Train hypernetwork implemented in stable-diffusion-webui with kohya_ss/sd-scripts.
pip install git+https://github.com/ddPn08/losalina.git
Warning
Saving in
safetensors
format is not available.
Commandline arguments
python sd-scripts/train_network.py \
--network_module losalina.hypernetwork \
--network_args "layer_structure=[1, 2, 2, 1]" \
"weight_init=KaimingNormal" \
"activation_func=mish" \
"add_layer_norm=True" \
"dropout_structure=[0, 0.3, 0.3, 0]" \
"last_layer_dropout=False"
Config file
network_module = "losalina.hypernetwork"
network_args = [
"layer_structure=[1, 2, 2, 1]",
"weight_init=KaimingNormal",
"activation_func=mish",
"add_layer_norm=True",
"dropout_structure=[0, 0.3, 0.3, 0]",
"last_layer_dropout=False",
]
layer_structure |
Layer structure of hypernetwork. | ex.) [1, 2, 2, 1] |
weight_init |
Weight initialization method. | Normal , XavierUniform , XavierNormal , KaimingUniform , KaimingNormal |
activation_func |
Activation function. | linear , relu , leakyrelu , elu , swish , tahh , sigmoid ... |
add_layer_norm |
Add layer normalization. | True , False |
dropout_structure |
Dropout structure. | ex.) [0, 0.3, 0.3, 0] |
last_layer_dropout |
Add dropout to last layer. | True , False |