Files
mai/moondream/generate_dataset.py

186 lines
6.3 KiB
Python

import dotenv
dotenv.load_dotenv()
import os
import sys
import torch
import datasets
import diffusers
import dotenv
import transformers
import argparse
from .hyperparams import MOONDREAM_REVISION
print(f"HF_HOME set to {os.getenv('HF_HOME')}")
# DATASET_SIZE = 10000
# ROWS_PER_DS = 1250
BATCH_SIZE = 4
PARQUET_BATCH_SIZE = 200
SKIP_PARQUET_BATCH = 203
parser = argparse.ArgumentParser()
parser.add_argument("-m", "--message")
args = parser.parse_args()
auth_token = os.getenv("HF_ACCESS_TOKEN")
if not auth_token:
print("huggingface access token not provided! please use the HF_ACCESS_TOKEN env var.")
sys.exit(1)
else:
print("huggingface access token loaded!")
tokenizer = transformers.AutoTokenizer.from_pretrained("vikhyatk/moondream2")
moondream = transformers.AutoModelForCausalLM.from_pretrained(
"vikhyatk/moondream2",
revision=MOONDREAM_REVISION,
trust_remote_code=True,
torch_dtype=torch.bfloat16,
device_map={"": "cuda"},
)
pipe = diffusers.StableDiffusion3Pipeline.from_pretrained(
"stabilityai/stable-diffusion-3.5-large",
torch_dtype=torch.bfloat16,
token=auth_token,
device_map="balanced",
)
def collate(batch):
images = []
keywords = []
for sample in batch:
images.append(sample["image"])
keywords.append([""])
return images, keywords
# flickr_dataset = datasets.load_dataset("nlphuji/flickr30k", split="test", streaming=True)\
# .select_columns(["image"])\
wiki_art_dataset = datasets.load_dataset("huggan/wikiart", split="train", streaming=True)\
.select_columns(["image"])
# anime_dataset_ft = datasets.Features({"image": datasets.Image(decode=True)})
# anime_dataset = datasets.load_dataset("animelover/danbooru2022", "1-full", trust_remote_code=True, split="train", streaming=True, features=anime_dataset_ft)\
# .select_columns(["image"])\
# .take(ROWS_PER_DS)\
# .add_column("question", ["Describe this image in one sentence. Include the word anime in the sentence."] * ROWS_PER_DS)\
# .add_column("keywords", [["anime"]] * ROWS_PER_DS)
# coco_dataset = datasets.load_dataset("detection-datasets/coco", split="train", streaming=True)\
# .select_columns(["image"])\
# .take(ROWS_PER_DS)\
# .add_column("question", ["Describe this image in one sentence."] * ROWS_PER_DS)\
# .add_column("keywords", [[""]] * ROWS_PER_DS)
# movie_poster_dataset = datasets.load_dataset("skvarre/movie_posters-100k", split="train", streaming=True)\
# .select_columns(["image"])\
# .take(ROWS_PER_DS)\
# .add_column("question", ["Describe this image in one sentence."] * ROWS_PER_DS)\
# .add_column("keywords", [[""]] * ROWS_PER_DS)
# cars_dataset = datasets.load_dataset("tanganke/stanford_cars", split="train", streaming=True)\
# .select_columns(["image"])\
# .take(ROWS_PER_DS)\
# .add_column("question", ["Describe this image in one sentence."] * ROWS_PER_DS)\
# .add_column("keywords", [[""]] * ROWS_PER_DS)
# website_dataset = datasets.load_dataset("silatus/1k_Website_Screenshots_and_Metadata", split="train", streaming=True)\
# .select_columns(["image"])\
# .take(ROWS_PER_DS)\
# .add_column("question", ["Describe this image in one sentence."] * ROWS_PER_DS)\
# .add_column("keywords", [[""]] * ROWS_PER_DS)
# movie_scene_dataset = datasets.load_dataset("unography/movie-scenes-resized-captioned", split="train", streaming=True)\
# .select_columns(["image"])\
# .take(ROWS_PER_DS)\
# .add_column("question", ["Describe this image in one sentence."] * ROWS_PER_DS)\
# .add_column("keywords", [[""]] * ROWS_PER_DS)
# ds = datasets.concatenate_datasets([
# flickr_dataset,
# wiki_art_dataset,
# anime_dataset,
# coco_dataset,
# movie_poster_dataset,
# cars_dataset,
# website_dataset,
# movie_scene_dataset,
# ]).cast_column("image", datasets.Image(decode=True)).skip(SKIP_PARQUET_BATCH * PARQUET_BATCH_SIZE)
ds = wiki_art_dataset.cast_column("image", datasets.Image(decode=True))
data_loader = torch.utils.data.DataLoader(
ds,
batch_size=BATCH_SIZE,
collate_fn=collate
)
temp_ds = {
"image": [],
"keywords": [],
"caption": [],
"generated_image": []
}
temp_ds_size = 0
ds_features = datasets.Features({
"image": datasets.Image(),
"keywords": datasets.Sequence(datasets.Value(dtype="string")),
"caption": datasets.Value(dtype="string"),
"generated_image": datasets.Image(),
})
generator = torch.Generator(device="cpu").manual_seed(12321313)
batch_count = SKIP_PARQUET_BATCH
for batch_index, batch in enumerate(data_loader):
images, keywords = batch
prompts = []
for i, img in enumerate(images):
caption = moondream.caption(img, length="normal")["caption"]
add_keywords = len(keywords[i]) > 0 and keywords[i][0] != ""
for k in keywords[i]:
if k and k in caption:
add_keywords = False
break
prompt = caption
if add_keywords:
prompt = f"{', '.join(keywords[i])}, {caption}"
prompts.append(prompt)
gen_imgs = pipe(
prompts,
num_inference_steps=28,
guidance_scale=3.5,
generator=generator,
max_sequence_length=512,
).images
temp_ds["image"].extend(images)
temp_ds["caption"].extend(prompts)
temp_ds["keywords"].extend(keywords)
temp_ds["generated_image"].extend(gen_imgs)
temp_ds_size += BATCH_SIZE
if temp_ds_size == PARQUET_BATCH_SIZE:
batch_ds = datasets.Dataset.from_dict(temp_ds, features=ds_features)
batch_ds.to_parquet(
f"data/batch_{batch_count}.parquet",
)
temp_ds_size = 0
temp_ds["image"].clear()
temp_ds["caption"].clear()
temp_ds["keywords"].clear()
temp_ds["generated_image"].clear()
batch_count += 1