Merge pull request #3975 from aria1th/force-push-patch-13

Save/loading AdamW optimizer (for hypernetworks)

Author: AUTOMATIC1111

modules/hypernetworks/hypernetwork.py

@@ -22,6 +22,8 @@
 from statistics import stdev, mean
 
 
+optimizer_dict = {optim_name : cls_obj for optim_name, cls_obj in inspect.getmembers(torch.optim, inspect.isclass) if optim_name != "Optimizer"}
+
 class HypernetworkModule(torch.nn.Module):
     multiplier = 1.0
     activation_dict = {
@@ -142,6 +144,8 @@ def __init__(self, name=None, enable_sizes=None, layer_structure=None, activatio
         self.use_dropout = use_dropout
         self.activate_output = activate_output
         self.last_layer_dropout = kwargs['last_layer_dropout'] if 'last_layer_dropout' in kwargs else True
+        self.optimizer_name = None
+        self.optimizer_state_dict = None
 
         for size in enable_sizes or []:
             self.layers[size] = (
@@ -163,6 +167,7 @@ def weights(self):
 
     def save(self, filename):
         state_dict = {}
+        optimizer_saved_dict = {}
 
         for k, v in self.layers.items():
             state_dict[k] = (v[0].state_dict(), v[1].state_dict())
@@ -178,8 +183,15 @@ def save(self, filename):
         state_dict['sd_checkpoint_name'] = self.sd_checkpoint_name
         state_dict['activate_output'] = self.activate_output
         state_dict['last_layer_dropout'] = self.last_layer_dropout
-        
+
+        if self.optimizer_name is not None:
+            optimizer_saved_dict['optimizer_name'] = self.optimizer_name
+
         torch.save(state_dict, filename)
+        if shared.opts.save_optimizer_state and self.optimizer_state_dict:
+            optimizer_saved_dict['hash'] = sd_models.model_hash(filename)
+            optimizer_saved_dict['optimizer_state_dict'] = self.optimizer_state_dict
+            torch.save(optimizer_saved_dict, filename + '.optim')
 
     def load(self, filename):
         self.filename = filename
@@ -202,6 +214,18 @@ def load(self, filename):
         print(f"Activate last layer is set to {self.activate_output}")
         self.last_layer_dropout = state_dict.get('last_layer_dropout', False)
 
+        optimizer_saved_dict = torch.load(self.filename + '.optim', map_location = 'cpu') if os.path.exists(self.filename + '.optim') else {}
+        self.optimizer_name = optimizer_saved_dict.get('optimizer_name', 'AdamW')
+        print(f"Optimizer name is {self.optimizer_name}")
+        if sd_models.model_hash(filename) == optimizer_saved_dict.get('hash', None):
+            self.optimizer_state_dict = optimizer_saved_dict.get('optimizer_state_dict', None)
+        else:
+            self.optimizer_state_dict = None
+        if self.optimizer_state_dict:
+            print("Loaded existing optimizer from checkpoint")
+        else:
+            print("No saved optimizer exists in checkpoint")
+
         for size, sd in state_dict.items():
             if type(size) == int:
                 self.layers[size] = (
@@ -223,7 +247,7 @@ def list_hypernetworks(path):
         name = os.path.splitext(os.path.basename(filename))[0]
         # Prevent a hypothetical "None.pt" from being listed.
         if name != "None":
-            res[name] = filename
+            res[name + f"({sd_models.model_hash(filename)})"] = filename
     return res
 
 
@@ -358,6 +382,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log
     shared.state.textinfo = "Initializing hypernetwork training..."
     shared.state.job_count = steps
 
+    hypernetwork_name = hypernetwork_name.rsplit('(', 1)[0]
     filename = os.path.join(shared.cmd_opts.hypernetwork_dir, f'{hypernetwork_name}.pt')
 
     log_directory = os.path.join(log_directory, datetime.datetime.now().strftime("%Y-%m-%d"), hypernetwork_name)
@@ -404,8 +429,19 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log
     weights = hypernetwork.weights()
     for weight in weights:
         weight.requires_grad = True
-    # if optimizer == "AdamW": or else Adam / AdamW / SGD, etc...
-    optimizer = torch.optim.AdamW(weights, lr=scheduler.learn_rate)
+    # Here we use optimizer from saved HN, or we can specify as UI option.
+    if (optimizer_name := hypernetwork.optimizer_name) in optimizer_dict:
+        optimizer = optimizer_dict[hypernetwork.optimizer_name](params=weights, lr=scheduler.learn_rate)
+    else:
+        print(f"Optimizer type {optimizer_name} is not defined!")
+        optimizer = torch.optim.AdamW(params=weights, lr=scheduler.learn_rate)
+        optimizer_name = 'AdamW'
+    if hypernetwork.optimizer_state_dict:  # This line must be changed if Optimizer type can be different from saved optimizer.
+        try:
+            optimizer.load_state_dict(hypernetwork.optimizer_state_dict)
+        except RuntimeError as e:
+            print("Cannot resume from saved optimizer!")
+            print(e)
 
     steps_without_grad = 0
 
@@ -467,7 +503,11 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log
             # Before saving, change name to match current checkpoint.
             hypernetwork_name_every = f'{hypernetwork_name}-{steps_done}'
             last_saved_file = os.path.join(hypernetwork_dir, f'{hypernetwork_name_every}.pt')
+            hypernetwork.optimizer_name = optimizer_name
+            if shared.opts.save_optimizer_state:
+                hypernetwork.optimizer_state_dict = optimizer.state_dict()
             save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, last_saved_file)
+            hypernetwork.optimizer_state_dict = None  # dereference it after saving, to save memory.
 
         textual_inversion.write_loss(log_directory, "hypernetwork_loss.csv", hypernetwork.step, len(ds), {
             "loss": f"{previous_mean_loss:.7f}",
@@ -530,8 +570,12 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log
     report_statistics(loss_dict)
 
     filename = os.path.join(shared.cmd_opts.hypernetwork_dir, f'{hypernetwork_name}.pt')
+    hypernetwork.optimizer_name = optimizer_name
+    if shared.opts.save_optimizer_state:
+        hypernetwork.optimizer_state_dict = optimizer.state_dict()
     save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, filename)
-
+    del optimizer
+    hypernetwork.optimizer_state_dict = None  # dereference it after saving, to save memory.
     return hypernetwork, filename
 
 def save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, filename):

modules/hypernetworks/ui.py

@@ -9,7 +9,7 @@
 from modules.hypernetworks import hypernetwork
 
 not_available = ["hardswish", "multiheadattention"]
-keys = ["linear"] + list(x for x in hypernetwork.HypernetworkModule.activation_dict.keys() if x not in not_available)
+keys = list(x for x in hypernetwork.HypernetworkModule.activation_dict.keys() if x not in not_available)
 
 def create_hypernetwork(name, enable_sizes, overwrite_old, layer_structure=None, activation_func=None, weight_init=None, add_layer_norm=False, use_dropout=False):
     # Remove illegal characters from name.

modules/shared.py

@@ -320,6 +320,7 @@ def options_section(section_identifier, options_dict):
 
 options_templates.update(options_section(('training', "Training"), {
     "unload_models_when_training": OptionInfo(False, "Move VAE and CLIP to RAM when training if possible. Saves VRAM."),
+    "save_optimizer_state": OptionInfo(False, "Saves Optimizer state as separate *.optim file. Training can be resumed with HN itself and matching optim file."),
     "dataset_filename_word_regex": OptionInfo("", "Filename word regex"),
     "dataset_filename_join_string": OptionInfo(" ", "Filename join string"),
     "training_image_repeats_per_epoch": OptionInfo(1, "Number of repeats for a single input image per epoch; used only for displaying epoch number", gr.Number, {"precision": 0}),