Upload IR_benchmark/run_for_laion.py with huggingface_hub

IR_benchmark/run_for_laion.py

@@ -0,0 +1,75 @@
+# CUDA_VISIBLE_DEVICES=1 python run_for_laion.py
+import os, sys 
+from src import Scorer
+from PIL import Image, ImageOps 
+import pyarrow.parquet as pq 
+import requests 
+from io import BytesIO
+import pandas as pd 
+import tarfile
+from tqdm import tqdm 
+import json
+
+# [0]: dataset 
+# ====================================================================
+def iter_tar_images(tar):
+    buffer = {}
+    for member in tar:
+        base, ext = os.path.splitext(member.name)
+        ext = ext.lower()
+        if base not in buffer:
+            buffer[base] = {}
+            
+        if ext == ".json":
+            f = tar.extractfile(member)
+            if f is None:
+                continue
+            meta = json.load(f)
+            buffer[base]["meta"] = meta
+            buffer[base]["meta_file"] = member.name
+        elif member.name.lower().endswith((".jpg", ".jpeg", ".png")):
+            f = tar.extractfile(member)
+            if f is None:
+                continue           
+            img = Image.open(f).convert("RGB")
+            img = ImageOps.exif_transpose(img)
+            buffer[base]["img"] = img 
+            buffer[base]["img_file"] = member.name
+
+        if base in buffer and "img" in buffer[base] and "meta" in buffer[base]:
+            yield buffer[base]["img"], buffer[base]["meta"], buffer[base]["img_file"], buffer[base]["meta_file"], base
+            del buffer[base]
+
+subset = "00000"
+data_rt = f"/scratch1/datasets/LAION/{subset}"
+filename = [
+    f  #     os.path.splitext(f)[0]
+    for f in os.listdir(data_rt)
+    if os.path.isfile(os.path.join(data_rt, f)) and ".tar" in f
+]
+filename = sorted(filename)
+
+save_rt = f"/scratch1/users/hsiang.chen/IR_Benchmark/data/LAION_label/{subset}"
+os.makedirs(save_rt, exist_ok=True)
+scorer = Scorer()
+for idx, tar_file in enumerate(filename):
+    results = []
+    ct = 0
+    tar = tarfile.open(os.path.join(data_rt, tar_file), "r")
+    print(f"data:{idx}/{len(filename)}, filename: {tar_file}, # of files: {len(tar.getmembers())}")
+    for img, meta, img_file, meta_file, filename in tqdm(iter_tar_images(tar)):
+        img_list = [img] # can be a list of multiple PIL images
+        output_score = scorer(img_list).tolist()
+        # print(img, meta, filename, output_score)
+        meta["depa_score"] = output_score[0]
+        meta["source"] = os.path.join(data_rt, tar_file)
+        meta["img_file"] = img_file
+        meta["meta_file"] = meta_file
+        results.append(meta)
+        ct += 1
+    tar.close()
+
+    splitname, splitext = os.path.splitext(tar_file)
+    df = pd.DataFrame(results)
+    df.to_parquet(os.path.join(save_rt, f"{splitname}.parquet"))
+