Files
lancedb/notebooks/diffusiondb/datagen.py
2023-06-01 10:23:29 -07:00

109 lines
3.1 KiB
Python
Executable File

#!/usr/bin/env python
#
# Copyright 2023 LanceDB Developers
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Dataset hf://poloclub/diffusiondb
"""
import io
from argparse import ArgumentParser
from multiprocessing import Pool
import lance
import lancedb
import pyarrow as pa
from datasets import load_dataset
from PIL import Image
from transformers import CLIPModel, CLIPProcessor, CLIPTokenizerFast
MODEL_ID = "openai/clip-vit-base-patch32"
device = "cuda"
tokenizer = CLIPTokenizerFast.from_pretrained(MODEL_ID)
model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32").to(device)
processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
schema = pa.schema(
[
pa.field("prompt", pa.string()),
pa.field("seed", pa.uint32()),
pa.field("step", pa.uint16()),
pa.field("cfg", pa.float32()),
pa.field("sampler", pa.string()),
pa.field("width", pa.uint16()),
pa.field("height", pa.uint16()),
pa.field("timestamp", pa.timestamp("s")),
pa.field("image_nsfw", pa.float32()),
pa.field("prompt_nsfw", pa.float32()),
pa.field("vector", pa.list_(pa.float32(), 512)),
pa.field("image", pa.binary()),
]
)
def pil_to_bytes(img) -> list[bytes]:
buf = io.BytesIO()
img.save(buf, format="PNG")
return buf.getvalue()
def generate_clip_embeddings(batch) -> pa.RecordBatch:
image = processor(text=None, images=batch["image"], return_tensors="pt")[
"pixel_values"
].to(device)
img_emb = model.get_image_features(image)
batch["vector"] = img_emb.cpu().tolist()
with Pool() as p:
batch["image_bytes"] = p.map(pil_to_bytes, batch["image"])
return batch
def datagen(args):
"""Generate DiffusionDB dataset, and use CLIP model to generate image embeddings."""
dataset = load_dataset("poloclub/diffusiondb", args.subset)
data = []
for b in dataset.map(
generate_clip_embeddings, batched=True, batch_size=256, remove_columns=["image"]
)["train"]:
b["image"] = b["image_bytes"]
del b["image_bytes"]
data.append(b)
tbl = pa.Table.from_pylist(data, schema=schema)
return tbl
def main():
parser = ArgumentParser()
parser.add_argument(
"-o", "--output", metavar="DIR", help="Output lance directory", required=True
)
parser.add_argument(
"-s",
"--subset",
choices=["2m_all", "2m_first_10k", "2m_first_100k"],
default="2m_first_10k",
help="subset of the hg dataset",
)
args = parser.parse_args()
batches = datagen(args)
lance.write_dataset(batches, args.output)
if __name__ == "__main__":
main()