This model is a fine-tuned version of facebook/opt-350m on the Gustavosta/Stable-Diffusion-Prompts dataset. It achieves the following results on the evaluation set:

Loss: 1.2987
eval_steps_per_second = 16.623
perplexity = 3.6644

example

output (on DALL-E 2, but as words are words, works anywhere)

Training and evaluation data

refer to the Gustavosta/Stable-Diffusion-Prompts dataset.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0001
train_batch_size: 8
eval_batch_size: 2
seed: 42
distributed_type: multi-GPU
num_devices: 2
gradient_accumulation_steps: 32
total_train_batch_size: 512
total_eval_batch_size: 4
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: cosine
lr_scheduler_warmup_ratio: 0.05
num_epochs: 10.0

Training results

Training Loss	Epoch	Step	Validation Loss
2.8568	0.95	16	2.5937
2.2487	1.95	32	2.1050
1.9011	2.95	48	1.8082
1.6837	3.95	64	1.6178
1.4887	4.95	80	1.4897
1.3812	5.95	96	1.4017
1.2944	6.95	112	1.3437
1.2574	7.95	128	1.3127
1.2325	8.95	144	1.3009
1.2223	9.95	160	1.2987

Framework versions

Transformers 4.25.0.dev0
Pytorch 1.13.0+cu117
Datasets 2.6.1
Tokenizers 0.13.1

pszemraj/opt-350m-magicprompt-SD

opt-350m-magicprompt-SD