From d8acd34f66ab35a91f10d66330bcc95a83bfcac6 Mon Sep 17 00:00:00 2001 From: AngelBottomless <35677394+aria1th@users.noreply.github.com> Date: Thu, 20 Oct 2022 23:43:03 +0900 Subject: [PATCH 1/4] generalized some functions and option for ignoring first layer --- modules/hypernetworks/hypernetwork.py | 23 +++++++++++++++-------- 1 file changed, 15 insertions(+), 8 deletions(-) diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index 7d617680f4d..3a44b3775a8 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -21,21 +21,27 @@ class HypernetworkModule(torch.nn.Module): multiplier = 1.0 - + activation_dict = {"relu": torch.nn.ReLU, "leakyrelu": torch.nn.LeakyReLU, "elu": torch.nn.ELU, + "swish": torch.nn.Hardswish} + def __init__(self, dim, state_dict=None, layer_structure=None, add_layer_norm=False, activation_func=None): super().__init__() assert layer_structure is not None, "layer_structure must not be None" assert layer_structure[0] == 1, "Multiplier Sequence should start with size 1!" assert layer_structure[-1] == 1, "Multiplier Sequence should end with size 1!" - + linears = [] for i in range(len(layer_structure) - 1): linears.append(torch.nn.Linear(int(dim * layer_structure[i]), int(dim * layer_structure[i+1]))) - if activation_func == "relu": - linears.append(torch.nn.ReLU()) - if activation_func == "leakyrelu": - linears.append(torch.nn.LeakyReLU()) + # if skip_first_layer because first parameters potentially contain negative values + if i < 1: continue + if activation_func in HypernetworkModule.activation_dict: + linears.append(HypernetworkModule.activation_dict[activation_func]()) + else: + print("Invalid key {} encountered as activation function!".format(activation_func)) + # if use_dropout: + linears.append(torch.nn.Dropout(p=0.3)) if add_layer_norm: linears.append(torch.nn.LayerNorm(int(dim * layer_structure[i+1]))) @@ -46,7 +52,7 @@ def __init__(self, dim, state_dict=None, layer_structure=None, add_layer_norm=Fa self.load_state_dict(state_dict) else: for layer in self.linear: - if not "ReLU" in layer.__str__(): + if isinstance(layer, torch.nn.Linear): layer.weight.data.normal_(mean=0.0, std=0.01) layer.bias.data.zero_() @@ -298,7 +304,8 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log return hypernetwork, filename scheduler = LearnRateScheduler(learn_rate, steps, ititial_step) - optimizer = torch.optim.AdamW(weights, lr=scheduler.learn_rate) + # if optimizer == "Adam": or else Adam / AdamW / etc... + optimizer = torch.optim.Adam(weights, lr=scheduler.learn_rate) pbar = tqdm.tqdm(enumerate(ds), total=steps - ititial_step) for i, entries in pbar: From a71e0212363979c7cbbb797c9fbd5f8cd03b29d3 Mon Sep 17 00:00:00 2001 From: AngelBottomless <35677394+aria1th@users.noreply.github.com> Date: Thu, 20 Oct 2022 23:48:52 +0900 Subject: [PATCH 2/4] only linear --- modules/hypernetworks/hypernetwork.py | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index 3a44b3775a8..905cbeefd33 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -35,13 +35,13 @@ def __init__(self, dim, state_dict=None, layer_structure=None, add_layer_norm=Fa for i in range(len(layer_structure) - 1): linears.append(torch.nn.Linear(int(dim * layer_structure[i]), int(dim * layer_structure[i+1]))) # if skip_first_layer because first parameters potentially contain negative values - if i < 1: continue + # if i < 1: continue if activation_func in HypernetworkModule.activation_dict: linears.append(HypernetworkModule.activation_dict[activation_func]()) else: print("Invalid key {} encountered as activation function!".format(activation_func)) # if use_dropout: - linears.append(torch.nn.Dropout(p=0.3)) + # linears.append(torch.nn.Dropout(p=0.3)) if add_layer_norm: linears.append(torch.nn.LayerNorm(int(dim * layer_structure[i+1]))) @@ -80,7 +80,7 @@ def forward(self, x): def trainables(self): layer_structure = [] for layer in self.linear: - if not "ReLU" in layer.__str__(): + if isinstance(layer, torch.nn.Linear): layer_structure += [layer.weight, layer.bias] return layer_structure @@ -304,8 +304,8 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log return hypernetwork, filename scheduler = LearnRateScheduler(learn_rate, steps, ititial_step) - # if optimizer == "Adam": or else Adam / AdamW / etc... - optimizer = torch.optim.Adam(weights, lr=scheduler.learn_rate) + # if optimizer == "AdamW": or else Adam / AdamW / SGD, etc... + optimizer = torch.optim.AdamW(weights, lr=scheduler.learn_rate) pbar = tqdm.tqdm(enumerate(ds), total=steps - ititial_step) for i, entries in pbar: From 108be15500aac590b4e00420635d7b61fccfa530 Mon Sep 17 00:00:00 2001 From: AngelBottomless <35677394+aria1th@users.noreply.github.com> Date: Fri, 21 Oct 2022 01:00:41 +0900 Subject: [PATCH 3/4] fix bugs and optimizations --- modules/hypernetworks/hypernetwork.py | 105 +++++++++++++++----------- 1 file changed, 59 insertions(+), 46 deletions(-) diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index 905cbeefd33..893ba110f3b 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -36,14 +36,14 @@ def __init__(self, dim, state_dict=None, layer_structure=None, add_layer_norm=Fa linears.append(torch.nn.Linear(int(dim * layer_structure[i]), int(dim * layer_structure[i+1]))) # if skip_first_layer because first parameters potentially contain negative values # if i < 1: continue + if add_layer_norm: + linears.append(torch.nn.LayerNorm(int(dim * layer_structure[i+1]))) if activation_func in HypernetworkModule.activation_dict: linears.append(HypernetworkModule.activation_dict[activation_func]()) else: print("Invalid key {} encountered as activation function!".format(activation_func)) # if use_dropout: # linears.append(torch.nn.Dropout(p=0.3)) - if add_layer_norm: - linears.append(torch.nn.LayerNorm(int(dim * layer_structure[i+1]))) self.linear = torch.nn.Sequential(*linears) @@ -115,11 +115,24 @@ def weights(self): for k, layers in self.layers.items(): for layer in layers: - layer.train() res += layer.trainables() return res + def eval(self): + for k, layers in self.layers.items(): + for layer in layers: + layer.eval() + for items in self.weights(): + items.requires_grad = False + + def train(self): + for k, layers in self.layers.items(): + for layer in layers: + layer.train() + for items in self.weights(): + items.requires_grad = True + def save(self, filename): state_dict = {} @@ -290,10 +303,6 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log shared.sd_model.first_stage_model.to(devices.cpu) hypernetwork = shared.loaded_hypernetwork - weights = hypernetwork.weights() - for weight in weights: - weight.requires_grad = True - losses = torch.zeros((32,)) last_saved_file = "" @@ -304,10 +313,10 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log return hypernetwork, filename scheduler = LearnRateScheduler(learn_rate, steps, ititial_step) - # if optimizer == "AdamW": or else Adam / AdamW / SGD, etc... - optimizer = torch.optim.AdamW(weights, lr=scheduler.learn_rate) + optimizer = torch.optim.AdamW(hypernetwork.weights(), lr=scheduler.learn_rate) pbar = tqdm.tqdm(enumerate(ds), total=steps - ititial_step) + hypernetwork.train() for i, entries in pbar: hypernetwork.step = i + ititial_step @@ -328,8 +337,9 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log losses[hypernetwork.step % losses.shape[0]] = loss.item() - optimizer.zero_grad() + optimizer.zero_grad(set_to_none=True) loss.backward() + del loss optimizer.step() mean_loss = losses.mean() if torch.isnan(mean_loss): @@ -346,44 +356,47 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log }) if hypernetwork.step > 0 and images_dir is not None and hypernetwork.step % create_image_every == 0: + torch.cuda.empty_cache() last_saved_image = os.path.join(images_dir, f'{hypernetwork_name}-{hypernetwork.step}.png') + with torch.no_grad(): + hypernetwork.eval() + shared.sd_model.cond_stage_model.to(devices.device) + shared.sd_model.first_stage_model.to(devices.device) + + p = processing.StableDiffusionProcessingTxt2Img( + sd_model=shared.sd_model, + do_not_save_grid=True, + do_not_save_samples=True, + ) - optimizer.zero_grad() - shared.sd_model.cond_stage_model.to(devices.device) - shared.sd_model.first_stage_model.to(devices.device) - - p = processing.StableDiffusionProcessingTxt2Img( - sd_model=shared.sd_model, - do_not_save_grid=True, - do_not_save_samples=True, - ) - - if preview_from_txt2img: - p.prompt = preview_prompt - p.negative_prompt = preview_negative_prompt - p.steps = preview_steps - p.sampler_index = preview_sampler_index - p.cfg_scale = preview_cfg_scale - p.seed = preview_seed - p.width = preview_width - p.height = preview_height - else: - p.prompt = entries[0].cond_text - p.steps = 20 - - preview_text = p.prompt - - processed = processing.process_images(p) - image = processed.images[0] if len(processed.images)>0 else None - - if unload: - shared.sd_model.cond_stage_model.to(devices.cpu) - shared.sd_model.first_stage_model.to(devices.cpu) - - if image is not None: - shared.state.current_image = image - image.save(last_saved_image) - last_saved_image += f", prompt: {preview_text}" + if preview_from_txt2img: + p.prompt = preview_prompt + p.negative_prompt = preview_negative_prompt + p.steps = preview_steps + p.sampler_index = preview_sampler_index + p.cfg_scale = preview_cfg_scale + p.seed = preview_seed + p.width = preview_width + p.height = preview_height + else: + p.prompt = entries[0].cond_text + p.steps = 20 + + preview_text = p.prompt + + processed = processing.process_images(p) + image = processed.images[0] if len(processed.images)>0 else None + + if unload: + shared.sd_model.cond_stage_model.to(devices.cpu) + shared.sd_model.first_stage_model.to(devices.cpu) + + if image is not None: + shared.state.current_image = image + image.save(last_saved_image) + last_saved_image += f", prompt: {preview_text}" + + hypernetwork.train() shared.state.job_no = hypernetwork.step From f89829ec3a0baceb445451ad98d4fb4323e922aa Mon Sep 17 00:00:00 2001 From: aria1th <35677394+aria1th@users.noreply.github.com> Date: Fri, 21 Oct 2022 01:37:11 +0900 Subject: [PATCH 4/4] Revert "fix bugs and optimizations" This reverts commit 108be15500aac590b4e00420635d7b61fccfa530. --- modules/hypernetworks/hypernetwork.py | 105 +++++++++++--------------- 1 file changed, 46 insertions(+), 59 deletions(-) diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index 893ba110f3b..905cbeefd33 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -36,14 +36,14 @@ def __init__(self, dim, state_dict=None, layer_structure=None, add_layer_norm=Fa linears.append(torch.nn.Linear(int(dim * layer_structure[i]), int(dim * layer_structure[i+1]))) # if skip_first_layer because first parameters potentially contain negative values # if i < 1: continue - if add_layer_norm: - linears.append(torch.nn.LayerNorm(int(dim * layer_structure[i+1]))) if activation_func in HypernetworkModule.activation_dict: linears.append(HypernetworkModule.activation_dict[activation_func]()) else: print("Invalid key {} encountered as activation function!".format(activation_func)) # if use_dropout: # linears.append(torch.nn.Dropout(p=0.3)) + if add_layer_norm: + linears.append(torch.nn.LayerNorm(int(dim * layer_structure[i+1]))) self.linear = torch.nn.Sequential(*linears) @@ -115,24 +115,11 @@ def weights(self): for k, layers in self.layers.items(): for layer in layers: + layer.train() res += layer.trainables() return res - def eval(self): - for k, layers in self.layers.items(): - for layer in layers: - layer.eval() - for items in self.weights(): - items.requires_grad = False - - def train(self): - for k, layers in self.layers.items(): - for layer in layers: - layer.train() - for items in self.weights(): - items.requires_grad = True - def save(self, filename): state_dict = {} @@ -303,6 +290,10 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log shared.sd_model.first_stage_model.to(devices.cpu) hypernetwork = shared.loaded_hypernetwork + weights = hypernetwork.weights() + for weight in weights: + weight.requires_grad = True + losses = torch.zeros((32,)) last_saved_file = "" @@ -313,10 +304,10 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log return hypernetwork, filename scheduler = LearnRateScheduler(learn_rate, steps, ititial_step) - optimizer = torch.optim.AdamW(hypernetwork.weights(), lr=scheduler.learn_rate) + # if optimizer == "AdamW": or else Adam / AdamW / SGD, etc... + optimizer = torch.optim.AdamW(weights, lr=scheduler.learn_rate) pbar = tqdm.tqdm(enumerate(ds), total=steps - ititial_step) - hypernetwork.train() for i, entries in pbar: hypernetwork.step = i + ititial_step @@ -337,9 +328,8 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log losses[hypernetwork.step % losses.shape[0]] = loss.item() - optimizer.zero_grad(set_to_none=True) + optimizer.zero_grad() loss.backward() - del loss optimizer.step() mean_loss = losses.mean() if torch.isnan(mean_loss): @@ -356,47 +346,44 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log }) if hypernetwork.step > 0 and images_dir is not None and hypernetwork.step % create_image_every == 0: - torch.cuda.empty_cache() last_saved_image = os.path.join(images_dir, f'{hypernetwork_name}-{hypernetwork.step}.png') - with torch.no_grad(): - hypernetwork.eval() - shared.sd_model.cond_stage_model.to(devices.device) - shared.sd_model.first_stage_model.to(devices.device) - - p = processing.StableDiffusionProcessingTxt2Img( - sd_model=shared.sd_model, - do_not_save_grid=True, - do_not_save_samples=True, - ) - if preview_from_txt2img: - p.prompt = preview_prompt - p.negative_prompt = preview_negative_prompt - p.steps = preview_steps - p.sampler_index = preview_sampler_index - p.cfg_scale = preview_cfg_scale - p.seed = preview_seed - p.width = preview_width - p.height = preview_height - else: - p.prompt = entries[0].cond_text - p.steps = 20 - - preview_text = p.prompt - - processed = processing.process_images(p) - image = processed.images[0] if len(processed.images)>0 else None - - if unload: - shared.sd_model.cond_stage_model.to(devices.cpu) - shared.sd_model.first_stage_model.to(devices.cpu) - - if image is not None: - shared.state.current_image = image - image.save(last_saved_image) - last_saved_image += f", prompt: {preview_text}" - - hypernetwork.train() + optimizer.zero_grad() + shared.sd_model.cond_stage_model.to(devices.device) + shared.sd_model.first_stage_model.to(devices.device) + + p = processing.StableDiffusionProcessingTxt2Img( + sd_model=shared.sd_model, + do_not_save_grid=True, + do_not_save_samples=True, + ) + + if preview_from_txt2img: + p.prompt = preview_prompt + p.negative_prompt = preview_negative_prompt + p.steps = preview_steps + p.sampler_index = preview_sampler_index + p.cfg_scale = preview_cfg_scale + p.seed = preview_seed + p.width = preview_width + p.height = preview_height + else: + p.prompt = entries[0].cond_text + p.steps = 20 + + preview_text = p.prompt + + processed = processing.process_images(p) + image = processed.images[0] if len(processed.images)>0 else None + + if unload: + shared.sd_model.cond_stage_model.to(devices.cpu) + shared.sd_model.first_stage_model.to(devices.cpu) + + if image is not None: + shared.state.current_image = image + image.save(last_saved_image) + last_saved_image += f", prompt: {preview_text}" shared.state.job_no = hypernetwork.step