File size: 4,623 Bytes
6490f58 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 | # /// script
# requires-python = ">=3.11"
# dependencies = [
# "coreai-core==1.0.0b3",
# "coreai-torch==0.4.3",
# "transformers==4.57.3",
# "torch<=2.13.0",
# ]
#
# [tool.uv]
# index-url = "https://pypi.org/simple"
# index-strategy = "unsafe-best-match"
# ///
"""CLIP ViT-B/32 as one .aimodel with separate text and image encoders.
Apple's models/clip recipe traces one function that needs an image and a fixed
batch of 3 prompts on every call. Embedding text or images alone needs each
tower on its own, so this exports:
- encode_text(input_ids, attention_mask) -> text_embeds, inputs (B, 77) int32,
padded to CLIP's 77-token context
- encode_image(pixel_values) -> image_embeds, input (B, 3, 224, 224)
Both batch dims are dynamic (1-64), and the embeddings are L2-normalized.
"""
import argparse
import shutil
import time
from pathlib import Path
import torch
import transformers
from coreai.runtime import AIModelAssetMetadata
from coreai_torch import TorchConverter, get_decomp_table
MODEL = "openai/clip-vit-base-patch32"
class TextEncoder(torch.nn.Module):
def __init__(self, model):
super().__init__()
self.text_model = model.text_model
self.text_projection = model.text_projection
def forward(self, input_ids, attention_mask):
pooled = self.text_model(input_ids=input_ids, attention_mask=attention_mask)[1]
embeds = self.text_projection(pooled)
return embeds / embeds.norm(p=2, dim=-1, keepdim=True)
class ImageEncoder(torch.nn.Module):
def __init__(self, model):
super().__init__()
self.vision_model = model.vision_model
self.visual_projection = model.visual_projection
def forward(self, pixel_values):
pooled = self.vision_model(pixel_values=pixel_values)[1]
embeds = self.visual_projection(pooled)
return embeds / embeds.norm(p=2, dim=-1, keepdim=True)
def export(module, kwargs, dynamic_shapes, dtype):
with torch.autocast(device_type="cpu", dtype=dtype):
exported = torch.export.export(module, args=(), kwargs=kwargs, dynamic_shapes=dynamic_shapes)
return exported.run_decompositions(get_decomp_table())
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--output-dir", required=True)
parser.add_argument("--dtype", choices=["float16", "float32"], default="float16")
args = parser.parse_args()
dtype = getattr(torch, args.dtype)
model = transformers.CLIPModel.from_pretrained(MODEL).eval().to(dtype)
tokenizer = transformers.CLIPTokenizer.from_pretrained(MODEL)
text = tokenizer(
["a photo of a cat", "a photo of a dog"],
return_tensors="pt",
padding="max_length",
max_length=77,
)
batch = torch.export.Dim("batch", min=1, max=64)
text_program = export(
TextEncoder(model),
{
"input_ids": text["input_ids"].to(torch.int32),
"attention_mask": text["attention_mask"].to(torch.int32),
},
{"input_ids": {0: batch}, "attention_mask": {0: batch}},
dtype,
)
image_program = export(
ImageEncoder(model),
{"pixel_values": torch.randn(2, 3, 224, 224).to(dtype)},
{"pixel_values": {0: batch}},
dtype,
)
converter = (
TorchConverter(mode=TorchConverter.Mode.RELEASE)
.add_exported_program(
exported_program=text_program,
input_names=["input_ids", "attention_mask"],
output_names=["text_embeds"],
entrypoint_name="encode_text",
)
.add_exported_program(
exported_program=image_program,
input_names=["pixel_values"],
output_names=["image_embeds"],
entrypoint_name="encode_image",
)
)
program = converter.to_coreai()
metadata = AIModelAssetMetadata()
metadata.author = "A. Radford et al."
metadata.license = "MIT"
metadata.model_description = (
"CLIP ViT-B/32 text and image encoders with L2-normalized embeddings. "
"Source: https://huggingface.co/openai/clip-vit-base-patch32"
)
metadata.creation_date = int(time.time())
path = Path(args.output_dir) / f"clip-vit-base-patch32-encoders_{args.dtype}.aimodel"
if path.exists():
shutil.rmtree(path)
path.parent.mkdir(parents=True, exist_ok=True)
program.save_asset(path, metadata)
# The tokenizer travels with the model, so the app needs nothing else
tokenizer.save_pretrained(path.parent / "clip-tokenizer")
print(f"[INFO] Saved {path}")
if __name__ == "__main__":
main()
|