com> Date: Tue, 10 Oct 2023 15:07:25 +0800 Subject: Fix lint --- extensions-builtin/Lora/lora_logger.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'extensions-builtin') diff --git a/extensions-builtin/Lora/lora_logger.py b/extensions-builtin/Lora/lora_logger.py index d50e90f0..d51de297 100644 --- a/extensions-builtin/Lora/lora_logger.py +++ b/extensions-builtin/Lora/lora_logger.py @@ -30,4 +30,4 @@ if not logger.handlers: handler.setFormatter( ColoredFormatter("[%(name)s]-%(levelname)s: %(message)s") ) - logger.addHandler(handler) \ No newline at end of file + logger.addHandler(handler) -- cgit v1.2.1 From a8cbe50c9fa324ed887089e4333452ecc4355c92 Mon Sep 17 00:00:00 2001 From: AUTOMATIC1111 <16777216c@gmail.com> Date: Sat, 14 Oct 2023 12:14:56 +0300 Subject: remove duplicated code --- extensions-builtin/Lora/networks.py | 31 ++----------------------------- 1 file changed, 2 insertions(+), 29 deletions(-) (limited to 'extensions-builtin') diff --git a/extensions-builtin/Lora/networks.py b/extensions-builtin/Lora/networks.py index 12f70576..d5f0f9f1 100644 --- a/extensions-builtin/Lora/networks.py +++ b/extensions-builtin/Lora/networks.py @@ -15,7 +15,7 @@ import torch from typing import Union from modules import shared, devices, sd_models, errors, scripts, sd_hijack -from modules.textual_inversion.textual_inversion import Embedding +import modules.textual_inversion.textual_inversion as textual_inversion from lora_logger import logger @@ -210,34 +210,7 @@ def load_network(name, network_on_disk): embeddings = {} for emb_name, data in bundle_embeddings.items(): - # textual inversion embeddings - if 'string_to_param' in data: - param_dict = data['string_to_param'] - param_dict = getattr(param_dict, '_parameters', param_dict) # fix for torch 1.12.1 loading saved file from torch 1.11 - assert len(param_dict) == 1, 'embedding file has multiple terms in it' - emb = next(iter(param_dict.items()))[1] - vec = emb.detach().to(devices.device, dtype=torch.float32) - shape = vec.shape[-1] - vectors = vec.shape[0] - elif type(data) == dict and 'clip_g' in data and 'clip_l' in data: # SDXL embedding - vec = {k: v.detach().to(devices.device, dtype=torch.float32) for k, v in data.items()} - shape = data['clip_g'].shape[-1] + data['clip_l'].shape[-1] - vectors = data['clip_g'].shape[0] - elif type(data) == dict and type(next(iter(data.values()))) == torch.Tensor: # diffuser concepts - assert len(data.keys()) == 1, 'embedding file has multiple terms in it' - - emb = next(iter(data.values())) - if len(emb.shape) == 1: - emb = emb.unsqueeze(0) - vec = emb.detach().to(devices.device, dtype=torch.float32) - shape = vec.shape[-1] - vectors = vec.shape[0] - else: - raise Exception(f"Couldn't identify {emb_name} in lora: {name} as neither textual inversion embedding nor diffuser concept.") - - embedding = Embedding(vec, emb_name) - embedding.vectors = vectors - embedding.shape = shape + embedding = textual_inversion.create_embedding_from_data(data, emb_name, filename=network_on_disk.filename + "/" + emb_name) embedding.loaded = None embeddings[emb_name] = embedding -- cgit v1.2.1