From e817fac5424f7f19a5f20071dc08ce4e483d0636 Mon Sep 17 00:00:00 2001 From: "Alex \"mcmonkey\" Goodwin" Date: Mon, 27 Mar 2023 22:29:23 -0700 Subject: [PATCH] better defaults --- modules/training.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/modules/training.py b/modules/training.py index 1949fa4e..7bcecb38 100644 --- a/modules/training.py +++ b/modules/training.py @@ -34,8 +34,8 @@ def create_train_interface(): learning_rate = gr.Textbox(label='Learning Rate', value='3e-4', info='Learning rate, in scientific notation. 3e-4 is a good starting base point. 1e-2 is extremely high, 1e-6 is extremely low.') # TODO: What is the actual maximum rank? Likely distinct per model. This might be better to somehow be on a log scale. - lora_rank = gr.Slider(label='LoRA Rank', value=8, minimum=0, maximum=1024, step=4, info='LoRA Rank, or dimension count. Higher values produce a larger file with better control over the model\'s content. Smaller values produce a smaller file with less overall control. Small values like 4 or 8 are great for stylistic guidance, high values like 128 or 256 are good for teaching content upgrades. Higher ranks also require higher VRAM.') - lora_alpha = gr.Slider(label='LoRA Alpha', value=16, minimum=0, maximum=2048, step=4, info='LoRA Alpha. This divided by the rank becomes the scaling of the LoRA. Higher means stronger. A good standard value is twice your Rank.') + lora_rank = gr.Slider(label='LoRA Rank', value=32, minimum=0, maximum=1024, step=4, info='LoRA Rank, or dimension count. Higher values produce a larger file with better control over the model\'s content. Smaller values produce a smaller file with less overall control. Small values like 4 or 8 are great for stylistic guidance, high values like 128 or 256 are good for teaching content upgrades. Higher ranks also require higher VRAM.') + lora_alpha = gr.Slider(label='LoRA Alpha', value=64, minimum=0, maximum=2048, step=4, info='LoRA Alpha. This divided by the rank becomes the scaling of the LoRA. Higher means stronger. A good standard value is twice your Rank.') # TODO: Better explain what this does, in terms of real world effect especially. lora_dropout = gr.Slider(label='LoRA Dropout', minimum=0.0, maximum=1.0, step=0.025, value=0.05, info='Percentage probability for dropout of LoRA layers.') cutoff_len = gr.Slider(label='Cutoff Length', minimum=0, maximum=2048, value=256, step=32, info='Cutoff length for text input. Essentially, how long of a line of text to feed in at a time. Higher values require drastically more VRAM.') @@ -52,7 +52,7 @@ def create_train_interface(): with gr.Row(): raw_text_file = gr.Dropdown(choices=get_dataset('training/datasets', 'txt'), value='None', label='Text File', info='The raw text file to use for training.') ui.create_refresh_button(raw_text_file, lambda : None, lambda : {'choices': get_dataset('training/datasets', 'txt')}, 'refresh-button') - overlap_len = gr.Slider(label='Overlap Length', minimum=0,maximum=512, value=32, step=8, info='Overlap length - ie how many tokens from the prior chunk of text to include into the next chunk. (The chunks themselves will be of a size determined by Cutoff Length above)') + overlap_len = gr.Slider(label='Overlap Length', minimum=0,maximum=512, value=128, step=16, info='Overlap length - ie how many tokens from the prior chunk of text to include into the next chunk. (The chunks themselves will be of a size determined by Cutoff Length above). Setting overlap to exactly half the cutoff length may be ideal.') with gr.Row(): start_button = gr.Button("Start LoRA Training")