If running on GCP use the following configurations for your VM instance, Deep Learning VM for PyTorch 2.3 with CUDA 12.1, M125, Debian 11, Python 3.10, with PyTorch 2.3 and fast.ai preinstalled

Training data for dreambooth needs to be on Huggingface Hub for easy access from any system. The dataset used in the code below is at https://huggingface.co/datasets/bhuv1-c/valid-warehouses-dataset.

```bash

!/bin/bash

Clone the diffusers repository

git clone https://github.com/bhuvvaan/dreambooth-training.git

cd dreambooth-training

Create a virtual environment

python3 -m venv diffusion-venv

Activate the virtual environment

source diffusion-venv/bin/activate

Install the diffusers package

pip install .

Install accelerate

pip install accelerate

Configure accelerate

accelerate config

cd examples/dreambooth

Install requirements for the dreambooth example

pip install -U -r requirements.txt

Login to Hugging Face

huggingface-cli login --token #your token here

export MODELNAME="CompVis/stable-diffusion-v1-4" # Model on the hub export INSTANCEHFREPO="bhuv1-c/fruits-for-intent" # Dataset on the hub export INSTANCEDIR="fruits-for-intent" # Local dataset directory export OUTPUT_DIR="md-intent-prediction-data-2" # Output folder on the hub

Download dataset from Hugging Face Hub (only images)

python3 <<EOF import os from huggingfacehub import snapshotdownload

Download all files

repodir = snapshotdownload( repoid="bhuv1-c/fruits-for-intent", # Correct string format localdir="fruits-for-intent", repotype="dataset", localdirusesymlinks=False # Prevents extra cache files )

Filter only image files

validexts = {'.png', '.jpg', '.jpeg', '.bmp', '.gif', '.webp'} for root, _, files in os.walk(repodir): for file in files: if not any(file.lower().endswith(ext) for ext in valid_exts): os.remove(os.path.join(root, file)) # Delete non-image files

Check if dataset is present before training

imagefiles = [f for f in os.listdir(repodir) if f.lower().endswith(tuple(validexts))] if not imagefiles: print("Error: No images found in dataset. Exiting.") exit(1)

print("Dataset downloaded and cleaned successfully!") EOF

Ensure dataset exists before proceeding

if [ ! -d "$INSTANCEDIR" ]; then echo "Error: Dataset folder '$INSTANCEDIR' does not exist!" exit 1 fi

Remove any .cache directories

if [ -d "$INSTANCEDIR/.cache" ]; then rm -rf "$INSTANCEDIR/.cache" echo "Removed .cache directory." fi

Run training

accelerate launch traindreambooth.py \ --pretrainedmodelnameorpath=$MODELNAME \ --instancedatadir=$INSTANCEDIR \ --outputdir=$OUTPUTDIR \ --traintextencoder \ --instanceprompt="a close-up photo of an apple, two lemon, one lime, one onion, three oranges and pear on a rustic wooden table"\ --classprompt="fruits on table" \ --resolution=256 \ --trainbatchsize=1 \ --gradientaccumulationsteps=4 \ --learningrate=3e-6 \ --lrscheduler="constant" \ --lrwarmupsteps=0 \ --maxtrainsteps=800 \ --pushto_hub

```

Hyperparameters can be changed according to the users convenience. A helpful guide on hyperparameters can be found at https://huggingface.co/blog/dreambooth.

For original diffusers readme, refer to https://github.com/huggingface/diffusers