From 166be3919b817cee5e702fd01c34afe9081b952c Mon Sep 17 00:00:00 2001
From: DepFA <35278260+dfaker@users.noreply.github.com>
Date: Thu, 20 Oct 2022 00:09:40 +0100
Subject: allow overwrite old hn

---
 modules/hypernetworks/ui.py | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

(limited to 'modules/hypernetworks/ui.py')

diff --git a/modules/hypernetworks/ui.py b/modules/hypernetworks/ui.py
index 08f75f15..f45345ea 100644
--- a/modules/hypernetworks/ui.py
+++ b/modules/hypernetworks/ui.py
@@ -10,9 +10,10 @@ from modules import sd_hijack, shared, devices
 from modules.hypernetworks import hypernetwork
 
 
-def create_hypernetwork(name, enable_sizes, layer_structure=None, add_layer_norm=False):
+def create_hypernetwork(name, enable_sizes, overwrite_old, layer_structure=None, add_layer_norm=False):
     fn = os.path.join(shared.cmd_opts.hypernetwork_dir, f"{name}.pt")
-    assert not os.path.exists(fn), f"file {fn} already exists"
+    if not overwrite_old:
+        assert not os.path.exists(fn), f"file {fn} already exists"
 
     if type(layer_structure) == str:
         layer_structure = tuple(map(int, re.sub(r'\D', '', layer_structure)))
-- 
cgit v1.2.3


From 6f98e89486f55b0e4657e96ce640cf1c4675d187 Mon Sep 17 00:00:00 2001
From: discus0434 <discus0434@gmail.com>
Date: Thu, 20 Oct 2022 00:10:45 +0000
Subject: update

---
 modules/hypernetworks/hypernetwork.py | 29 +++++++++++++++--------
 modules/hypernetworks/ui.py           |  3 ++-
 modules/ui.py                         | 43 +++++++++++++++++++----------------
 3 files changed, 44 insertions(+), 31 deletions(-)

(limited to 'modules/hypernetworks/ui.py')

diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py
index 74300122..7d617680 100644
--- a/modules/hypernetworks/hypernetwork.py
+++ b/modules/hypernetworks/hypernetwork.py
@@ -22,16 +22,20 @@ from modules.textual_inversion.learn_schedule import LearnRateScheduler
 class HypernetworkModule(torch.nn.Module):
     multiplier = 1.0
 
-    def __init__(self, dim, state_dict=None, layer_structure=None, add_layer_norm=False):
+    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 mut not be None"
+        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 add_layer_norm:
                 linears.append(torch.nn.LayerNorm(int(dim * layer_structure[i+1])))
 
@@ -42,8 +46,9 @@ class HypernetworkModule(torch.nn.Module):
             self.load_state_dict(state_dict)
         else:
             for layer in self.linear:
-                layer.weight.data.normal_(mean=0.0, std=0.01)
-                layer.bias.data.zero_()
+                if not "ReLU" in layer.__str__():
+                    layer.weight.data.normal_(mean=0.0, std=0.01)
+                    layer.bias.data.zero_()
 
         self.to(devices.device)
 
@@ -69,7 +74,8 @@ class HypernetworkModule(torch.nn.Module):
     def trainables(self):
         layer_structure = []
         for layer in self.linear:
-            layer_structure += [layer.weight, layer.bias]
+            if not "ReLU" in layer.__str__():
+                layer_structure += [layer.weight, layer.bias]
         return layer_structure
 
 
@@ -81,7 +87,7 @@ class Hypernetwork:
     filename = None
     name = None
 
-    def __init__(self, name=None, enable_sizes=None, layer_structure=None, add_layer_norm=False):
+    def __init__(self, name=None, enable_sizes=None, layer_structure=None, add_layer_norm=False, activation_func=None):
         self.filename = None
         self.name = name
         self.layers = {}
@@ -90,11 +96,12 @@ class Hypernetwork:
         self.sd_checkpoint_name = None
         self.layer_structure = layer_structure
         self.add_layer_norm = add_layer_norm
+        self.activation_func = activation_func
 
         for size in enable_sizes or []:
             self.layers[size] = (
-                HypernetworkModule(size, None, self.layer_structure, self.add_layer_norm),
-                HypernetworkModule(size, None, self.layer_structure, self.add_layer_norm),
+                HypernetworkModule(size, None, self.layer_structure, self.add_layer_norm, self.activation_func),
+                HypernetworkModule(size, None, self.layer_structure, self.add_layer_norm, self.activation_func),
             )
 
     def weights(self):
@@ -117,6 +124,7 @@ class Hypernetwork:
         state_dict['name'] = self.name
         state_dict['layer_structure'] = self.layer_structure
         state_dict['is_layer_norm'] = self.add_layer_norm
+        state_dict['activation_func'] = self.activation_func
         state_dict['sd_checkpoint'] = self.sd_checkpoint
         state_dict['sd_checkpoint_name'] = self.sd_checkpoint_name
 
@@ -131,12 +139,13 @@ class Hypernetwork:
 
         self.layer_structure = state_dict.get('layer_structure', [1, 2, 1])
         self.add_layer_norm = state_dict.get('is_layer_norm', False)
+        self.activation_func = state_dict.get('activation_func', None)
 
         for size, sd in state_dict.items():
             if type(size) == int:
                 self.layers[size] = (
-                    HypernetworkModule(size, sd[0], self.layer_structure, self.add_layer_norm),
-                    HypernetworkModule(size, sd[1], self.layer_structure, self.add_layer_norm),
+                    HypernetworkModule(size, sd[0], self.layer_structure, self.add_layer_norm, self.activation_func),
+                    HypernetworkModule(size, sd[1], self.layer_structure, self.add_layer_norm, self.activation_func),
                 )
 
         self.name = state_dict.get('name', self.name)
diff --git a/modules/hypernetworks/ui.py b/modules/hypernetworks/ui.py
index 08f75f15..83f9547b 100644
--- a/modules/hypernetworks/ui.py
+++ b/modules/hypernetworks/ui.py
@@ -10,7 +10,7 @@ from modules import sd_hijack, shared, devices
 from modules.hypernetworks import hypernetwork
 
 
-def create_hypernetwork(name, enable_sizes, layer_structure=None, add_layer_norm=False):
+def create_hypernetwork(name, enable_sizes, layer_structure=None, add_layer_norm=False, activation_func=None):
     fn = os.path.join(shared.cmd_opts.hypernetwork_dir, f"{name}.pt")
     assert not os.path.exists(fn), f"file {fn} already exists"
 
@@ -22,6 +22,7 @@ def create_hypernetwork(name, enable_sizes, layer_structure=None, add_layer_norm
         enable_sizes=[int(x) for x in enable_sizes],
         layer_structure=layer_structure,
         add_layer_norm=add_layer_norm,
+        activation_func=activation_func,
     )
     hypernet.save(fn)
 
diff --git a/modules/ui.py b/modules/ui.py
index d2e24880..8751fa9c 100644
--- a/modules/ui.py
+++ b/modules/ui.py
@@ -5,43 +5,44 @@ import json
 import math
 import mimetypes
 import os
+import platform
 import random
+import subprocess as sp
 import sys
 import tempfile
 import time
 import traceback
-import platform
-import subprocess as sp
 from functools import partial, reduce
 
+import gradio as gr
+import gradio.routes
+import gradio.utils
 import numpy as np
+import piexif
 import torch
 from PIL import Image, PngImagePlugin
-import piexif
 
-import gradio as gr
-import gradio.utils
-import gradio.routes
-
-from modules import sd_hijack, sd_models, localization
+from modules import localization, sd_hijack, sd_models
 from modules.paths import script_path
-from modules.shared import opts, cmd_opts, restricted_opts
+from modules.shared import cmd_opts, opts, restricted_opts
+
 if cmd_opts.deepdanbooru:
     from modules.deepbooru import get_deepbooru_tags
-import modules.shared as shared
-from modules.sd_samplers import samplers, samplers_for_img2img
-from modules.sd_hijack import model_hijack
+
+import modules.codeformer_model
+import modules.generation_parameters_copypaste
+import modules.gfpgan_model
+import modules.hypernetworks.ui
+import modules.images_history as img_his
 import modules.ldsr_model
 import modules.scripts
-import modules.gfpgan_model
-import modules.codeformer_model
+import modules.shared as shared
 import modules.styles
-import modules.generation_parameters_copypaste
+import modules.textual_inversion.ui
 from modules import prompt_parser
 from modules.images import save_image
-import modules.textual_inversion.ui
-import modules.hypernetworks.ui
-import modules.images_history as img_his
+from modules.sd_hijack import model_hijack
+from modules.sd_samplers import samplers, samplers_for_img2img
 
 # this is a fix for Windows users. Without it, javascript files will be served with text/html content-type and the browser will not show any UI
 mimetypes.init()
@@ -268,8 +269,8 @@ def calc_time_left(progress, threshold, label, force_display):
         time_since_start = time.time() - shared.state.time_start
         eta = (time_since_start/progress)
         eta_relative = eta-time_since_start
-        if (eta_relative > threshold and progress > 0.02) or force_display:           
-            return label + time.strftime('%H:%M:%S', time.gmtime(eta_relative))        
+        if (eta_relative > threshold and progress > 0.02) or force_display:
+            return label + time.strftime('%H:%M:%S', time.gmtime(eta_relative))
         else:
             return ""
 
@@ -1219,6 +1220,7 @@ def create_ui(wrap_gradio_gpu_call):
                     new_hypernetwork_sizes = gr.CheckboxGroup(label="Modules", value=["768", "320", "640", "1280"], choices=["768", "320", "640", "1280"])
                     new_hypernetwork_layer_structure = gr.Textbox("1, 2, 1", label="Enter hypernetwork layer structure", placeholder="1st and last digit must be 1. ex:'1, 2, 1'")
                     new_hypernetwork_add_layer_norm = gr.Checkbox(label="Add layer normalization")
+                    new_hypernetwork_activation_func = gr.Dropdown(value="relu", label="Select activation function of hypernetwork", choices=["relu", "leakyrelu"])
 
                     with gr.Row():
                         with gr.Column(scale=3):
@@ -1303,6 +1305,7 @@ def create_ui(wrap_gradio_gpu_call):
                 new_hypernetwork_sizes,
                 new_hypernetwork_layer_structure,
                 new_hypernetwork_add_layer_norm,
+                new_hypernetwork_activation_func,
             ],
             outputs=[
                 train_hypernetwork_name,
-- 
cgit v1.2.3


From 51e3dc9ccad157d7161b697a246e26c868d46a7c Mon Sep 17 00:00:00 2001
From: timntorres <timothynarcisotorres@gmail.com>
Date: Fri, 21 Oct 2022 02:11:12 -0700
Subject: Sanitize hypernet name input.

---
 modules/hypernetworks/ui.py | 3 +++
 1 file changed, 3 insertions(+)

(limited to 'modules/hypernetworks/ui.py')

diff --git a/modules/hypernetworks/ui.py b/modules/hypernetworks/ui.py
index 266f04f6..e6f50a1f 100644
--- a/modules/hypernetworks/ui.py
+++ b/modules/hypernetworks/ui.py
@@ -11,6 +11,9 @@ from modules.hypernetworks import hypernetwork
 
 
 def create_hypernetwork(name, enable_sizes, overwrite_old, layer_structure=None, add_layer_norm=False, activation_func=None):
+    # Remove illegal characters from name.
+    name = "".join( x for x in name if (x.isalnum() or x in "._- "))
+
     fn = os.path.join(shared.cmd_opts.hypernetwork_dir, f"{name}.pt")
     if not overwrite_old:
         assert not os.path.exists(fn), f"file {fn} already exists"
-- 
cgit v1.2.3


From 0e8ca8e7af05be22d7d2c07a47c3c7febe0f0ab6 Mon Sep 17 00:00:00 2001
From: discus0434 <discus0434@gmail.com>
Date: Sat, 22 Oct 2022 11:07:00 +0000
Subject: add dropout

---
 modules/hypernetworks/hypernetwork.py | 68 +++++++++++++++++++++--------------
 modules/hypernetworks/ui.py           | 10 +++---
 modules/ui.py                         | 43 +++++++++++-----------
 3 files changed, 70 insertions(+), 51 deletions(-)

(limited to 'modules/hypernetworks/ui.py')

diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py
index 905cbeef..e493f366 100644
--- a/modules/hypernetworks/hypernetwork.py
+++ b/modules/hypernetworks/hypernetwork.py
@@ -1,47 +1,60 @@
+import csv
 import datetime
 import glob
 import html
 import os
 import sys
 import traceback
-import tqdm
-import csv
 
+import modules.textual_inversion.dataset
 import torch
-
-from ldm.util import default
-from modules import devices, shared, processing, sd_models
-import torch
-from torch import einsum
+import tqdm
 from einops import rearrange, repeat
-import modules.textual_inversion.dataset
+from ldm.util import default
+from modules import devices, processing, sd_models, shared
 from modules.textual_inversion import textual_inversion
 from modules.textual_inversion.learn_schedule import LearnRateScheduler
+from torch import einsum
 
 
 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):
+    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, activation_func=None, add_layer_norm=False, use_dropout=False):
         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!"
-        
+        assert activation_func not in self.activation_dict.keys() + "linear", f"Valid activation funcs: 'linear', 'relu', 'leakyrelu', 'elu', 'swish'"
+
         linears = []
         for i in range(len(layer_structure) - 1):
+
+            # Add a fully-connected layer
             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 activation_func in HypernetworkModule.activation_dict:
-                linears.append(HypernetworkModule.activation_dict[activation_func]())
+
+            # Add an activation func
+            if activation_func == "linear":
+                pass
+            elif activation_func in self.activation_dict:
+                linears.append(self.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))
+                raise NotImplementedError(
+                    "Valid activation funcs: 'linear', 'relu', 'leakyrelu', 'elu', 'swish'"
+                )
+
+            # Add dropout
+            if use_dropout:
+                linears.append(torch.nn.Dropout(p=0.3))
+
+            # Add layer normalization
             if add_layer_norm:
                 linears.append(torch.nn.LayerNorm(int(dim * layer_structure[i+1])))
 
@@ -93,7 +106,7 @@ class Hypernetwork:
     filename = None
     name = None
 
-    def __init__(self, name=None, enable_sizes=None, layer_structure=None, add_layer_norm=False, activation_func=None):
+    def __init__(self, name=None, enable_sizes=None, layer_structure=None, activation_func=None, add_layer_norm=False, use_dropout=False):
         self.filename = None
         self.name = name
         self.layers = {}
@@ -101,13 +114,14 @@ class Hypernetwork:
         self.sd_checkpoint = None
         self.sd_checkpoint_name = None
         self.layer_structure = layer_structure
-        self.add_layer_norm = add_layer_norm
         self.activation_func = activation_func
+        self.add_layer_norm = add_layer_norm
+        self.use_dropout = use_dropout
 
         for size in enable_sizes or []:
             self.layers[size] = (
-                HypernetworkModule(size, None, self.layer_structure, self.add_layer_norm, self.activation_func),
-                HypernetworkModule(size, None, self.layer_structure, self.add_layer_norm, self.activation_func),
+                HypernetworkModule(size, None, self.layer_structure, self.activation_func, self.add_layer_norm, self.use_dropout),
+                HypernetworkModule(size, None, self.layer_structure, self.activation_func, self.add_layer_norm, self.use_dropout),
             )
 
     def weights(self):
@@ -129,8 +143,9 @@ class Hypernetwork:
         state_dict['step'] = self.step
         state_dict['name'] = self.name
         state_dict['layer_structure'] = self.layer_structure
-        state_dict['is_layer_norm'] = self.add_layer_norm
         state_dict['activation_func'] = self.activation_func
+        state_dict['is_layer_norm'] = self.add_layer_norm
+        state_dict['use_dropout'] = self.use_dropout
         state_dict['sd_checkpoint'] = self.sd_checkpoint
         state_dict['sd_checkpoint_name'] = self.sd_checkpoint_name
 
@@ -144,8 +159,9 @@ class Hypernetwork:
         state_dict = torch.load(filename, map_location='cpu')
 
         self.layer_structure = state_dict.get('layer_structure', [1, 2, 1])
-        self.add_layer_norm = state_dict.get('is_layer_norm', False)
         self.activation_func = state_dict.get('activation_func', None)
+        self.add_layer_norm = state_dict.get('is_layer_norm', False)
+        self.use_dropout = state_dict.get('use_dropout', False)
 
         for size, sd in state_dict.items():
             if type(size) == int:
diff --git a/modules/hypernetworks/ui.py b/modules/hypernetworks/ui.py
index 1a5a27d8..5f6f17b6 100644
--- a/modules/hypernetworks/ui.py
+++ b/modules/hypernetworks/ui.py
@@ -3,14 +3,13 @@ import os
 import re
 
 import gradio as gr
-
-import modules.textual_inversion.textual_inversion
 import modules.textual_inversion.preprocess
-from modules import sd_hijack, shared, devices
+import modules.textual_inversion.textual_inversion
+from modules import devices, sd_hijack, shared
 from modules.hypernetworks import hypernetwork
 
 
-def create_hypernetwork(name, enable_sizes, layer_structure=None, add_layer_norm=False, activation_func=None):
+def create_hypernetwork(name, enable_sizes, layer_structure=None, activation_func=None, add_layer_norm=False, use_dropout=False):
     fn = os.path.join(shared.cmd_opts.hypernetwork_dir, f"{name}.pt")
     assert not os.path.exists(fn), f"file {fn} already exists"
 
@@ -21,8 +20,9 @@ def create_hypernetwork(name, enable_sizes, layer_structure=None, add_layer_norm
         name=name,
         enable_sizes=[int(x) for x in enable_sizes],
         layer_structure=layer_structure,
-        add_layer_norm=add_layer_norm,
         activation_func=activation_func,
+        add_layer_norm=add_layer_norm,
+        use_dropout=use_dropout,
     )
     hypernet.save(fn)
 
diff --git a/modules/ui.py b/modules/ui.py
index 716f14b8..d4b32c05 100644
--- a/modules/ui.py
+++ b/modules/ui.py
@@ -5,43 +5,44 @@ import json
 import math
 import mimetypes
 import os
+import platform
 import random
+import subprocess as sp
 import sys
 import tempfile
 import time
 import traceback
-import platform
-import subprocess as sp
 from functools import partial, reduce
 
+import gradio as gr
+import gradio.routes
+import gradio.utils
 import numpy as np
+import piexif
 import torch
 from PIL import Image, PngImagePlugin
-import piexif
 
-import gradio as gr
-import gradio.utils
-import gradio.routes
-
-from modules import sd_hijack, sd_models, localization
+from modules import localization, sd_hijack, sd_models
 from modules.paths import script_path
-from modules.shared import opts, cmd_opts, restricted_opts
+from modules.shared import cmd_opts, opts, restricted_opts
+
 if cmd_opts.deepdanbooru:
     from modules.deepbooru import get_deepbooru_tags
-import modules.shared as shared
-from modules.sd_samplers import samplers, samplers_for_img2img
-from modules.sd_hijack import model_hijack
+
+import modules.codeformer_model
+import modules.generation_parameters_copypaste
+import modules.gfpgan_model
+import modules.hypernetworks.ui
+import modules.images_history as img_his
 import modules.ldsr_model
 import modules.scripts
-import modules.gfpgan_model
-import modules.codeformer_model
+import modules.shared as shared
 import modules.styles
-import modules.generation_parameters_copypaste
+import modules.textual_inversion.ui
 from modules import prompt_parser
 from modules.images import save_image
-import modules.textual_inversion.ui
-import modules.hypernetworks.ui
-import modules.images_history as img_his
+from modules.sd_hijack import model_hijack
+from modules.sd_samplers import samplers, samplers_for_img2img
 
 # this is a fix for Windows users. Without it, javascript files will be served with text/html content-type and the browser will not show any UI
 mimetypes.init()
@@ -1223,8 +1224,9 @@ def create_ui(wrap_gradio_gpu_call):
                     new_hypernetwork_name = gr.Textbox(label="Name")
                     new_hypernetwork_sizes = gr.CheckboxGroup(label="Modules", value=["768", "320", "640", "1280"], choices=["768", "320", "640", "1280"])
                     new_hypernetwork_layer_structure = gr.Textbox("1, 2, 1", label="Enter hypernetwork layer structure", placeholder="1st and last digit must be 1. ex:'1, 2, 1'")
+                    new_hypernetwork_activation_func = gr.Dropdown(value="relu", label="Select activation function of hypernetwork", choices=["linear", "relu", "leakyrelu", "elu", "swish"])
                     new_hypernetwork_add_layer_norm = gr.Checkbox(label="Add layer normalization")
-                    new_hypernetwork_activation_func = gr.Dropdown(value="relu", label="Select activation function of hypernetwork", choices=["linear", "relu", "leakyrelu"])
+                    new_hypernetwork_use_dropout = gr.Checkbox(label="Use dropout")
 
                     with gr.Row():
                         with gr.Column(scale=3):
@@ -1308,8 +1310,9 @@ def create_ui(wrap_gradio_gpu_call):
                 new_hypernetwork_name,
                 new_hypernetwork_sizes,
                 new_hypernetwork_layer_structure,
-                new_hypernetwork_add_layer_norm,
                 new_hypernetwork_activation_func,
+                new_hypernetwork_add_layer_norm,
+                new_hypernetwork_use_dropout
             ],
             outputs=[
                 train_hypernetwork_name,
-- 
cgit v1.2.3


From de096d0ce752c96e45508dcc7b9e84f7dbe10cca Mon Sep 17 00:00:00 2001
From: AngelBottomless <35677394+aria1th@users.noreply.github.com>
Date: Tue, 25 Oct 2022 14:48:49 +0900
Subject: Weight initialization and More activation func

add weight init

add weight init option in create_hypernetwork

fstringify hypernet info

save weight initialization info for further debugging

fill bias with zero for He/Xavier

initialize LayerNorm with Normal

fix loading weight_init
---
 modules/hypernetworks/hypernetwork.py | 47 ++++++++++++++++++++++++++++-------
 modules/hypernetworks/ui.py           |  4 ++-
 modules/ui.py                         |  4 ++-
 3 files changed, 44 insertions(+), 11 deletions(-)

(limited to 'modules/hypernetworks/ui.py')

diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py
index d647ea55..afbcdff8 100644
--- a/modules/hypernetworks/hypernetwork.py
+++ b/modules/hypernetworks/hypernetwork.py
@@ -5,6 +5,7 @@ import html
 import os
 import sys
 import traceback
+import inspect
 
 import modules.textual_inversion.dataset
 import torch
@@ -15,10 +16,12 @@ from modules import devices, processing, sd_models, shared
 from modules.textual_inversion import textual_inversion
 from modules.textual_inversion.learn_schedule import LearnRateScheduler
 from torch import einsum
+from torch.nn.init import normal_, xavier_normal_, xavier_uniform_, kaiming_normal_, kaiming_uniform_, zeros_
 
 from collections import defaultdict, deque
 from statistics import stdev, mean
 
+
 class HypernetworkModule(torch.nn.Module):
     multiplier = 1.0
     activation_dict = {
@@ -26,9 +29,12 @@ class HypernetworkModule(torch.nn.Module):
         "leakyrelu": torch.nn.LeakyReLU,
         "elu": torch.nn.ELU,
         "swish": torch.nn.Hardswish,
+        "tanh": torch.nn.Tanh,
+        "sigmoid": torch.nn.Sigmoid,
     }
+    activation_dict.update({cls_name: cls_obj for cls_name, cls_obj in inspect.getmembers(torch.nn.modules.activation) if inspect.isclass(cls_obj) and cls_obj.__module__ == 'torch.nn.modules.activation'})
 
-    def __init__(self, dim, state_dict=None, layer_structure=None, activation_func=None, add_layer_norm=False, use_dropout=False):
+    def __init__(self, dim, state_dict=None, layer_structure=None, activation_func=None, weight_init='Normal', add_layer_norm=False, use_dropout=False):
         super().__init__()
 
         assert layer_structure is not None, "layer_structure must not be None"
@@ -65,9 +71,24 @@ class HypernetworkModule(torch.nn.Module):
         else:
             for layer in self.linear:
                 if type(layer) == torch.nn.Linear or type(layer) == torch.nn.LayerNorm:
-                    layer.weight.data.normal_(mean=0.0, std=0.01)
-                    layer.bias.data.zero_()
-
+                    w, b = layer.weight.data, layer.bias.data
+                    if weight_init == "Normal" or type(layer) == torch.nn.LayerNorm:
+                        normal_(w, mean=0.0, std=0.01)
+                        normal_(b, mean=0.0, std=0.005)
+                    elif weight_init == 'XavierUniform':
+                        xavier_uniform_(w)
+                        zeros_(b)
+                    elif weight_init == 'XavierNormal':
+                        xavier_normal_(w)
+                        zeros_(b)
+                    elif weight_init == 'KaimingUniform':
+                        kaiming_uniform_(w, nonlinearity='leaky_relu' if 'leakyrelu' == activation_func else 'relu')
+                        zeros_(b)
+                    elif weight_init == 'KaimingNormal':
+                        kaiming_normal_(w, nonlinearity='leaky_relu' if 'leakyrelu' == activation_func else 'relu')
+                        zeros_(b)
+                    else:
+                        raise KeyError(f"Key {weight_init} is not defined as initialization!")
         self.to(devices.device)
 
     def fix_old_state_dict(self, state_dict):
@@ -105,7 +126,7 @@ class Hypernetwork:
     filename = None
     name = None
 
-    def __init__(self, name=None, enable_sizes=None, layer_structure=None, activation_func=None, add_layer_norm=False, use_dropout=False):
+    def __init__(self, name=None, enable_sizes=None, layer_structure=None, activation_func=None, weight_init=None, add_layer_norm=False, use_dropout=False):
         self.filename = None
         self.name = name
         self.layers = {}
@@ -114,13 +135,14 @@ class Hypernetwork:
         self.sd_checkpoint_name = None
         self.layer_structure = layer_structure
         self.activation_func = activation_func
+        self.weight_init = weight_init
         self.add_layer_norm = add_layer_norm
         self.use_dropout = use_dropout
 
         for size in enable_sizes or []:
             self.layers[size] = (
-                HypernetworkModule(size, None, self.layer_structure, self.activation_func, self.add_layer_norm, self.use_dropout),
-                HypernetworkModule(size, None, self.layer_structure, self.activation_func, self.add_layer_norm, self.use_dropout),
+                HypernetworkModule(size, None, self.layer_structure, self.activation_func, self.weight_init, self.add_layer_norm, self.use_dropout),
+                HypernetworkModule(size, None, self.layer_structure, self.activation_func, self.weight_init, self.add_layer_norm, self.use_dropout),
             )
 
     def weights(self):
@@ -144,6 +166,7 @@ class Hypernetwork:
         state_dict['layer_structure'] = self.layer_structure
         state_dict['activation_func'] = self.activation_func
         state_dict['is_layer_norm'] = self.add_layer_norm
+        state_dict['weight_initialization'] = self.weight_init
         state_dict['use_dropout'] = self.use_dropout
         state_dict['sd_checkpoint'] = self.sd_checkpoint
         state_dict['sd_checkpoint_name'] = self.sd_checkpoint_name
@@ -158,15 +181,21 @@ class Hypernetwork:
         state_dict = torch.load(filename, map_location='cpu')
 
         self.layer_structure = state_dict.get('layer_structure', [1, 2, 1])
+        print(self.layer_structure)
         self.activation_func = state_dict.get('activation_func', None)
+        print(f"Activation function is {self.activation_func}")
+        self.weight_init = state_dict.get('weight_initialization', 'Normal')
+        print(f"Weight initialization is {self.weight_init}")
         self.add_layer_norm = state_dict.get('is_layer_norm', False)
+        print(f"Layer norm is set to {self.add_layer_norm}")
         self.use_dropout = state_dict.get('use_dropout', False)
+        print(f"Dropout usage is set to {self.use_dropout}" )
 
         for size, sd in state_dict.items():
             if type(size) == int:
                 self.layers[size] = (
-                    HypernetworkModule(size, sd[0], self.layer_structure, self.activation_func, self.add_layer_norm, self.use_dropout),
-                    HypernetworkModule(size, sd[1], self.layer_structure, self.activation_func, self.add_layer_norm, self.use_dropout),
+                    HypernetworkModule(size, sd[0], self.layer_structure, self.activation_func, self.weight_init, self.add_layer_norm, self.use_dropout),
+                    HypernetworkModule(size, sd[1], self.layer_structure, self.activation_func, self.weight_init, self.add_layer_norm, self.use_dropout),
                 )
 
         self.name = state_dict.get('name', self.name)
diff --git a/modules/hypernetworks/ui.py b/modules/hypernetworks/ui.py
index 2b472d87..2c6c0470 100644
--- a/modules/hypernetworks/ui.py
+++ b/modules/hypernetworks/ui.py
@@ -8,8 +8,9 @@ import modules.textual_inversion.textual_inversion
 from modules import devices, sd_hijack, shared
 from modules.hypernetworks import hypernetwork
 
+keys = list(hypernetwork.HypernetworkModule.activation_dict.keys())
 
-def create_hypernetwork(name, enable_sizes, overwrite_old, layer_structure=None, activation_func=None, add_layer_norm=False, use_dropout=False):
+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.
     name = "".join( x for x in name if (x.isalnum() or x in "._- "))
 
@@ -25,6 +26,7 @@ def create_hypernetwork(name, enable_sizes, overwrite_old, layer_structure=None,
         enable_sizes=[int(x) for x in enable_sizes],
         layer_structure=layer_structure,
         activation_func=activation_func,
+        weight_init=weight_init,
         add_layer_norm=add_layer_norm,
         use_dropout=use_dropout,
     )
diff --git a/modules/ui.py b/modules/ui.py
index 03528968..8e343258 100644
--- a/modules/ui.py
+++ b/modules/ui.py
@@ -1238,7 +1238,8 @@ def create_ui(wrap_gradio_gpu_call):
                     new_hypernetwork_name = gr.Textbox(label="Name")
                     new_hypernetwork_sizes = gr.CheckboxGroup(label="Modules", value=["768", "320", "640", "1280"], choices=["768", "320", "640", "1280"])
                     new_hypernetwork_layer_structure = gr.Textbox("1, 2, 1", label="Enter hypernetwork layer structure", placeholder="1st and last digit must be 1. ex:'1, 2, 1'")
-                    new_hypernetwork_activation_func = gr.Dropdown(value="relu", label="Select activation function of hypernetwork", choices=["linear", "relu", "leakyrelu", "elu", "swish"])
+                    new_hypernetwork_activation_func = gr.Dropdown(value="relu", label="Select activation function of hypernetwork", choices=modules.hypernetworks.ui.keys)
+                    new_hypernetwork_initialization_option = gr.Dropdown(value = "Normal", label="Select Layer weights initialization. relu-like - Kaiming, sigmoid-like - Xavier is recommended", choices=["Normal", "KaimingUniform", "KaimingNormal", "XavierUniform", "XavierNormal"])
                     new_hypernetwork_add_layer_norm = gr.Checkbox(label="Add layer normalization")
                     new_hypernetwork_use_dropout = gr.Checkbox(label="Use dropout")
                     overwrite_old_hypernetwork = gr.Checkbox(value=False, label="Overwrite Old Hypernetwork")
@@ -1342,6 +1343,7 @@ def create_ui(wrap_gradio_gpu_call):
                 overwrite_old_hypernetwork,
                 new_hypernetwork_layer_structure,
                 new_hypernetwork_activation_func,
+                new_hypernetwork_initialization_option,
                 new_hypernetwork_add_layer_norm,
                 new_hypernetwork_use_dropout
             ],
-- 
cgit v1.2.3


From 462e6ba6675bd14c0f82e465423a0eedfff82372 Mon Sep 17 00:00:00 2001
From: AngelBottomless <35677394+aria1th@users.noreply.github.com>
Date: Thu, 27 Oct 2022 15:40:24 +0900
Subject: Disable unavailable or duplicate options

---
 modules/hypernetworks/ui.py | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

(limited to 'modules/hypernetworks/ui.py')

diff --git a/modules/hypernetworks/ui.py b/modules/hypernetworks/ui.py
index 2c6c0470..c2d4b51c 100644
--- a/modules/hypernetworks/ui.py
+++ b/modules/hypernetworks/ui.py
@@ -8,7 +8,8 @@ import modules.textual_inversion.textual_inversion
 from modules import devices, sd_hijack, shared
 from modules.hypernetworks import hypernetwork
 
-keys = list(hypernetwork.HypernetworkModule.activation_dict.keys())
+not_available = ["hardswish", "multiheadattention"]
+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.
-- 
cgit v1.2.3


From f361e804ebaa5af4a10711ece2522869fb64a4c6 Mon Sep 17 00:00:00 2001
From: AngelBottomless <35677394+aria1th@users.noreply.github.com>
Date: Sat, 29 Oct 2022 08:36:50 +0900
Subject: Re enable linear

---
 modules/hypernetworks/ui.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

(limited to 'modules/hypernetworks/ui.py')

diff --git a/modules/hypernetworks/ui.py b/modules/hypernetworks/ui.py
index c2d4b51c..aad09ffc 100644
--- a/modules/hypernetworks/ui.py
+++ b/modules/hypernetworks/ui.py
@@ -9,7 +9,7 @@ from modules import devices, sd_hijack, shared
 from modules.hypernetworks import hypernetwork
 
 not_available = ["hardswish", "multiheadattention"]
-keys = list(x for x in hypernetwork.HypernetworkModule.activation_dict.keys() if x not in not_available)
+keys = ["linear"] + 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.
-- 
cgit v1.2.3


From 20194fd9752a280306fb66b57b258609b0918c46 Mon Sep 17 00:00:00 2001
From: AngelBottomless <35677394+aria1th@users.noreply.github.com>
Date: Sat, 29 Oct 2022 16:56:42 +0900
Subject: We have duplicate linear now

---
 modules/hypernetworks/ui.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

(limited to 'modules/hypernetworks/ui.py')

diff --git a/modules/hypernetworks/ui.py b/modules/hypernetworks/ui.py
index aad09ffc..c2d4b51c 100644
--- a/modules/hypernetworks/ui.py
+++ b/modules/hypernetworks/ui.py
@@ -9,7 +9,7 @@ from modules import devices, sd_hijack, shared
 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.
-- 
cgit v1.2.3


From 5f1dfbbc959855fd90ba80c0c76301d2063772fa Mon Sep 17 00:00:00 2001
From: Vladimir Mandic <mandic00@live.com>
Date: Sat, 24 Dec 2022 18:02:22 -0500
Subject: implement train api

---
 modules/api/api.py                    | 94 ++++++++++++++++++++++++++++++++++-
 modules/api/models.py                 |  9 ++++
 modules/hypernetworks/hypernetwork.py | 26 ++++++++++
 modules/hypernetworks/ui.py           | 31 ++----------
 4 files changed, 132 insertions(+), 28 deletions(-)

(limited to 'modules/hypernetworks/ui.py')

diff --git a/modules/api/api.py b/modules/api/api.py
index b43dd16b..1ceba75d 100644
--- a/modules/api/api.py
+++ b/modules/api/api.py
@@ -10,13 +10,17 @@ from fastapi.security import HTTPBasic, HTTPBasicCredentials
 from secrets import compare_digest
 
 import modules.shared as shared
-from modules import sd_samplers, deepbooru
+from modules import sd_samplers, deepbooru, sd_hijack
 from modules.api.models import *
 from modules.processing import StableDiffusionProcessingTxt2Img, StableDiffusionProcessingImg2Img, process_images
 from modules.extras import run_extras, run_pnginfo
+from modules.textual_inversion.textual_inversion import create_embedding, train_embedding
+from modules.textual_inversion.preprocess import preprocess
+from modules.hypernetworks.hypernetwork import create_hypernetwork, train_hypernetwork
 from PIL import PngImagePlugin,Image
 from modules.sd_models import checkpoints_list
 from modules.realesrgan_model import get_realesrgan_models
+from modules import devices
 from typing import List
 
 def upscaler_to_index(name: str):
@@ -97,6 +101,11 @@ class Api:
         self.add_api_route("/sdapi/v1/artist-categories", self.get_artists_categories, methods=["GET"], response_model=List[str])
         self.add_api_route("/sdapi/v1/artists", self.get_artists, methods=["GET"], response_model=List[ArtistItem])
         self.add_api_route("/sdapi/v1/refresh-checkpoints", self.refresh_checkpoints, methods=["POST"])
+        self.add_api_route("/sdapi/v1/create/embedding", self.create_embedding, methods=["POST"], response_model=CreateResponse)
+        self.add_api_route("/sdapi/v1/create/hypernetwork", self.create_hypernetwork, methods=["POST"], response_model=CreateResponse)
+        self.add_api_route("/sdapi/v1/preprocess", self.preprocess, methods=["POST"], response_model=PreprocessResponse)
+        self.add_api_route("/sdapi/v1/train/embedding", self.train_embedding, methods=["POST"], response_model=TrainResponse)
+        self.add_api_route("/sdapi/v1/train/hypernetwork", self.train_hypernetwork, methods=["POST"], response_model=TrainResponse)
 
     def add_api_route(self, path: str, endpoint, **kwargs):
         if shared.cmd_opts.api_auth:
@@ -326,6 +335,89 @@ class Api:
     def refresh_checkpoints(self):
         shared.refresh_checkpoints()
 
+    def create_embedding(self, args: dict):
+        try:
+            shared.state.begin()
+            filename = create_embedding(**args) # create empty embedding
+            sd_hijack.model_hijack.embedding_db.load_textual_inversion_embeddings() # reload embeddings so new one can be immediately used
+            shared.state.end()
+            return CreateResponse(info = "create embedding filename: {filename}".format(filename = filename))
+        except AssertionError as e:
+            shared.state.end()
+            return TrainResponse(info = "create embedding error: {error}".format(error = e))
+
+    def create_hypernetwork(self, args: dict):
+        try:
+            shared.state.begin()
+            filename = create_hypernetwork(**args) # create empty embedding
+            shared.state.end()
+            return CreateResponse(info = "create hypernetwork filename: {filename}".format(filename = filename))
+        except AssertionError as e:
+            shared.state.end()
+            return TrainResponse(info = "create hypernetwork error: {error}".format(error = e))
+
+    def preprocess(self, args: dict):
+        try:
+            shared.state.begin()
+            preprocess(**args) # quick operation unless blip/booru interrogation is enabled
+            shared.state.end()
+            return PreprocessResponse(info = 'preprocess complete')
+        except KeyError as e:
+            shared.state.end()
+            return PreprocessResponse(info = "preprocess error: invalid token: {error}".format(error = e))
+        except AssertionError as e:
+            shared.state.end()
+            return PreprocessResponse(info = "preprocess error: {error}".format(error = e))
+        except FileNotFoundError as e:
+            shared.state.end()
+            return PreprocessResponse(info = 'preprocess error: {error}'.format(error = e))
+
+    def train_embedding(self, args: dict):
+        try:
+            shared.state.begin()
+            apply_optimizations = shared.opts.training_xattention_optimizations
+            error = None
+            filename = ''
+            if not apply_optimizations:
+                sd_hijack.undo_optimizations()
+            try:
+                embedding, filename = train_embedding(**args) # can take a long time to complete
+            except Exception as e:
+                error = e
+            finally:
+                if not apply_optimizations:
+                    sd_hijack.apply_optimizations()
+                shared.state.end()
+            return TrainResponse(info = "train embedding complete: filename: {filename} error: {error}".format(filename = filename, error = error))
+        except AssertionError as msg:
+            shared.state.end()
+            return TrainResponse(info = "train embedding error: {msg}".format(msg = msg))
+
+    def train_hypernetwork(self, args: dict):
+        try:
+            shared.state.begin()
+            initial_hypernetwork = shared.loaded_hypernetwork
+            apply_optimizations = shared.opts.training_xattention_optimizations
+            error = None
+            filename = ''
+            if not apply_optimizations:
+                sd_hijack.undo_optimizations()
+            try:
+                hypernetwork, filename = train_hypernetwork(*args)
+            except Exception as e:
+                error = e
+            finally:
+                shared.loaded_hypernetwork = initial_hypernetwork
+                shared.sd_model.cond_stage_model.to(devices.device)
+                shared.sd_model.first_stage_model.to(devices.device)
+                if not apply_optimizations:
+                    sd_hijack.apply_optimizations()
+                shared.state.end()
+            return TrainResponse(info = "train embedding complete: filename: {filename} error: {error}".format(filename = filename, error = error))
+        except AssertionError as msg:
+            shared.state.end()
+            return TrainResponse(info = "train embedding error: {error}".format(error = error))
+
     def launch(self, server_name, port):
         self.app.include_router(self.router)
         uvicorn.run(self.app, host=server_name, port=port)
diff --git a/modules/api/models.py b/modules/api/models.py
index a22bc6b3..c446ce7a 100644
--- a/modules/api/models.py
+++ b/modules/api/models.py
@@ -175,6 +175,15 @@ class InterrogateRequest(BaseModel):
 class InterrogateResponse(BaseModel):
     caption: str = Field(default=None, title="Caption", description="The generated caption for the image.")
 
+class TrainResponse(BaseModel):
+    info: str = Field(title="Train info", description="Response string from train embedding or hypernetwork task.")
+
+class CreateResponse(BaseModel):
+    info: str = Field(title="Create info", description="Response string from create embedding or hypernetwork task.")
+
+class PreprocessResponse(BaseModel):
+    info: str = Field(title="Preprocess info", description="Response string from preprocessing task.")
+
 fields = {}
 for key, metadata in opts.data_labels.items():
     value = opts.data.get(key)
diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py
index c406ffb3..3182ff03 100644
--- a/modules/hypernetworks/hypernetwork.py
+++ b/modules/hypernetworks/hypernetwork.py
@@ -378,6 +378,32 @@ def report_statistics(loss_info:dict):
             print(e)
 
 
+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.
+    name = "".join( x for x in name if (x.isalnum() or x in "._- "))
+
+    fn = os.path.join(shared.cmd_opts.hypernetwork_dir, f"{name}.pt")
+    if not overwrite_old:
+        assert not os.path.exists(fn), f"file {fn} already exists"
+
+    if type(layer_structure) == str:
+        layer_structure = [float(x.strip()) for x in layer_structure.split(",")]
+
+    hypernet = modules.hypernetworks.hypernetwork.Hypernetwork(
+        name=name,
+        enable_sizes=[int(x) for x in enable_sizes],
+        layer_structure=layer_structure,
+        activation_func=activation_func,
+        weight_init=weight_init,
+        add_layer_norm=add_layer_norm,
+        use_dropout=use_dropout,
+    )
+    hypernet.save(fn)
+
+    shared.reload_hypernetworks()
+
+    return fn
+
 
 def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, gradient_step, data_root, log_directory, training_width, training_height, steps, shuffle_tags, tag_drop_out, latent_sampling_method, create_image_every, save_hypernetwork_every, template_file, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height):
     # images allows training previews to have infotext. Importing it at the top causes a circular import problem.
diff --git a/modules/hypernetworks/ui.py b/modules/hypernetworks/ui.py
index c2d4b51c..e7f9e593 100644
--- a/modules/hypernetworks/ui.py
+++ b/modules/hypernetworks/ui.py
@@ -3,39 +3,16 @@ import os
 import re
 
 import gradio as gr
-import modules.textual_inversion.preprocess
-import modules.textual_inversion.textual_inversion
+import modules.hypernetworks.hypernetwork
 from modules import devices, sd_hijack, shared
-from modules.hypernetworks import hypernetwork
 
 not_available = ["hardswish", "multiheadattention"]
-keys = list(x for x in hypernetwork.HypernetworkModule.activation_dict.keys() if x not in not_available)
+keys = list(x for x in modules.hypernetworks.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.
-    name = "".join( x for x in name if (x.isalnum() or x in "._- "))
+    filename = modules.hypernetworks.hypernetwork.create_hypernetwork(name, enable_sizes, overwrite_old, layer_structure, activation_func, weight_init, add_layer_norm, use_dropout)
 
-    fn = os.path.join(shared.cmd_opts.hypernetwork_dir, f"{name}.pt")
-    if not overwrite_old:
-        assert not os.path.exists(fn), f"file {fn} already exists"
-
-    if type(layer_structure) == str:
-        layer_structure = [float(x.strip()) for x in layer_structure.split(",")]
-
-    hypernet = modules.hypernetworks.hypernetwork.Hypernetwork(
-        name=name,
-        enable_sizes=[int(x) for x in enable_sizes],
-        layer_structure=layer_structure,
-        activation_func=activation_func,
-        weight_init=weight_init,
-        add_layer_norm=add_layer_norm,
-        use_dropout=use_dropout,
-    )
-    hypernet.save(fn)
-
-    shared.reload_hypernetworks()
-
-    return gr.Dropdown.update(choices=sorted([x for x in shared.hypernetworks.keys()])), f"Created: {fn}", ""
+    return gr.Dropdown.update(choices=sorted([x for x in shared.hypernetworks.keys()])), f"Created: {filename}", ""
 
 
 def train_hypernetwork(*args):
-- 
cgit v1.2.3


From a4a5475cfa3c68af6cb046081002a72f862ce4be Mon Sep 17 00:00:00 2001
From: aria1th <35677394+aria1th@users.noreply.github.com>
Date: Tue, 10 Jan 2023 14:56:57 +0900
Subject: Variable dropout rate

Implements variable dropout rate from #4549

Fixes hypernetwork multiplier being able to modified during training, also fixes user-errors by setting multiplier value to lower values for training.

Changes function name to match torch.nn.module standard

Fixes RNG reset issue when generating previews by restoring RNG state
---
 modules/hypernetworks/hypernetwork.py | 101 +++++++++++++++++++++++++---------
 modules/hypernetworks/ui.py           |   4 +-
 modules/ui.py                         |   4 +-
 3 files changed, 81 insertions(+), 28 deletions(-)

(limited to 'modules/hypernetworks/ui.py')

diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py
index ea3f1db9..300d3975 100644
--- a/modules/hypernetworks/hypernetwork.py
+++ b/modules/hypernetworks/hypernetwork.py
@@ -39,7 +39,7 @@ class HypernetworkModule(torch.nn.Module):
     activation_dict.update({cls_name.lower(): cls_obj for cls_name, cls_obj in inspect.getmembers(torch.nn.modules.activation) if inspect.isclass(cls_obj) and cls_obj.__module__ == 'torch.nn.modules.activation'})
 
     def __init__(self, dim, state_dict=None, layer_structure=None, activation_func=None, weight_init='Normal',
-                 add_layer_norm=False, use_dropout=False, activate_output=False, last_layer_dropout=False):
+                 add_layer_norm=False, activate_output=False, dropout_structure=None):
         super().__init__()
 
         assert layer_structure is not None, "layer_structure must not be None"
@@ -64,9 +64,12 @@ class HypernetworkModule(torch.nn.Module):
             if add_layer_norm:
                 linears.append(torch.nn.LayerNorm(int(dim * layer_structure[i+1])))
 
-            # Add dropout except last layer
-            if use_dropout and (i < len(layer_structure) - 3 or last_layer_dropout and i < len(layer_structure) - 2):
-                linears.append(torch.nn.Dropout(p=0.3))
+            # Everything should be now parsed into dropout structure, and applied here.
+            # Since we only have dropouts after layers, dropout structure should start with 0 and end with 0.
+            if dropout_structure is not None and dropout_structure[i+1] > 0:
+                assert 0 < dropout_structure[i+1] < 1, "Dropout probability should be 0 or float between 0 and 1!"
+                linears.append(torch.nn.Dropout(p=dropout_structure[i+1]))
+            # Code explanation : [1, 2, 1] -> dropout is missing when last_layer_dropout is false. [1, 2, 2, 1] -> [0, 0.3, 0, 0], when its True, [0, 0.3, 0.3, 0].
 
         self.linear = torch.nn.Sequential(*linears)
 
@@ -113,7 +116,7 @@ class HypernetworkModule(torch.nn.Module):
             state_dict[to] = x
 
     def forward(self, x):
-        return x + self.linear(x) * self.multiplier
+        return x + self.linear(x) * (HypernetworkModule.multiplier if not self.training else 1)
 
     def trainables(self):
         layer_structure = []
@@ -126,6 +129,21 @@ class HypernetworkModule(torch.nn.Module):
 def apply_strength(value=None):
     HypernetworkModule.multiplier = value if value is not None else shared.opts.sd_hypernetwork_strength
 
+#param layer_structure : sequence used for length, use_dropout : controlling boolean, last_layer_dropout : for compatibility check.
+def parse_dropout_structure(layer_structure, use_dropout, last_layer_dropout):
+    if layer_structure is None:
+        layer_structure = [1, 2, 1]
+    if not use_dropout:
+        return [0] * len(layer_structure)
+    dropout_values = [0]
+    dropout_values.extend([0.3] * (len(layer_structure) - 3))
+    if last_layer_dropout:
+        dropout_values.append(0.3)
+    else:
+        dropout_values.append(0)
+    dropout_values.append(0)
+    return dropout_values
+
 
 class Hypernetwork:
     filename = None
@@ -144,18 +162,22 @@ class Hypernetwork:
         self.add_layer_norm = add_layer_norm
         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.last_layer_dropout = kwargs.get('last_layer_dropout', True)
+        self.dropout_structure = kwargs.get('dropout_structure', None)
+        if self.dropout_structure is None:
+            self.dropout_structure = parse_dropout_structure(self.layer_structure, self.use_dropout, self.last_layer_dropout)
         self.optimizer_name = None
         self.optimizer_state_dict = None
+        self.optional_info = None
 
         for size in enable_sizes or []:
             self.layers[size] = (
                 HypernetworkModule(size, None, self.layer_structure, self.activation_func, self.weight_init,
-                                   self.add_layer_norm, self.use_dropout, self.activate_output, last_layer_dropout=self.last_layer_dropout),
+                                   self.add_layer_norm, self.activate_output, dropout_structure=self.dropout_structure),
                 HypernetworkModule(size, None, self.layer_structure, self.activation_func, self.weight_init,
-                                   self.add_layer_norm, self.use_dropout, self.activate_output, last_layer_dropout=self.last_layer_dropout),
+                                   self.add_layer_norm, self.activate_output, dropout_structure=self.dropout_structure),
             )
-        self.eval_mode()
+        self.eval()
 
     def weights(self):
         res = []
@@ -164,14 +186,14 @@ class Hypernetwork:
                 res += layer.parameters()
         return res
 
-    def train_mode(self):
+    def train(self, mode=True):
         for k, layers in self.layers.items():
             for layer in layers:
-                layer.train()
+                layer.train(mode=mode)
                 for param in layer.parameters():
-                    param.requires_grad = True
+                    param.requires_grad = mode
 
-    def eval_mode(self):
+    def eval(self):
         for k, layers in self.layers.items():
             for layer in layers:
                 layer.eval()
@@ -191,11 +213,13 @@ class Hypernetwork:
         state_dict['activation_func'] = self.activation_func
         state_dict['is_layer_norm'] = self.add_layer_norm
         state_dict['weight_initialization'] = self.weight_init
-        state_dict['use_dropout'] = self.use_dropout
         state_dict['sd_checkpoint'] = self.sd_checkpoint
         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
+        state_dict['use_dropout'] = self.use_dropout
+        state_dict['dropout_structure'] = self.dropout_structure
+        state_dict['last_layer_dropout'] = (self.dropout_structure[-2] != 0) if self.dropout_structure is not None else self.last_layer_dropout
+        state_dict['optional_info'] = self.optional_info if self.optional_info else None
 
         if self.optimizer_name is not None:
             optimizer_saved_dict['optimizer_name'] = self.optimizer_name
@@ -215,43 +239,56 @@ class Hypernetwork:
 
         self.layer_structure = state_dict.get('layer_structure', [1, 2, 1])
         print(self.layer_structure)
+        optional_info = state_dict.get('optional_info', None)
+        if optional_info is not None:
+            print(f"INFO:\n {optional_info}\n")
+            self.optional_info = optional_info
         self.activation_func = state_dict.get('activation_func', None)
         print(f"Activation function is {self.activation_func}")
         self.weight_init = state_dict.get('weight_initialization', 'Normal')
         print(f"Weight initialization is {self.weight_init}")
         self.add_layer_norm = state_dict.get('is_layer_norm', False)
         print(f"Layer norm is set to {self.add_layer_norm}")
-        self.use_dropout = state_dict.get('use_dropout', False)
+        self.dropout_structure = state_dict.get('dropout_structure', None)
+        self.use_dropout = True if self.dropout_structure is not None and any(self.dropout_structure) else state_dict.get('use_dropout', False)
         print(f"Dropout usage is set to {self.use_dropout}" )
         self.activate_output = state_dict.get('activate_output', True)
         print(f"Activate last layer is set to {self.activate_output}")
         self.last_layer_dropout = state_dict.get('last_layer_dropout', False)
+        # Dropout structure should have same length as layer structure, Every digits should be in [0,1), and last digit must be 0.
+        if self.dropout_structure is None:
+            print("Using previous dropout structure")
+            self.dropout_structure = parse_dropout_structure(self.layer_structure, self.use_dropout, self.last_layer_dropout)
+        print(f"Dropout structure is set to {self.dropout_structure}")
 
         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:
+            self.optimizer_name = optimizer_saved_dict.get('optimizer_name', 'AdamW')
             print("Loaded existing optimizer from checkpoint")
+            print(f"Optimizer name is {self.optimizer_name}")
         else:
+            self.optimizer_name = "AdamW"
             print("No saved optimizer exists in checkpoint")
 
         for size, sd in state_dict.items():
             if type(size) == int:
                 self.layers[size] = (
                     HypernetworkModule(size, sd[0], self.layer_structure, self.activation_func, self.weight_init,
-                                       self.add_layer_norm, self.use_dropout, self.activate_output, last_layer_dropout=self.last_layer_dropout),
+                                       self.add_layer_norm, self.activate_output, self.dropout_structure),
                     HypernetworkModule(size, sd[1], self.layer_structure, self.activation_func, self.weight_init,
-                                       self.add_layer_norm, self.use_dropout, self.activate_output, last_layer_dropout=self.last_layer_dropout),
+                                       self.add_layer_norm, self.activate_output, self.dropout_structure),
                 )
 
         self.name = state_dict.get('name', self.name)
         self.step = state_dict.get('step', 0)
         self.sd_checkpoint = state_dict.get('sd_checkpoint', None)
         self.sd_checkpoint_name = state_dict.get('sd_checkpoint_name', None)
+        self.eval()
 
 
 def list_hypernetworks(path):
@@ -379,9 +416,10 @@ def report_statistics(loss_info:dict):
             print(e)
 
 
-def create_hypernetwork(name, enable_sizes, overwrite_old, layer_structure=None, activation_func=None, weight_init=None, add_layer_norm=False, use_dropout=False):
+def create_hypernetwork(name, enable_sizes, overwrite_old, layer_structure=None, activation_func=None, weight_init=None, add_layer_norm=False, use_dropout=False, dropout_structure=None):
     # Remove illegal characters from name.
     name = "".join( x for x in name if (x.isalnum() or x in "._- "))
+    assert name, "Name cannot be empty!"
 
     fn = os.path.join(shared.cmd_opts.hypernetwork_dir, f"{name}.pt")
     if not overwrite_old:
@@ -390,6 +428,11 @@ def create_hypernetwork(name, enable_sizes, overwrite_old, layer_structure=None,
     if type(layer_structure) == str:
         layer_structure = [float(x.strip()) for x in layer_structure.split(",")]
 
+    if use_dropout and dropout_structure and type(dropout_structure) == str:
+        dropout_structure = [float(x.strip()) for x in dropout_structure.split(",")]
+    else:
+        dropout_structure = [0] * len(layer_structure)
+
     hypernet = modules.hypernetworks.hypernetwork.Hypernetwork(
         name=name,
         enable_sizes=[int(x) for x in enable_sizes],
@@ -398,6 +441,7 @@ def create_hypernetwork(name, enable_sizes, overwrite_old, layer_structure=None,
         weight_init=weight_init,
         add_layer_norm=add_layer_norm,
         use_dropout=use_dropout,
+        dropout_structure=dropout_structure
     )
     hypernet.save(fn)
 
@@ -480,7 +524,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, gradient_step,
         shared.sd_model.first_stage_model.to(devices.cpu)
 
     weights = hypernetwork.weights()
-    hypernetwork.train_mode()
+    hypernetwork.train()
 
     # Here we use optimizer from saved HN, or we can specify as UI option.
     if hypernetwork.optimizer_name in optimizer_dict:
@@ -594,7 +638,11 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, gradient_step,
                 if images_dir is not None and steps_done % create_image_every == 0:
                     forced_filename = f'{hypernetwork_name}-{steps_done}'
                     last_saved_image = os.path.join(images_dir, forced_filename)
-                    hypernetwork.eval_mode()
+                    hypernetwork.eval()
+                    rng_state = torch.get_rng_state()
+                    cuda_rng_state = None
+                    if torch.cuda.is_available():
+                        cuda_rng_state = torch.cuda.get_rng_state_all()
                     shared.sd_model.cond_stage_model.to(devices.device)
                     shared.sd_model.first_stage_model.to(devices.device)
 
@@ -627,7 +675,10 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, gradient_step,
                     if unload:
                         shared.sd_model.cond_stage_model.to(devices.cpu)
                         shared.sd_model.first_stage_model.to(devices.cpu)
-                    hypernetwork.train_mode()
+                    torch.set_rng_state(rng_state)
+                    if torch.cuda.is_available():
+                        torch.cuda.set_rng_state_all(cuda_rng_state)
+                    hypernetwork.train()
                     if image is not None:
                         shared.state.current_image = image
                         last_saved_image, last_text_info = images.save_image(image, images_dir, "", p.seed, p.prompt, shared.opts.samples_format, processed.infotexts[0], p=p, forced_filename=forced_filename, save_to_dirs=False)
@@ -649,7 +700,7 @@ Last saved image: {html.escape(last_saved_image)}<br/>
     finally:
         pbar.leave = False
         pbar.close()
-        hypernetwork.eval_mode()
+        hypernetwork.eval()
         #report_statistics(loss_dict)
 
     filename = os.path.join(shared.cmd_opts.hypernetwork_dir, f'{hypernetwork_name}.pt')
diff --git a/modules/hypernetworks/ui.py b/modules/hypernetworks/ui.py
index e7f9e593..81e3f519 100644
--- a/modules/hypernetworks/ui.py
+++ b/modules/hypernetworks/ui.py
@@ -9,8 +9,8 @@ from modules import devices, sd_hijack, shared
 not_available = ["hardswish", "multiheadattention"]
 keys = list(x for x in modules.hypernetworks.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):
-    filename = modules.hypernetworks.hypernetwork.create_hypernetwork(name, enable_sizes, overwrite_old, layer_structure, activation_func, weight_init, add_layer_norm, use_dropout)
+def create_hypernetwork(name, enable_sizes, overwrite_old, layer_structure=None, activation_func=None, weight_init=None, add_layer_norm=False, use_dropout=False, dropout_structure=None):
+    filename = modules.hypernetworks.hypernetwork.create_hypernetwork(name, enable_sizes, overwrite_old, layer_structure, activation_func, weight_init, add_layer_norm, use_dropout, dropout_structure)
 
     return gr.Dropdown.update(choices=sorted([x for x in shared.hypernetworks.keys()])), f"Created: {filename}", ""
 
diff --git a/modules/ui.py b/modules/ui.py
index b6079aec..9b9081b5 100644
--- a/modules/ui.py
+++ b/modules/ui.py
@@ -1268,6 +1268,7 @@ def create_ui():
                     new_hypernetwork_initialization_option = gr.Dropdown(value = "Normal", label="Select Layer weights initialization. Recommended: Kaiming for relu-like, Xavier for sigmoid-like, Normal otherwise", choices=["Normal", "KaimingUniform", "KaimingNormal", "XavierUniform", "XavierNormal"], elem_id="train_new_hypernetwork_initialization_option")
                     new_hypernetwork_add_layer_norm = gr.Checkbox(label="Add layer normalization", elem_id="train_new_hypernetwork_add_layer_norm")
                     new_hypernetwork_use_dropout = gr.Checkbox(label="Use dropout", elem_id="train_new_hypernetwork_use_dropout")
+                    new_hypernetwork_dropout_structure = gr.Textbox("0, 0, 0", label="Enter hypernetwork Dropout structure (or empty). Recommended : 0~0.35 incrementing sequence: 0, 0.05, 0.15", placeholder="1st and last digit must be 0 and values should be between 0 and 1. ex:'0, 0.01, 0'")
                     overwrite_old_hypernetwork = gr.Checkbox(value=False, label="Overwrite Old Hypernetwork", elem_id="train_overwrite_old_hypernetwork")
 
                     with gr.Row():
@@ -1414,7 +1415,8 @@ def create_ui():
                 new_hypernetwork_activation_func,
                 new_hypernetwork_initialization_option,
                 new_hypernetwork_add_layer_norm,
-                new_hypernetwork_use_dropout
+                new_hypernetwork_use_dropout,
+                new_hypernetwork_dropout_structure
             ],
             outputs=[
                 train_hypernetwork_name,
-- 
cgit v1.2.3