|
import pyarrow.parquet as pq |
|
import pyarrow as pa |
|
from PIL import Image |
|
import io |
|
import pandas as pd |
|
|
|
def is_image_corrupt(image_dict): |
|
try: |
|
image_bytes = image_dict['bytes'] |
|
image = Image.open(io.BytesIO(image_bytes)) |
|
image.verify() |
|
return False |
|
except (IOError, SyntaxError, KeyError) as e: |
|
return True |
|
|
|
def process_parquet_file(file_path, output_file_path): |
|
table = pq.read_table(file_path) |
|
|
|
df = table.to_pandas() |
|
|
|
image_column = 'image' |
|
|
|
clean_indices = [] |
|
corrupt_indices = [] |
|
|
|
for i, image_dict in enumerate(df[image_column]): |
|
if is_image_corrupt(image_dict): |
|
corrupt_indices.append(i) |
|
else: |
|
clean_indices.append(i) |
|
|
|
clean_df = df.iloc[clean_indices] |
|
|
|
clean_table = pa.Table.from_pandas(clean_df) |
|
|
|
pq.write_table(clean_table, output_file_path) |
|
|
|
print(f"File {file_path}: {len(corrupt_indices)} corrupt images were removed.") |
|
|
|
for i in range(5): |
|
input_file_path = f'data/train-0000{i}-of-00005.parquet' |
|
output_file_path = f'data/train-0000{i}-of-00005.parquet' |
|
process_parquet_file(input_file_path, output_file_path) |
|
|