konsumer
commented
3 years ago It's much better to use your own Google Drive folder to hold code & output. It's faster (after initial download) and you can browse the files on Drive.
Here is some fancy setup-code that will mount your google drive and collect the code & models, and wrap everything up so it's easy to work with:
# this is where you are keeping things, on your google-drive
# path to code
PATH = "MyDrive/FALdetector"
# where to store output files
OUT = "MyDrive/FALdetector/out"
# where to look for input files (if applicable)
IN = "MyDrive/FALdetector/in"
from google.colab import files, drive
from os import path, system, remove, chdir, getcwd, mkdir
import subprocess
from shutil import unpack_archive, rmtree, move
import requests
from urllib.parse import urlparse
import site
from IPython.display import display, Image as IM
def download(url, fname=None):
if fname == None:
a = urlparse(url)
fname = path.basename(a.path)
print(f"Downloading {url}")
# clean up previous download
try:
remove(fname)
except:
pass
with open(fname, "wb") as handle:
response = requests.get(url, stream=True)
if not response.ok:
return print(response)
for block in response.iter_content(1024):
if not block:
break
handle.write(block)
return fname
# show an array of images
def show_images(images):
for image in images:
display(IM(image))
print("Setting things up...")
# you will need to auth to your own google drive:
drive.mount("/content/drive")
full_path = path.join("/content/drive", PATH)
full_out = path.join("/content/drive", OUT)
full_in = path.join("/content/drive", IN)
# go to root-dir
chdir("/content")
# grab the repo
if (not path.exists(full_path)):
f = download("https://github.com/PeterWang512/FALdetector/archive/refs/heads/master.zip")
unpack_archive(f)
move("FALdetector-master", full_path)
remove(f)
# download models
chdir(path.join(full_path, "weights"))
if not path.exists("global.pth"):
download("https://www.dropbox.com/s/rb8zpvrbxbbutxc/global.pth?dl=1")
if not path.exists("local.pth"):
download("https://www.dropbox.com/s/pby9dhpr6cqziyl/local.pth?dl=1")
# CLASSIFY
# they are structured a bit funny for import, so I have to do some extra stuff
chdir(full_path)
import global_classifier
import numpy as np
import torch
import torchvision.transforms as transforms
from PIL import Image
from networks.drn_seg import DRNSeg
from utils.tools import *
from utils.visualize import *
def global_classify(image, no_crop=False):
model = global_classifier.load_classifier(path.join(full_path, "weights", "global.pth"), 0)
return global_classifier.classify_fake(model, image, no_crop)
def local_classify(fname, crop=False, dest_folder=full_out):
img_path = fname
model_path = path.join(full_path, "weights", "local.pth")
no_crop = not crop
device = 'cuda:{}'.format(0)
model = DRNSeg(2)
state_dict = torch.load(model_path, map_location=device)
model.load_state_dict(state_dict['model'])
model.to(device)
model.eval()
tf = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(
mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
im_w, im_h = Image.open(img_path).size
if no_crop:
face = Image.open(img_path).convert('RGB')
else:
faces = face_detection(img_path, verbose=False)
if len(faces) == 0:
print("no face detected by dlib, exiting")
sys.exit()
face, box = faces[0]
face = resize_shorter_side(face, 400)[0]
face_tens = tf(face).to(device)
# Warping field prediction
with torch.no_grad():
flow = model(face_tens.unsqueeze(0))[0].cpu().numpy()
flow = np.transpose(flow, (1, 2, 0))
h, w, _ = flow.shape
# Undoing the warps
modified = face.resize((w, h), Image.BICUBIC)
modified_np = np.asarray(modified)
reverse_np = warp(modified_np, flow)
reverse = Image.fromarray(reverse_np)
finname = path.splitext(path.basename(fname))[0]
finput = os.path.join(dest_folder, f'{finname}-cropped_input.jpg')
fwarped = os.path.join(dest_folder, f'{finname}-warped.jpg')
fheat = os.path.join(dest_folder, f'{finname}-heatmap.jpg')
flow_magn = np.sqrt(flow[:, :, 0]**2 + flow[:, :, 1]**2)
# Saving the results
modified.save(finput, quality=90)
reverse.save(fwarped, quality=90)
save_heatmap_cv(modified_np, flow_magn, fheat)
return (finput, fwarped, fheat)Put this at the bottom of the same code-block to ask to upload files:
chdir(full_out)
print("Please upload the image(s) you'd like to analyze: ")
images = files.upload()
# output analysis
for image in images:
print(f"Global: {image}")
full_image = path.join(full_out, image)
prob = global_classify(full_image)
print("Probibility being modified by Photoshop FAL: {:.2f}%".format(prob*100))
print(f"Local: {image}")
show_images(local_classify(full_image))Put this at the bottom instead, to download a list of URLs:
# INPUT URLS
chdir(full_out)
# put yours in here
urls = [
"https://organicthemes.com/demo/profile/files/2018/05/profile-pic.jpg"
]
for url in urls:
image = download(url)
print(f"Global: {image}")
full_image = path.join(full_out, image)
prob = global_classify(full_image)
print("Probibility being modified by Photoshop FAL: {:.2f}%".format(prob*100))
print(f"Local: {image}")
show_images(local_classify(full_image))Put this at the bottom instead, to find all the images in your google drive:
from glob import glob
for image in glob(path.join(full_in, '*')):
print(f"Global: {image}")
full_image = path.join(full_out, image)
prob = global_classify(full_image)
print("Probibility being modified by Photoshop FAL: {:.2f}%".format(prob*100))
print(f"Local: {image}")
show_images(local_classify(full_image))IN (at the top) should be a flat-directory, with no sub-dirs, filled with images.
On all 3, you can remove the show_images() part to just output to Drive (in OUT dir, in Drive) and not show them:
local_classify(full_image)
Jean-Pierre-Polnareff
LinhVi
zahra12638
I don't know if it's helpful to anyone, but I made a google-collab notebook, so you can play around with it very easily, for free:
https://colab.research.google.com/drive/1AQ0XSKWjzJBhGXXJ0XrA4DckFdv6Ul5N?usp=sharing#scrollTo=2eYISfxy-8qe
Works great!