LoRA inspector

• LoRA inspector
  • Install
  • Usage
    • Inspect
    • Save meta
    • Average weights
    • Tag frequency
    • Dataset
    • Definition
  • Update metadata
    • Usage
  • Changelog
  • Development
  • Future
  • Reference

Inspect LoRA files for meta info (and hopefully quantitative analysis of the LoRA weights).

• view training parameters
• extract metadata to be stored (we can store it in JSON currently)
• only safetensors are supported (want to support all LoRA files)
• only metadata from kohya-ss LoRA (want to parse all metadata in LoRA files)

---

NOTE this is a work in progress and not meant for production use. NOTE

---

Install

Clone this repo or download the python script file.

Requires dependencies:

torch
safetensors
tqdm

Can install them one of the following:

• Add this script to your training directory and use the virtual environment (venv). RECOMMENDED
• Make/use with a venv/conda
• pip install safetensors tqdm (See Get started for instructions on how to install PyTorch)

Usage

Inspect

$ python lora-inspector.py --help
usage: lora-inspector.py [-h] [-s] [-w] [-t] [-d] lora_file_or_dir

positional arguments:
  lora_file_or_dir  Directory containing the lora files

options:
  -h, --help        show this help message and exit
  -s, --save_meta   Should we save the metadata to a file?
  -w, --weights     Show the average magnitude and strength of the weights
  -t, --tags        Show the most common tags in the training set
  -d, --dataset     Show the dataset metadata including directory names and number of images

You can add a directory or file:

$ python lora-inspector.py /mnt/900/training/sets/landscape-2023-11-06-200718-e4d7120b -w
/mnt/900/training/sets/landscape-2023-11-06-200718-e4d7120b/landscape-2023-11-06-200718-e4d7120b-000015.safetensors
Date: 2023-11-06T20:16:34 Title: landscape
License: CreativeML Open RAIL-M Author: rockerBOO
Description: High quality landscape photos
Resolution: 512x512 Architecture: stable-diffusion-v1/lora
Network Dim/Rank: 16.0 Alpha: 8.0 Dropout: 0.3 dtype: torch.float32
Module: networks.lora : {'block_dims': '4,4,4,4,4,4,4,4,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8', 'block_alphas': '16,16,16,16,16,16,16,16,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32', 'block_dropout': '0.01, 0.010620912260804992, 0.01248099020159499, 0.015572268683063176, 0.01988151037617019, 0.02539026244641935, 0.032074935571726845, 0.03990690495552037, 0.04885263290251277, 0.058873812432261884, 0.0699275313155418, 0.08196645583109653, 0.09493903345590124, 0.10878971362098, 0.12345918558747097, 0.13888463242431537, 0.155, 0.17173627983648962, 0.18902180461412393, 0.20678255506208312, 0.22494247692026895, 0.2434238066153228, 0.26214740425618505, 0.2810330925232585', 'dropout': 0.3}
Learning Rate (LR): 2e-06 UNet LR: 1.0 TE LR: 1.0
Optimizer: prodigyopt.prodigy.Prodigy(weight_decay=0.1,betas=(0.9, 0.9999),d_coef=1.5,use_bias_correction=True)
Scheduler: cosine  Warmup steps: 0
Epoch: 15 Batches per epoch: 57 Gradient accumulation steps: 24
Train images: 57 Regularization images: 0
Noise offset: 0.05 Adaptive noise scale: 0.01 IP noise gamma: 0.1  Multires noise discount: 0.3
Min SNR gamma: 5.0 Zero terminal SNR: True Debiased Estimation: True
UNet weight average magnitude: 0.7865518983141094
UNet weight average strength: 0.00995593195090544
No Text Encoder found in this LoRA
----------------------
/mnt/900/training/sets/landscape-2023-11-06-200718-e4d7120b/landscape-2023-11-06-200718-e4d7120b.safetensors
Date: 2023-11-06T20:27:12 Title: landscape
License: CreativeML Open RAIL-M Author: rockerBOO
Description: High quality landscape photos
Resolution: 512x512 Architecture: stable-diffusion-v1/lora
Network Dim/Rank: 16.0 Alpha: 8.0 Dropout: 0.3 dtype: torch.float32
Module: networks.lora : {'block_dims': '4,4,4,4,4,4,4,4,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8', 'block_alphas': '16,16,16,16,16,16,16,16,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32', 'block_dropout': '0.01, 0.010620912260804992, 0.01248099020159499, 0.015572268683063176, 0.01988151037617019, 0.02539026244641935, 0.032074935571726845, 0.03990690495552037, 0.04885263290251277, 0.058873812432261884, 0.0699275313155418, 0.08196645583109653, 0.09493903345590124, 0.10878971362098, 0.12345918558747097, 0.13888463242431537, 0.155, 0.17173627983648962, 0.18902180461412393, 0.20678255506208312, 0.22494247692026895, 0.2434238066153228, 0.26214740425618505, 0.2810330925232585', 'dropout': 0.3}
Learning Rate (LR): 2e-06 UNet LR: 1.0 TE LR: 1.0
Optimizer: prodigyopt.prodigy.Prodigy(weight_decay=0.1,betas=(0.9, 0.9999),d_coef=1.5,use_bias_correction=True)
Scheduler: cosine  Warmup steps: 0
Epoch: 30 Batches per epoch: 57 Gradient accumulation steps: 24
Train images: 57 Regularization images: 0
Noise offset: 0.05 Adaptive noise scale: 0.01 IP noise gamma: 0.1  Multires noise discount: 0.3
Min SNR gamma: 5.0 Zero terminal SNR: True Debiased Estimation: True
UNet weight average magnitude: 0.8033398082829257
UNet weight average strength: 0.010114916750103732
No Text Encoder found in this LoRA
----------------------

$ python lora-inspector.py /mnt/900/lora/testing/landscape-2023-11-06-200718-e4d7120b.safetensors
/mnt/900/lora/testing/landscape-2023-11-06-200718-e4d7120b.safetensors
Date: 2023-11-06T20:27:12 Title: landscape
License: CreativeML Open RAIL-M Author: rockerBOO
Description: High quality landscape photos
Resolution: 512x512 Architecture: stable-diffusion-v1/lora
Network Dim/Rank: 16.0 Alpha: 8.0 Dropout: 0.3 dtype: torch.float32
Module: networks.lora : {'block_dims': '4,4,4,4,4,4,4,4,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8', 'block_alphas': '16,16,16,16,16,16,16,16,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32,32', 'block_dropout': '0.01, 0.010620912260804992, 0.01248099020159499, 0.015572268683063176, 0.01988151037617019, 0.02539026244641935, 0.032074935571726845, 0.03990690495552037, 0.04885263290251277, 0.058873812432261884, 0.0699275313155418, 0.08196645583109653, 0.09493903345590124, 0.10878971362098, 0.12345918558747097, 0.13888463242431537, 0.155, 0.17173627983648962, 0.18902180461412393, 0.20678255506208312, 0.22494247692026895, 0.2434238066153228, 0.26214740425618505, 0.2810330925232585', 'dropout': 0.3}
Learning Rate (LR): 2e-06 UNet LR: 1.0 TE LR: 1.0
Optimizer: prodigyopt.prodigy.Prodigy(weight_decay=0.1,betas=(0.9, 0.9999),d_coef=1.5,use_bias_correction=True)
Scheduler: cosine  Warmup steps: 0
Epoch: 30 Batches per epoch: 57 Gradient accumulation steps: 24
Train images: 57 Regularization images: 0
Noise offset: 0.05 Adaptive noise scale: 0.01 IP noise gamma: 0.1  Multires noise discount: 0.3
Min SNR gamma: 5.0 Zero terminal SNR: True Debiased Estimation: True
UNet weight average magnitude: 0.8033398082829257
UNet weight average strength: 0.010114916750103732
No Text Encoder found in this LoRA
----------------------

Save meta

We also have support for saving the meta that is extracted and converted from strings. We can then save those to a JSON file. These will save the metadata into meta/alorafile.safetensors-{session_id}.json in the current working directory.

$ python lora-inspector.py ~/loras/alorafile.safetensors --save_meta

$ python lora-inspector.py /mnt/900/training/cyberpunk-anime-21-min-snr/unet-1.15-te-1.15-noise-0.1-steps--linear-DAdaptation-networks.lora/last.safetensors --save_meta
/mnt/900/training/cyberpunk-anime-21-min-snr/unet-1.15-te-1.15-noise-0.1-steps--linear-DAdaptation-networks.lora/last.safetensors
train images: 1005 regularization images: 32000
learning rate: 1.15 unet: 1.15 text encoder: 1.15
epoch: 1 batches: 2025
optimizer: dadaptation.dadapt_adam.DAdaptAdam lr scheduler: linear
network dim/rank: 8.0 alpha: 4.0 module: networks.lora
----------------------

Average weights

Find the average magnitude and average strength of your weights. Compare these with other LoRAs to see how powerful or not so powerful your weights are. NOTE Weights shown are not conclusive to a good value. They are an initial example.

$ python lora-inspector.py /mnt/900/lora/studioGhibliStyle_offset.safetensors -w
UNet weight average magnitude: 4.299801171795097
UNet weight average strength: 0.01127891692482733
Text Encoder weight average magnitude: 3.128134997225176
Text Encoder weight average strength: 0.00769676965767913

Tag frequency

Shows the frequency of a tag (words separated by commas). Trigger words are generally the most frequent, as they would use that word across the whole training dataset.

$ python lora-inspector.py -t /mnt/900/lora/booscapes.safetensors
...
-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
Tags
-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
4k photo”                         23
spectacular mountains             17
award winning nature photo        16
ryan dyar                         14
image credit nasa nat geo         11
sunset in a valley                11
garden                            10
british columbia                  10
dramatic autumn landscape         10
autumn mountains                  10
an amazing landscape image        10
austria                           9
nature scenery                    9
pristine water                    9
boreal forest                     9
scenic view of river              9
alpes                             9
mythical floral hills             8
misty environment                 8
a photo of a lake on a sunny day  8
majestic beautiful world          8
breathtaking stars                8
lush valley                       7
dramatic scenery                  7
solar storm                       7
siberia                           7
cosmic skies                      7
dolomites                         7
oregon                            6
landscape photography 4k          6
very long spires                  6
beautiful forests and trees       6
wildscapes                        6
mountain behind meadow            6
colorful wildflowers              6
photo of green river              6
beautiful night sky               6
switzerland                       6
natural dynamic range color       6
middle earth                      6
jessica rossier color scheme      6
arizona                           6
enchanting and otherworldly       6

Dataset

A pretty basic view of the dataset with the directories and number of images.

$ python lora-inspector.py -d /mnt/900/lora/booscapes.safetensors
Dataset dirs: 2
    [source] 50 images
    [p7] 4 images

Definition

• epoch: an epoch is seeing the entire dataset once

• Batches per epoch: how many batches per each epoch (does not include gradient accumulation steps)

• Gradient accumulation steps: gradient accumulation steps

• Train images: number of training images you have

• Regularization images: number of regularization images

• Scheduler: the learning rate scheduler (cosine, cosine_with_restart, linear, constant, …)

• Optimizer: the optimizer (Adam, Prodigy, DAdaptation, Lion, …)

• Network dim/rank: the rank of the LoRA network

• Alpha: the alpha to the rank of the LoRA network

• Module: the python module that created the network

• Noise offset: noise offset option

• Adaptive noise scale: adaptive noise scale

• IP noise gamma: Input Perturbation noise gamma Input Perturbation Reduces Exposure Bias in Diffusion Models

  • …we propose a very simple but effective training regularization, consisting in perturbing the ground truth samples to simulate the inference time prediction errors.

• multires noise discount: multires noise discount (See Multi-Resolution Noise for Diffusion Model Training)

• multires noise scale: multires noise scale

• average magnitude: square each weight, add them up, get the square root

• average strength: abs each weight, add them up, get average

• debiased estimation loss: Debias the Training of Diffusion Models

Update metadata

Simple script to update your metadata values. Helpful for changing ss_output_name for applications that use this value to set a good name for it.

To see your current metadata values, save the metadata using lora-inspector.py --save_meta ... and inspect the JSON file.

$ python update_metadata.py --help
usage: update_metadata.py [-h] [--key KEY] [--value VALUE] safetensors_file

positional arguments:
  safetensors_file

options:
  -h, --help        show this help message and exit
  --key KEY         Key to change in the metadata
  --value VALUE     Value to set to the metadata

Usage

$ python update_metadata.py /mnt/900/lora/testing/armored-core-2023-08-02-173642-ddb4785e.safetensors --key ss_output_name --value mechBOO_v2
Updated ss_output_name with mechBOO_v2
Saved to /mnt/900/lora/testing/armored-core-2023-08-02-173642-ddb4785e.safetensors

Changelog

• 2023-11-11 — Add debiased estimation loss, dtype (precision)
• 2023-10-27 — Add IP noise gamma
• 2023-08-27 — Add max_grad_norm, scale weight norms, gradient accumulation steps, dropout, and datasets
• 2023-08-08 — Add simple metadata updater script
• 2023-07-31 — Add SDXL support
• 2023-07-17 — Add network dropout, scale weight norms, adaptive noise scale, and steps
• 2023-07-06 — Add Tag Frequency
• 2023-04-12 — Add gradient norm, gradient checkpoint metadata
• 2023-04-03 — Add clip_skip, segment off LoCon/conv layers in average weights
• 2023-04-03 — Add noise_offset, min_snr_gamma (when added to kohya-ss), and network_args (for LoCon values)
• 2023-04-02 — Add --weights which allows you to see the average magnitude and strength of your LoRA UNet and Text Encoder weights.

Development

Formatted using black.

Future

What else do you want to see? Make an issue or a PR.

Use cases/ideas that this can expand into:

• Extract metadata from LoRA files to be used elsewhere
• Put the metadata into a database or search engine to find specific trainings
• Find possible issues with the training due to the metadata
• Compare LoRA files together

Reference

• https://github.com/Zyin055/Inspect-Embedding-Training
• https://github.com/kohya-ss/sd-scripts