def run(input_dir, output_dir, allow_api):
""" Runs inference of the Scene Description Model on images located
in the input_dir. Stores results in a .csv file located in the
output_dir. If allowed, run() sends selected images to content
moderation models of the Amazon Rekognition, Google Vision Cloud
and Microsoft Computer Vision APIs. Images with the top
prediction`s score below the API_THRESHOLD constant are sent to
APIs. APIs` results are also stored in the output_dir.
Parameters
----------
input_dir : str
Path to the dir with images to run inference on.
output_dir : str
Path to the dir where all results will be stored.
allow_api : bool
If True selected images are sent to the APIs.
"""
print("[NSFW DETECTION] Loading model...")
model = predict.load_model(MODEL_PATH)
nsfw_classification = dict()
imgs_for_api = set()
try:
nsfw_classification = predict.classify(model,
input_dir,
image_dim=NETWORK_INPUT_SIZE)
except ValueError:
print("[Error] Input directory does not contain any valid images")
with open(os.path.join(output_dir, "moderate_image.csv"), "w") \
as output_file:
for img in nsfw_classification.keys():
img_name = img[img.rfind('\\') + 1:]
if (allow_api and
max(nsfw_classification[img].values()) < API_THRESHOLD):
imgs_for_api.add(img_name)
img_name = img[img.rfind('\\') + 1:]
output = f"{img_name}"
for cat, score in nsfw_classification[img].items():
output += f",{cat}:{score:.2f}"
output += "\n"
output_file.write(output)
print(f"[NSFW DETECTION] Analyzed {len(nsfw_classification)} images, "
"results are stored in "
f"{os.path.join(output_dir, 'moderate_image.csv')}")
print(f"[NSFW DETECTION] Passing {len(imgs_for_api)} images for"
" further analysis to APIs...")
run_aws_api(imgs_for_api, input_dir, output_dir, "moderate_image")
run_ms_api(imgs_for_api, input_dir, output_dir, "moderate_image")
run_gg_api(imgs_for_api, input_dir, output_dir, "moderate_image")
print("-------------------------------------------------------\n")
def guess_nsfw_info(image_path):
"""
Guess NSFW content from an image with nsfw_model.
:param image_path
"""
try:
model = predict.load_model(NSFW_MODEL)
img_dict = predict.classify(model, image_path)[image_path]
return (
float(img_dict["p**n"]),
float(img_dict["hentai"]),
float(img_dict["sexy"]),
)
except Exception as error:
logger.error(error, exc_info=True)
def image_data_convert(file_list, network, resize_method, include_top):
print('画像から特徴量を抽出中…')
output_features = []
if network == 'nsfw':
nsfw_model = predict.load_model('./nsfw.299x299.h5')
if network == 'inception_resnet_v2':
img_list = []
for j in tqdm(range(len(file_list))):
img_file = file_list[j]
orig_img = Image.open('./thumbnails/' + img_file + '.jpg')
resized_img = resize_image(orig_img, 299, 299, resize_method)
img_list.append(resized_img)
img_list = np.array(img_list)
inputs = Input(shape=(299, 299, 3))
model = InceptionResNetV2(include_top=include_top,
weights='imagenet',
input_tensor=inputs)
output_features = model.predict(img_list)
print("画像から特徴量への変換終了")
elif network == 'mobilenet_v2':
img_list = []
for j in tqdm(range(len(file_list))):
img_file = file_list[j]
orig_img = Image.open('./thumbnails/' + img_file + '.jpg')
resized_img = resize_image(orig_img, 224, 224, resize_method)
img_list.append(resized_img)
img_list = np.array(img_list)
inputs = Input(shape=(224, 224, 3))
model = MobileNetV2(include_top=include_top,
weights='imagenet',
input_tensor=inputs) #一度include_topをtrueにしてテスト
output_features = model.predict(img_list)
print("画像から特徴量への変換終了")
else:
return None
final_out = []
for i in range(len(output_features)):
final_out.append(output_features[i].flatten())
final_out = np.array(final_out)
return final_out
def check_nsfw(image_list):
"""
:param image_list: list of image paths
"""
# Painfully slow importing time
from nsfw_detector import predict
logger.info("Checking for NSFW content")
for image in image_list:
try:
model = predict.load_model(NSFW_MODEL)
img_dict = predict.classify(model, image)[image]
nsfw_tuple = (
float(img_dict["p**n"]),
float(img_dict["hentai"]),
float(img_dict["sexy"]),
)
except Exception as error:
logger.error(error, exc_info=True)
raise NSFWContent("Error guessing NSFW")
logger.info(nsfw_tuple)
if any(guessed > 0.2 for guessed in nsfw_tuple):
raise NSFWContent("NSFW guessed from %s: %s", image, nsfw_tuple)
from nsfw_detector import predict
import os, fcntl
import time
import shutil
import random
import requests
from flask import Flask, render_template, request
import base64
model = predict.load_model('/Imgfit/ImageReview-v3/nsfw_model/nsfw.299x299.h5')
img_root = './images'
def getRandomSet(bits):
num_set = [chr(i) for i in range(48, 58)]
char_set = [chr(i) for i in range(97, 123)]
total_set = num_set + char_set
value_set = "".join(random.sample(total_set, bits))
return value_set
def predict_img(imgpath):
name_rand = imgpath.split('/')[2].split('.')[0] + '.txt'
print('outname', name_rand)
try:
print('start')
time.sleep(float(random.randint(0, 5) / 10))
res = predict.classify(model, imgpath)
print(res) # save
res2 = res[imgpath]
def image_to_nsfw_score(file_list,
batch_size=100,
use_nsfw_score=True,
meta_path='view_data',
image_path='thumbnails'):
output_features = []
pd_output_features = {}
pd_output_features['nsfw_score_hentai'] = []
pd_output_features['nsfw_score_porn'] = []
pd_output_features['nsfw_score_sexy'] = []
if use_nsfw_score:
nsfw_model = predict.load_model('./nsfw_mobilenet2.224x224.h5')
for i in range(-(-len(file_list) // batch_size)):
print('画像のNSFWスコアを計算中...{0}/{1}'.format(
i + 1, -(-len(file_list) // batch_size)))
if i == -(-len(file_list) // batch_size) - 1:
batch = file_list[batch_size * i:]
else:
batch = file_list[batch_size * i:batch_size * i + batch_size]
image_paths = [
image_path + '/' + video_id + '.jpg' for video_id in batch
]
video_ids = [video_id for video_id in batch]
#既にデータファイル側に記録されていればそれを使う
data_used_nsfw = 0
for video_id in batch:
with open(meta_path + '/{}'.format(video_id) + '.json',
mode='r',
encoding='utf-8') as f:
json_data = json.load(f)
if 'hentai' in json_data:
data_used_nsfw += 1
if data_used_nsfw == len(batch):
nsfw_score_dict = {}
for video_id in batch:
nsfw_score_dict[image_path + '/' + video_id + '.jpg'] = {
'hentai': json_data['hentai'],
'p**n': json_data['p**n'],
'sexy': json_data['sexy']
}
else:
nsfw_score_dict = predict.classify(
nsfw_model, image_paths) #GantMan氏のモデルを使用。
one_batch_features = []
#時間だけが異なるデータが重複しているのでその対策
duplicating = {}
for j in batch:
duplicating[j] = batch.count(j)
for key in nsfw_score_dict:
video_id = video_ids[image_paths.index(key)]
one_video_feature = []
#後で利用できるようにjsonファイルを更新しておく
with open(meta_path + '/{}'.format(video_id) + '.json',
mode='r',
encoding='utf-8') as f:
json_data = json.load(f)
if 'hentai' not in json_data:
json_data['hentai'] = nsfw_score_dict[key]['hentai']
if 'p**n' not in json_data:
json_data['p**n'] = nsfw_score_dict[key]['p**n']
if 'sexy' not in json_data:
json_data['sexy'] = nsfw_score_dict[key]['sexy']
with open(meta_path + '/{}'.format(video_id) + '.json',
mode='w',
encoding='utf-8') as f:
json.dump(json_data, f, ensure_ascii=False, indent=4)
#それぞれ0~1の間でその種類に属する確率を表している数値。
one_video_feature.append(
nsfw_score_dict[key]['hentai']) #性的なイラスト
one_video_feature.append(nsfw_score_dict[key]['p**n']) #実写のポルノ
one_video_feature.append(
nsfw_score_dict[key]['sexy']) #ポルノというほどではないがやや性的な画像
#そのほか、'drawings'（健全なアニメ等のイラスト）, 'neutral'（普通の実写画像）というNSFWではない要素もあるが、
#今回は必要ないので無視。
for k in range(duplicating[video_id]):
one_batch_features.append(one_video_feature)
pd_output_features['nsfw_score_hentai'].append(
nsfw_score_dict[key]['hentai'])
pd_output_features['nsfw_score_porn'].append(
nsfw_score_dict[key]['p**n'])
pd_output_features['nsfw_score_sexy'].append(
nsfw_score_dict[key]['sexy'])
output_features.extend(one_batch_features)
output_features = np.array(output_features)
return output_features, pd_output_features