GitHub

use Hugging Face TFCLIPModel to calculate CLIP score.

clip_model.tflite, input_ids.bin, attention_mask.bin, and pixel_values.bin are generated with the test_clip_model.ipynb
bazel build --config android_arm64 clip_score to build android arm64 binary (with ${ANDROID_NDK_HOME}, ${ANDROID_NDK_VERSION} and ${ANDROID_NDK_API_LEVEL} set correctly)

Name		Name	Last commit message	Last commit date
Latest commit History 12 Commits
.bazelrc		.bazelrc
.bazelversion		.bazelversion
BUILD		BUILD
README.md		README.md
WORKSPACE		WORKSPACE
build_def.bzl		build_def.bzl
calculate_val_set_clip_score.py		calculate_val_set_clip_score.py
clip_score.cc		clip_score.cc
gen_token_ids_and_clip_scores.py		gen_token_ids_and_clip_scores.py
open_clip_scores.tsv		open_clip_scores.tsv
read_tfrecord.cc		read_tfrecord.cc
test_clip_model.ipynb		test_clip_model.ipynb
val_clip_scores.tsv		val_clip_scores.tsv
val_toke_ids_and_clip_scores.tfrecord		val_toke_ids_and_clip_scores.tfrecord
val_toke_ids_and_clip_scores.tsv		val_toke_ids_and_clip_scores.tsv
verify_decoder_fp16.ipynb		verify_decoder_fp16.ipynb
verify_text_encoder_fp16.ipynb		verify_text_encoder_fp16.ipynb
verify_text_encoder_fp32.ipynb		verify_text_encoder_fp32.ipynb
verify_unet_fp16.ipynb		verify_unet_fp16.ipynb

Provide feedback