updates

app.py

@@ -1,49 +1,86 @@
 from __future__ import annotations
-from IPython.display import clear_output
-from IPython import get_ipython
+# Downloading files of the server
 import os
-#os.system('pip install -r requirements.txt')
+import requests
+def download_file(url, save_path):
+    response = requests.get(url)
+    with open(save_path, 'wb') as file:
+        file.write(response.content)
+file_names = [
+    'cloee-1.wav',
+    'julian-bedtime-style-1.wav',
+    'julian-bedtime-style-2.wav',
+    'pirate_by_coqui.wav',
+    'thera-1.wav'
+]
+base_url = 'https://raw.githubusercontent.com/ruslanmv/ai-story-server/main/voices/'
+save_folder = 'voices/'
+if not os.path.exists(save_folder):
+    os.makedirs(save_folder)
+for file_name in file_names:
+    url = base_url + file_name
+    save_path = os.path.join(save_folder, file_name)
+    download_file(url, save_path)
+    print(f'Downloaded {file_name}')
+requirements_url = 'https://raw.githubusercontent.com/ruslanmv/ai-story-server/main/requirements.txt'
+save_path = 'requirements.txt'
+download_file(requirements_url, save_path)  
 #os.system('pip install gradio==3.48.0')
+os.system('pip install -r requirements.txt')
 os.system('pip install python-dotenv')
-# In[1]:
-#Use GPU
-import subprocess
-gpu_info = subprocess.getoutput('nvidia-smi')
-if 'failed' in gpu_info:
+os.system('pip install ipython')
+from IPython.display import clear_output
+clear_output()
+import os
+import shutil
+from IPython.display import clear_output
+# Use GPU
+def is_nvidia_smi_available():
+    return shutil.which("nvidia-smi") is not None
+if is_nvidia_smi_available():
+    gpu_info = os.popen("nvidia-smi").read()
+    if gpu_info.find('failed') >= 0:
+        print('Not connected to a GPU')
+        is_gpu = False
+    else:
+        print(gpu_info)
+        is_gpu = True
+else:
+    print('nvidia-smi command not found')
     print('Not connected to a GPU')
     is_gpu = False
-else:
-    print(gpu_info)
-    is_gpu = True
-# In[2]:
-# In[3]:
 import os
 import dotenv
 # Load the environment variables from the .env file
+# You can change the default secret
+with open(".env", "w") as env_file:
+    env_file.write("SECRET_TOKEN=secret")
 dotenv.load_dotenv()
 # Access the value of the SECRET_TOKEN variable
 secret_token = os.getenv("SECRET_TOKEN")
-# In[7]:
 import os
 #download for mecab
+# Check if unidic is installed
 os.system("python -m unidic download")
-# In[5]:
+
 # By using XTTS you agree to CPML license https://coqui.ai/cpml
 os.environ["COQUI_TOS_AGREED"] = "1"
 # NOTE: for streaming will require gradio audio streaming fix
 # pip install --upgrade -y gradio==0.50.2 git+https://github.com/gorkemgoknar/gradio.git@patch-1
 #Now you’re ready to install 🤗 Transformers with the following command:
-#For CPU-support only, Transformers and PyTorch with:
-os.system('pip install transformers[torch]') 
 if not is_gpu:
+  #For CPU-support only, Transformers and PyTorch with:
+  os.system('pip install transformers[tf-cpu]')
+  #os.system('pip install  transformers[torch] accelerate==0.26.1')
   #pip install 'transformers[tf-cpu]'   #Transformers and TensorFlow 2.0:
   os.system('pip install llama-cpp-python==0.2.11')
 else:
-    # we need to compile a CUBLAS version 
+    os.system('pip install transformers[torch]')
+    # we need to compile a CUBLAS version
     # Or get it from  https://jllllll.github.io/llama-cpp-python-cuBLAS-wheels/
     os.system('CMAKE_ARGS="-DLLAMA_CUBLAS=on" pip install llama-cpp-python==0.2.11')
+clear_output()
 
-# In[8]:
 import textwrap
 from scipy.io.wavfile import write
 from pydub import AudioSegment
@@ -52,19 +89,15 @@ import numpy as np
 import torch
 import nltk  # we'll use this to split into sentences
 nltk.download("punkt")
-
 import noisereduce as nr
 import subprocess
 import langid
 import uuid
 import emoji
 import pathlib
-
 import datetime
-
 from scipy.io.wavfile import write
 from pydub import AudioSegment
-
 import re
 import io, wave
 import librosa
@@ -73,23 +106,15 @@ from TTS.api import TTS
 from TTS.tts.configs.xtts_config import XttsConfig
 from TTS.tts.models.xtts import Xtts
 from TTS.utils.generic_utils import get_user_data_dir
-
-
 import gradio as gr
 import os
 import time
-
 import gradio as gr
-from transformers import pipeline
 import numpy as np
-
+from transformers import pipeline
 from gradio_client import Client
 from huggingface_hub import InferenceClient
-
-
-
-# In[9]:
-
+clear_output()
 
 # This will trigger downloading model
 print("Downloading if not downloaded Coqui XTTS V2")
@@ -98,16 +123,13 @@ model_name = "tts_models/multilingual/multi-dataset/xtts_v2"
 ModelManager().download_model(model_name)
 model_path = os.path.join(get_user_data_dir("tts"), model_name.replace("/", "--"))
 print("XTTS downloaded")
-
 if is_gpu:
     use_deepspeed=True
 else:
     use_deepspeed=False
-    
 print("Loading XTTS")
 config = XttsConfig()
 config.load_json(os.path.join(model_path, "config.json"))
-
 model = Xtts.init_from_config(config)
 model.load_checkpoint(
     config,
@@ -116,23 +138,11 @@ model.load_checkpoint(
     eval=True,
     use_deepspeed=use_deepspeed,
 )
-
-#if is_gpu:
-#    model.cuda()
-
 print("Done loading TTS")
-
-
-# In[60]:
-
-
 #####llm_model = os.environ.get("LLM_MODEL", "mistral") # or "zephyr"
-
 title = "Voice chat with Zephyr/Mistral and Coqui XTTS"
-
 DESCRIPTION = """# Voice chat with Zephyr/Mistral and Coqui XTTS"""
 css = """.toast-wrap { display: none !important } """
-
 from huggingface_hub import HfApi
 
 HF_TOKEN = os.environ.get("HF_TOKEN")
@@ -174,40 +184,14 @@ pirate_system_message = f"You as {character_name}. {character_scenario} Print ou
 ROLE_PROMPTS["Pirate"]= pirate_system_message
 ##"You are an AI assistant with Zephyr model by Mistral and Hugging Face and speech from Coqui XTTS . User will you give you a task. Your goal is to complete the task as faithfully as you can. While performing the task think step-by-step and justify your steps, your answers should be clear and short sentences"
 
-# In[49]:
-
-
-
-
-
-# In[15]:
-
-
 ### WILL USE LOCAL MISTRAL OR ZEPHYR
-import os
 from huggingface_hub import hf_hub_download
-
 print("Downloading LLM")
-
-# Get the current directory
-current_dir = os.getcwd()
-# Append the current directory to the zephyr_model_path
-zephyr_model_path = os.path.join(current_dir, "zephyr-7b-beta.Q5_K_M.gguf")
+print("Downloading Zephyr")
+# use new gguf format
+zephyr_model_path = "./zephyr-7b-beta.Q5_K_M.gguf"
 if not os.path.isfile(zephyr_model_path):
-    print("Downloading Zephyr")
-    hf_hub_download(repo_id="TheBloke/zephyr-7B-beta-GGUF", local_dir=current_dir, filename="zephyr-7b-beta.Q5_K_M.gguf")
-else:
-    print("Zephyr it is already downloaded")
-
-
-# In[ ]:
-
-
-
-
-
-# In[16]:
-
+    hf_hub_download(repo_id="TheBloke/zephyr-7B-beta-GGUF", local_dir=".", filename="zephyr-7b-beta.Q5_K_M.gguf")
 
 from llama_cpp import Llama
 # set GPU_LAYERS to 15 if you have a 8GB GPU so both models can fit in
@@ -223,17 +207,14 @@ LLAMA_VERBOSE=False
 
 llm_zephyr = Llama(model_path=zephyr_model_path,
                    n_gpu_layers=GPU_LAYERS,
-                   max_new_tokens=512, 
-                   context_window=4096, 
+                   max_new_tokens=512,
+                   context_window=4096,
                    n_ctx=4096,
                    n_batch=128,
-                   verbose=LLAMA_VERBOSE)
-
+                   )
+llm_zephyr.verbose = LLAMA_VERBOSE
 print("Running LLM Zephyr")
-
-
-# In[17]:
-
+clear_output()
 
 def split_sentences(text, max_len):
     # Apply custom rules to enforce sentence breaks with double punctuation
@@ -251,7 +232,7 @@ def split_sentences(text, max_len):
             sentence_list.extend(wrapped)
         else:
             sentence_list.append(sent)
-    
+
     return sentence_list
 
 
@@ -292,7 +273,7 @@ def pcm_to_wav(pcm_data, sample_rate=24000, channels=1, bit_depth=16):
 
     # Prepare the WAV file headers
     wav_header = struct.pack('<4sI4s', b'RIFF', chunk_size, b'WAVE')  # 'RIFF' chunk descriptor
-    fmt_subchunk = struct.pack('<4sIHHIIHH', 
+    fmt_subchunk = struct.pack('<4sIHHIIHH',
                                b'fmt ', fmt_subchunk_size, 1, channels,
                                sample_rate, sample_rate * channels * bit_depth // 8,
                                channels * bit_depth // 8, bit_depth)
@@ -301,12 +282,8 @@ def pcm_to_wav(pcm_data, sample_rate=24000, channels=1, bit_depth=16):
 
     return wav_header + fmt_subchunk + data_subchunk + pcm_data
 
-
-# In[23]:
-
-
 def generate_local_llm(
-    prompt,    
+    prompt,
     history,
     system_message=None,
     temperature=0.8,
@@ -344,13 +321,13 @@ def generate_local_llm(
 
             if "<|user|>" in character:
                 # end of context
-                return 
-                
+                return
+
             if emoji.is_emoji(character):
                 # Bad emoji not a meaning messes chat from next lines
                 return
-                
-            
+
+
             output += response["choices"][0]["text"].replace("<|assistant|>","").replace("<|user|>","")
             yield output
 
@@ -366,16 +343,6 @@ def generate_local_llm(
 
     return output
 
-
-# In[28]:
-
-
-get_ipython().system('pip install OpenAI')
-
-
-# In[103]:
-
-
 def generate_stream(prompt, model="mixtral-8x7b"):
     base_url = "https://ruslanmv-hf-llm-api.hf.space"
     api_key = "sk-xxxxx"
@@ -436,16 +403,12 @@ def generate_local(
     return output
 
 
-# In[ ]:
-
-
-# In[17]:
 
 
 def get_latents(speaker_wav,voice_cleanup=False):
     if (voice_cleanup):
         try:
-            cleanup_filter="lowpass=8000,highpass=75,areverse,silenceremove=start_periods=1:start_silence=0:start_threshold=0.02,areverse,silenceremove=start_periods=1:start_silence=0:start_threshold=0.02" 
+            cleanup_filter="lowpass=8000,highpass=75,areverse,silenceremove=start_periods=1:start_silence=0:start_threshold=0.02,areverse,silenceremove=start_periods=1:start_silence=0:start_threshold=0.02"
             resample_filter="-ac 1 -ar 22050"
             out_filename = speaker_wav + str(uuid.uuid4()) + ".wav"  #ffmpeg to know output format
             #we will use newer ffmpeg as that has afftn denoise filter
@@ -459,7 +422,7 @@ def get_latents(speaker_wav,voice_cleanup=False):
             print("Error: failed filtering, use original microphone input")
     else:
             speaker_wav=speaker_wav
-            
+
     # create as function as we can populate here with voice cleanup/filtering
     (
         gpt_cond_latent,
@@ -485,15 +448,15 @@ def wave_header_chunk(frame_input=b"", channels=1, sample_width=2, sample_rate=2
 #Config will have more correct languages, they may be added before we append here
 ##["en","es","fr","de","it","pt","pl","tr","ru","nl","cs","ar","zh-cn","ja"]
 
-xtts_supported_languages=config.languages  
+xtts_supported_languages=config.languages
 def detect_language(prompt):
     # Fast language autodetection
     if len(prompt)>15:
         language_predicted=langid.classify(prompt)[0].strip() # strip need as there is space at end!
-        if language_predicted == "zh": 
+        if language_predicted == "zh":
             #we use zh-cn on xtts
             language_predicted = "zh-cn"
-            
+
         if language_predicted not in xtts_supported_languages:
             print(f"Detected a language not supported by xtts :{language_predicted}, switching to english for now")
             gr.Warning(f"Language detected '{language_predicted}' can not be spoken properly 'yet' ")
@@ -508,10 +471,6 @@ def detect_language(prompt):
 
     return language
 
-
-# In[18]:
-
-
 def get_voice_streaming(prompt, language, latent_tuple, suffix="0"):
     gpt_cond_latent, speaker_embedding = latent_tuple
 
@@ -559,16 +518,6 @@ def get_voice_streaming(prompt, language, latent_tuple, suffix="0"):
     except:
         return None
 
-
-# In[ ]:
-
-
-
-
-
-# In[19]:
-
-
 # Will be triggered on text submit (will send to generate_speech)
 def add_text(history, text):
     history = [] if history is None else history
@@ -593,7 +542,7 @@ def add_file(history, file):
 
 
 def get_sentence(history, chatbot_role):
-    
+
     history = [["", None]] if history is None else history
 
     history[-1][1] = ""
@@ -606,14 +555,14 @@ def get_sentence(history, chatbot_role):
     stored_sentence_hash = None
 
     print(chatbot_role)
-    
+
     for character in generate_local(history[-1][0], history[:-1], system_message=ROLE_PROMPTS[chatbot_role]):
         history[-1][1] = character.replace("<|assistant|>","")
         # It is coming word by word
 
         text_to_generate = nltk.sent_tokenize(history[-1][1].replace("\n", " ").replace("<|assistant|>"," ").replace("<|ass>","").replace("[/ASST]","").replace("[/ASSI]","").replace("[/ASS]","").replace("","").strip())
         if len(text_to_generate) > 1:
-            
+
             dif = len(text_to_generate) - len(sentence_list)
 
             if dif == 1 and len(sentence_list) != 0:
@@ -632,22 +581,22 @@ def get_sentence(history, chatbot_role):
                 stored_sentence_hash = None
             else:
                 sentence = text_to_generate[len(sentence_list)]
-                
+
             # too short sentence just append to next one if there is any
-            # this is for proper language detection 
+            # this is for proper language detection
             if len(sentence)<=15 and stored_sentence_hash is None and stored_sentence is None:
                 if sentence[-1] in [".","!","?"]:
                     if stored_sentence_hash != hash(sentence):
                         stored_sentence = sentence
-                        stored_sentence_hash = hash(sentence) 
+                        stored_sentence_hash = hash(sentence)
                         print("Storing:",stored_sentence)
                         continue
-            
-            
+
+
             sentence_hash = hash(sentence)
             if stored_sentence_hash is not None and sentence_hash == stored_sentence_hash:
                 continue
-            
+
             if sentence_hash not in sentence_hash_list:
                 sentence_hash_list.append(sentence_hash)
                 sentence_list.append(sentence)
@@ -663,28 +612,22 @@ def get_sentence(history, chatbot_role):
                 last_sentence = stored_sentence + last_sentence
                 stored_sentence = stored_sentence_hash = None
                 print("Last Sentence with stored:",last_sentence)
-        
+
             sentence_hash_list.append(sentence_hash)
             sentence_list.append(last_sentence)
             print("Last Sentence: ", last_sentence)
-    
+
             yield (last_sentence, history)
     except:
         print("ERROR on last sentence history is :", history)
 
 
-# In[19]:
-
-
 from scipy.io.wavfile import write
 from pydub import AudioSegment
 
 second_of_silence = AudioSegment.silent() # use default
 second_of_silence.export("sil.wav", format='wav')
-
-
-
-# In[20]:
+clear_output()
 
 
 def generate_speech_from_history(history, chatbot_role, sentence):
@@ -692,7 +635,7 @@ def generate_speech_from_history(history, chatbot_role, sentence):
     # total_wav_bytestream = b""
     if len(sentence)==0:
         print("EMPTY SENTENCE")
-        return 
+        return
     # Sometimes prompt </s> coming on output remove it
     # Some post process for speech only
     sentence = sentence.replace("</s>", "")
@@ -714,20 +657,20 @@ def generate_speech_from_history(history, chatbot_role, sentence):
     #if (sentence[-1] in ["!", "?", ".", ","]) or (sentence[-2] in ["!", "?", ".", ","]):
     #    # just add a space
     #    sentence = sentence[:-1] + " " + sentence[-1]
-        
+
     # regex does the job well
     sentence = re.sub("([^\x00-\x7F]|\w)([\.。?!]+)",r"\1 \2",sentence)
-    
+
     print("Sentence for speech:", sentence)
 
     results = []
-    
+
     try:
         if len(sentence) < SENTENCE_SPLIT_LENGTH:
             # no problem continue on
             sentence_list = [sentence]
         else:
-            # Until now nltk likely split sentences properly but we need additional 
+            # Until now nltk likely split sentences properly but we need additional
             # check for longer sentence and split at last possible position
             # Do whatever necessary, first break at hypens then spaces and then even split very long words
             # sentence_list=textwrap.wrap(sentence,SENTENCE_SPLIT_LENGTH)
@@ -738,7 +681,7 @@ def generate_speech_from_history(history, chatbot_role, sentence):
             if any(c.isalnum() for c in sentence):
                 if language=="autodetect":
                     #on first call autodetect, nexts sentence calls will use same language
-                    language = detect_language(sentence) 
+                    language = detect_language(sentence)
                 #exists at least 1 alphanumeric (utf-8)
 
                 #print("Inserting data to get_voice_streaming:")
@@ -766,7 +709,7 @@ def generate_speech_from_history(history, chatbot_role, sentence):
                         continue
 
                 # Filter output for better voice
-                filter_output=True 
+                filter_output=True
                 if filter_output:
                     try:
                         data_s16 = np.frombuffer(sentence_wav_bytestream, dtype=np.int16, count=len(sentence_wav_bytestream)//2, offset=0)
@@ -776,7 +719,7 @@ def generate_speech_from_history(history, chatbot_role, sentence):
                         sentence_wav_bytestream = sentence_wav_bytestream.tobytes()
                     except:
                         print("failed to remove noise")
-                        
+
                 # Directly encode the WAV bytestream to base64
                 base64_audio = base64.b64encode(pcm_to_wav(sentence_wav_bytestream)).decode('utf8')
 
@@ -804,9 +747,6 @@ def generate_speech_from_history(history, chatbot_role, sentence):
     return results
 
 
-# In[21]:
-
-
 latent_map = {}
 try:
     # get the current working directory
@@ -822,16 +762,12 @@ try:
 
 except Exception as e:
     print("Error:", str(e))
-
-
-# In[ ]:
-
-
+    
 # Define the main function for the API endpoint that takes the input text and chatbot role
 def generate_story_and_speech(secret_token, input_text, chatbot_role):
     if secret_token != SECRET_TOKEN:
         raise gr.Error(
-            f'Invalid secret token. Please fork the original space if you want to use it for yourself.')
+            f'Invalid secret token.  Secret Token: secret')
     # Initialize a list of lists for history with the user input as the first entry
     history = [[input_text, None]]
     story_sentences = get_sentence(history, chatbot_role)  # get_sentence function generates text
@@ -849,7 +785,7 @@ def generate_story_and_speech(secret_token, input_text, chatbot_role):
         # Convert the list of lists back into a list of tuples for the history
         history_tuples = [tuple(entry) for entry in last_history]
 
-        return generate_speech_from_history(history_tuples, chatbot_role, story_text) 
+        return generate_speech_from_history(history_tuples, chatbot_role, story_text)
 
     else:
         return []
@@ -862,5 +798,4 @@ demo = gr.Interface(
 )
 
 demo.queue()
-demo.launch(debug=True)
-
+demo.launch(debug=True)

requirements.txt

@@ -20,4 +20,6 @@ noisereduce==3.0.0
 #deepspeed
 #deepspeed==0.12.6
 deepspeed==0.10.0
-ipython
+ipython
+python-dotenv
+OpenAI