Add application file

Dockerfile

@@ -0,0 +1,41 @@
+FROM python:3.8.10
+
+WORKDIR /content
+RUN apt-get update -y && apt-get upgrade -y && apt-get install -y sudo && apt-get install -y python3-pip && pip3 install --upgrade pip
+RUN apt-get install -y gnupg wget htop sudo git git-lfs software-properties-common build-essential cmake curl
+RUN apt-get install -y ffmpeg libavcodec-dev libavformat-dev libavdevice-dev libgl1 libgtk2.0-0 jq libdc1394-22-dev libraw1394-dev libopenblas-base
+
+
+RUN pip3 install pandas scipy matplotlib torch torchvision torchaudio gradio altair imageio-ffmpeg pocketsphinx jq "numpy==1.23.1"
+
+
+RUN sudo apt remove cmake
+
+RUN sudo apt-get install build-essential libssl-dev
+RUN wget https://github.com/Kitware/CMake/releases/download/v3.20.0/cmake-3.20.0.tar.gz
+RUN tar -zxvf cmake-3.20.0.tar.gz
+RUN cd cmake-3.20.0
+RUN ./bootstrap
+RUN make
+RUN sudo make install
+
+RUN cmake --version 
+
+
+
+RUN git lfs install
+RUN git clone https://huggingface.co/camenduru/pocketsphinx-20.04-t4 pocketsphinx && cd pocketsphinx && cmake -S . -B build && cmake --build build --target install
+
+RUN git clone https://huggingface.co/camenduru/one-shot-talking-face-20.04-t4 one-shot-talking-face && cd one-shot-talking-face && pip install -r requirements.txt && chmod 755 OpenFace/FeatureExtraction
+RUN mkdir /content/out
+
+COPY app.py /content/app.py
+
+RUN git clone https://github.com/TencentARC/GFPGAN.git && cd GFPGAN && pip install basicsr && pip install facexlib && pip install -r requirements.txt && python setup.py develop && pip install realesrgan
+
+RUN git clone https://github.com/chi0tzp/PyVideoFramesExtractor && cd PyVideoFramesExtractor && pip install -r requirements.txt
+
+
+EXPOSE 7860
+
+CMD ["python3", "app.py"]

README.md

@@ -1,8 +1,8 @@
 ---
 title: One Shot Talking Face From Text
-emoji: 📈
-colorFrom: pink
-colorTo: indigo
+emoji: 🐠
+colorFrom: yellow
+colorTo: pink
 sdk: docker
 pinned: false
 ---

app.py

@@ -0,0 +1,220 @@
+import gradio as gr
+# import os, subprocess, torchaudio
+# import torch
+from PIL import Image
+from gtts import gTTS
+import tempfile
+from pydub import AudioSegment
+from pydub.generators import Sine
+# from fairseq.checkpoint_utils import load_model_ensemble_and_task_from_hf_hub
+# from fairseq.models.text_to_speech.hub_interface import TTSHubInterface
+import soundfile
+
+import dlib
+import cv2
+import imageio
+import os
+import gradio as gr
+import os, subprocess, torchaudio
+from PIL import Image
+import ffmpeg
+
+
+
+block = gr.Blocks()
+
+def merge_frames():
+  path = '/content/video_results/restored_imgs'
+  image_folder = os.fsencode(path)
+  print(image_folder)
+  filenames = []
+
+  for file in os.listdir(image_folder):
+      filename = os.fsdecode(file)
+      if filename.endswith( ('.jpg', '.png', '.gif') ):
+          filenames.append(filename)
+
+  filenames.sort() # this iteration technique has no built in order, so sort the frames
+  images = list(map(lambda filename: imageio.imread("/content/video_results/restored_imgs/"+filename), filenames))
+
+  
+  imageio.mimsave('/content/video_output.mp4', images, fps=25.0) # modify the frame duration as needed
+
+
+block = gr.Blocks()
+
+
+
+def audio_video():
+  
+  input_video = ffmpeg.input('/content/video_output.mp4')
+
+  input_audio = ffmpeg.input('/content/audio.wav')
+
+  ffmpeg.concat(input_video, input_audio, v=1, a=1).output('final_output.mp4').run()
+
+
+def compute_aspect_preserved_bbox(bbox, increase_area, h, w):
+    left, top, right, bot = bbox
+    width = right - left
+    height = bot - top
+
+    width_increase = max(increase_area, ((1 + 2 * increase_area) * height - width) / (2 * width))
+    height_increase = max(increase_area, ((1 + 2 * increase_area) * width - height) / (2 * height))
+
+    left_t = int(left - width_increase * width)
+    top_t = int(top - height_increase * height)
+    right_t = int(right + width_increase * width)
+    bot_t = int(bot + height_increase * height)
+
+    left_oob = -min(0, left_t)
+    right_oob = right - min(right_t, w)
+    top_oob = -min(0, top_t)
+    bot_oob = bot - min(bot_t, h)
+
+    if max(left_oob, right_oob, top_oob, bot_oob) > 0:
+        max_w = max(left_oob, right_oob)
+        max_h = max(top_oob, bot_oob)
+        if max_w > max_h:
+            return left_t + max_w, top_t + max_w, right_t - max_w, bot_t - max_w
+        else:
+            return left_t + max_h, top_t + max_h, right_t - max_h, bot_t - max_h
+
+    else:
+        return (left_t, top_t, right_t, bot_t)
+
+def crop_src_image(src_img, detector=None):
+    if  detector is None:
+        detector = dlib.get_frontal_face_detector()
+    save_img='/content/image_pre.png'
+    img = cv2.imread(src_img)
+    faces = detector(img, 0)
+    h, width, _ = img.shape
+    if len(faces) > 0:
+        bbox = [faces[0].left(), faces[0].top(),faces[0].right(), faces[0].bottom()]
+        l = bbox[3]-bbox[1]
+        bbox[1]= bbox[1]-l*0.1
+        bbox[3]= bbox[3]-l*0.1
+        bbox[1] = max(0,bbox[1])
+        bbox[3] = min(h,bbox[3])
+        bbox = compute_aspect_preserved_bbox(tuple(bbox), 0.5, img.shape[0], img.shape[1])
+        img = img[bbox[1] :bbox[3] , bbox[0]:bbox[2]]
+        img = cv2.resize(img, (256, 256))
+        cv2.imwrite(save_img,img)
+    else:
+        img = cv2.resize(img,(256,256))
+        cv2.imwrite(save_img, img)
+
+
+
+def pad_image(image):
+    w, h = image.size
+    if w == h:
+        return image
+    elif w > h:
+        new_image = Image.new(image.mode, (w, w), (0, 0, 0))
+        new_image.paste(image, (0, (w - h) // 2))
+        return new_image
+    else:
+        new_image = Image.new(image.mode, (h, h), (0, 0, 0))
+        new_image.paste(image, ((h - w) // 2, 0))
+        return new_image
+
+def calculate(image_in, audio_in):
+    waveform, sample_rate = torchaudio.load(audio_in)
+    torchaudio.save("/content/audio.wav", waveform, sample_rate, encoding="PCM_S", bits_per_sample=16)
+    image = Image.open(image_in)
+    image = pad_image(image)
+    image.save("image.png")
+
+    pocketsphinx_run = subprocess.run(['pocketsphinx', '-phone_align', 'yes', 'single', '/content/audio.wav'], check=True, capture_output=True)
+    jq_run = subprocess.run(['jq', '[.w[]|{word: (.t | ascii_upcase | sub("<S>"; "sil") | sub("<SIL>"; "sil") | sub("\\\(2\\\)"; "") | sub("\\\(3\\\)"; "") | sub("\\\(4\\\)"; "") | sub("\\\[SPEECH\\\]"; "SIL") | sub("\\\[NOISE\\\]"; "SIL")), phones: [.w[]|{ph: .t | sub("\\\+SPN\\\+"; "SIL") | sub("\\\+NSN\\\+"; "SIL"), bg: (.b*100)|floor, ed: (.b*100+.d*100)|floor}]}]'], input=pocketsphinx_run.stdout, capture_output=True)
+    with open("test.json", "w") as f:
+        f.write(jq_run.stdout.decode('utf-8').strip())
+
+    os.system(f"cd /content/one-shot-talking-face && python3 -B test_script.py --img_path /content/results/restored_imgs/image_pre.png --audio_path /content/audio.wav --phoneme_path /content/test.json --save_dir /content/train")
+    return "/content/train/image_audio.mp4"
+
+
+def one_shot_talking(image_in,audio_in):
+
+  #Pre-processing of image
+  crop_src_image(image_in)
+
+  #Improve quality of input image
+  !python /content/GFPGAN/inference_gfpgan.py --upscale 2 -i /content/image_pre.png -o /content/results  --bg_upsampler realesrgan
+
+  image_in_one_shot='/content/results/restored_imgs/image_pre.png'
+
+  #One Shot Talking Face algorithm
+  calculate(image_in_one_shot,audio_in)
+
+  #Video Quality Improvement
+
+  #1. Extract the frames from the video file using PyVideoFramesExtractor
+  !python /content/PyVideoFramesExtractor/extract.py --video=/content/train/image_pre_audio.mp4
+
+  #2. Improve image quality using GFPGAN on each frames
+  !python /content/GFPGAN/inference_gfpgan.py --upscale 2 -i /content/extracted_frames/image_pre_audio_frames -o /content/video_results  --bg_upsampler realesrgan
+
+  #3. Merge all the frames to a one video using imageio
+  merge_frames()
+
+  audio_video()
+  return "Sucessufull"
+
+
+def one_shot(image,input_text,gender): 
+   if gender == 'Female' or gender == 'female':
+     tts = gTTS(input_text)
+     with tempfile.NamedTemporaryFile(suffix='.mp3', delete=False) as f:
+          tts.write_to_fp(f)
+          f.seek(0)
+          sound = AudioSegment.from_file(f.name, format="mp3")
+          sound.export("/content/audio.wav", format="wav")
+     one_shot_talking(image,'audio.wav') 
+       
+   elif gender == 'Male' or gender == 'male':
+      print(gender)
+      models, cfg, task = load_model_ensemble_and_task_from_hf_hub(
+          "Voicemod/fastspeech2-en-male1",
+          arg_overrides={"vocoder": "hifigan", "fp16": False}
+      )
+
+      model = models[0].cuda()
+      TTSHubInterface.update_cfg_with_data_cfg(cfg, task.data_cfg)
+      generator = task.build_generator([model], cfg)
+      # next(model.parameters()).device
+
+      sample = TTSHubInterface.get_model_input(task, input_text)
+      sample["net_input"]["src_tokens"] = sample["net_input"]["src_tokens"].cuda()
+      sample["net_input"]["src_lengths"] = sample["net_input"]["src_lengths"].cuda()
+      sample["speaker"] = sample["speaker"].cuda()
+
+      wav, rate = TTSHubInterface.get_prediction(task, model, generator, sample)
+      # soundfile.write("/content/audio_before.wav", wav, rate)
+      soundfile.write("/content/audio_before.wav", wav.cpu().clone().numpy(), rate)
+      cmd='ffmpeg -i /content/audio_before.wav -filter:a "atempo=0.7" -vn /content/audio.wav'
+      os.system(cmd)
+      one_shot_talking(image,'audio.wav')
+    
+
+
+
+input_value = "Hello How are you?"
+
+
+
+image = gr.Image(show_label=True, type="filepath",label="Input Image")
+input_text=gr.Textbox(lines=3, value=input_value, label="Input Text")
+gender = gr.Radio(["Female","Male"],value="Female",label="Gender")
+output = gr.Video(show_label=True,label="Output")
+
+demo = gr.Interface(
+    one_shot,
+    [image,input_text,gender],
+    [output],
+    title="One Shot Talking Face from Text",
+)
+demo.launch(enable_queue = False)
+

requirements.txt

@@ -0,0 +1,5 @@
+gtts
+soundfile
+fairseq
+huggingface-hub
+g2p_en