inference.py

from os import listdir, path
import numpy as np
import scipy, cv2, os, sys, argparse, audio
import dlib, json, h5py, subprocess
from tqdm import tqdm

from keras.callbacks import ModelCheckpoint
from keras.models import Model

args = parser.parse_args()
args.img_size = 96

if args.face.split('.')[1] in ['jpg', 'png', 'jpeg']:
	args.static = True

def rect_to_bb(d):
	x = d.rect.left()
	y = d.rect.top()
	w = d.rect.right() - x
	h = d.rect.bottom() - y
	return (x, y, w, h)

def calcMaxArea(rects):
	max_cords = (-1,-1,-1,-1)
	max_area = 0
	max_rect = None
	for i in range(len(rects)):
		cur_rect = rects[i]
		(x,y,w,h) = rect_to_bb(cur_rect)
		if w*h > max_area:
			max_area = w*h
			max_cords = (x,y,w,h)
			max_rect = cur_rect	
	return max_cords, max_rect
	
def face_detect(images):
	detector = dlib.cnn_face_detection_model_v1(args.face_det_checkpoint)

	batch_size = args.face_det_batch_size

	predictions = []
	for i in tqdm(range(0, len(images), batch_size)):
		predictions.extend(detector(images[i:i + batch_size]))
	
	results = []
	pady1, pady2, padx1, padx2 = args.pads
	for rects, image in zip(predictions, images):
		(x, y, w, h), max_rect = calcMaxArea(rects)
		if x == -1:
			results.append([None, (-1,-1,-1,-1), False])
			continue
		y1 = max(0, y + pady1)
		y2 = min(image.shape[0], y + h + pady2)
		x1 = max(0, x + padx1)
		x2 = min(image.shape[1], x + w + padx2)
		face = image[y1:y2, x1:x2, ::-1] # RGB ---> BGR

		results.append([face, (y1, y2, x1, x2), True])
	
	del detector
	return results 

def datagen(frames, mels):
	img_batch, mel_batch, frame_batch, coords_batch = [], [], [], []

	if not args.static:
		face_det_results = face_detect([f[...,::-1] for f in frames]) # BGR2RGB for CNN face detection
	else:
		face_det_results = face_detect([frames[0][...,::-1]])

	for i, m in enumerate(mels):
		idx = 0 if args.static else i%len(frames)
		frame_to_save = frames[idx].copy()
		face, coords, valid_frame = face_det_results[idx].copy()
		if not valid_frame:
			print ("Face not detected, skipping frame {}".format(i))
			continue

		face = cv2.resize(face, (args.img_size, args.img_size))

		img_batch.append(face)
		mel_batch.append(m)
		frame_batch.append(frame_to_save)
		coords_batch.append(coords)

		if len(img_batch) >= args.lipgan_batch_size:
			img_batch, mel_batch = np.asarray(img_batch), np.asarray(mel_batch)

			img_masked = img_batch.copy()
			img_masked[:, args.img_size//2:] = 0

			img_batch = np.concatenate((img_batch, img_masked), axis=3) / 255.
			mel_batch = np.reshape(mel_batch, [len(mel_batch), mel_batch.shape[1], mel_batch.shape[2], 1])

			yield img_batch, mel_batch, frame_batch, coords_batch
			img_batch, mel_batch, frame_batch, coords_batch = [], [], [], []

	if len(img_batch) > 0:
		img_batch, mel_batch = np.asarray(img_batch), np.asarray(mel_batch)

		img_masked = img_batch.copy()
		img_masked[:, args.img_size//2:] = 0

		img_batch = np.concatenate((img_batch, img_masked), axis=3) / 255.
		mel_batch = np.reshape(mel_batch, [len(mel_batch), mel_batch.shape[1], mel_batch.shape[2], 1])

		yield img_batch, mel_batch, frame_batch, coords_batch

fps = args.fps
mel_step_size = 27
mel_idx_multiplier = 80./fps

if args.model == 'residual':
	from generator import create_model_residual as create_model
else:
	from generator import create_model as create_model

def main():
	if args.face.split('.')[1] in ['jpg', 'png', 'jpeg']:
		full_frames = [cv2.imread(args.face)]
	else:
		video_stream = cv2.VideoCapture(args.face)
		
		full_frames = []
		while 1:
			still_reading, frame = video_stream.read()
			if not still_reading:
				video_stream.release()
				break
			full_frames.append(frame)
			if len(full_frames) % 2000 == 0: print(len(full_frames))

			if len(full_frames) * (1./fps) >= args.max_sec: break

		print ("Number of frames available for inference: "+str(len(full_frames)))

	wav = audio.load_wav(args.audio, 16000)
	mel = audio.melspectrogram(wav)
	print(mel.shape)

	if np.isnan(mel.reshape(-1)).sum() > 0:
		raise ValueError('Mel contains nan!')

	mel_chunks = []
	i = 0
	while 1:
		start_idx = int(i * mel_idx_multiplier)
		if start_idx + mel_step_size > len(mel[0]):
			break
		mel_chunks.append(mel[:, start_idx : start_idx + mel_step_size])
		i += 1

	print("Length of mel chunks: {}".format(len(mel_chunks)))

	batch_size = args.lipgan_batch_size
	gen = datagen(full_frames.copy(), mel_chunks)

	for i, (img_batch, mel_batch, frames, coords) in enumerate(tqdm(gen, 
											total=int(np.ceil(float(len(mel_chunks))/batch_size)))):
		if i == 0:
			model = create_model(args, mel_step_size)
			print ("Model Created")

			model.load_weights(args.checkpoint_path)
			print ("Model loaded")

			frame_h, frame_w = full_frames[0].shape[:-1]
			out = cv2.VideoWriter(path.join(args.results_dir, 'result.avi'), 
									cv2.VideoWriter_fourcc(*'DIVX'), fps, (frame_w, frame_h))

		pred = model.predict([img_batch, mel_batch])
		pred = pred * 255
		
		for p, f, c in zip(pred, frames, coords):
			y1, y2, x1, x2 = c
			p = cv2.resize(p, (x2 - x1, y2 - y1))

			f[y1:y2, x1:x2] = p
			out.write(f)

	out.release()

	command = 'ffmpeg -i {} -i {} -strict -2 -q:v 1 {}'.format(args.audio, path.join(args.results_dir, 'result.avi'), 
														path.join(args.results_dir, 'result_voice.avi'))
	subprocess.call(command, shell=True)

if __name__ == '__main__':
	main()