# Usage ### python main.py --mode interface ### python main.py videos/Spirituality_1_clip.mp4 -n 3 --mode inference --model gemini import gradio as gr import os import whisper import cv2 import json import tempfile import torch import transformers from transformers import pipeline import re import time from torch import cuda, bfloat16 from moviepy.editor import VideoFileClip from image_caption import Caption from pathlib import Path from langchain import PromptTemplate from langchain import LLMChain from langchain.llms import HuggingFacePipeline from difflib import SequenceMatcher import argparse import shutil from PIL import Image import google.generativeai as genai from huggingface_hub import InferenceClient from openai import OpenAI class VideoClassifier: global audio_time , setup_time , caption_time , classification_time audio_time = 0 setup_time = 0 caption_time = 0 classification_time = 0 def __init__(self, no_of_frames, mode='interface',model='gemini'): self.no_of_frames = no_of_frames self.mode = mode self.model_name = model.strip().lower() print(self.model_name) os.environ["TOKENIZERS_PARALLELISM"] = "false" if self.model_name=='mistral': print("Setting up Mistral model for Class Selection") self.setup_mistral_model() else : print("Setting up Gemini model for Class Selection") self.setup_gemini_model() self.setup_paths() self.hf_key = os.environ.get("HF_KEY", None) """chatgpt 3.5""" # self.chatgpt_client = OpenAI(api_key="sk-proj-KY1qI7zTpsUiJhMUHuNdT3BlbkFJLOjVnTUSpYJi87yUtSEI") self.chatgpt_client= OpenAI(api_key="sk-proj-TVoFQ4X9apDUs0V6zCDIT3BlbkFJmWRNMgJ6fapge12zygzG") # self.whisper_model = whisper.load_model("base") def setup_paths(self): self.path = './results' if os.path.exists(self.path): shutil.rmtree(self.path) os.mkdir(self.path) def setup_gemini_model(self): self.genai = genai self.genai.configure(api_key="AIzaSyAFG94rVbm9eWepO5uPGsMha8XJ-sHbMdA") self.genai_model = genai.GenerativeModel('gemini-pro') self.whisper_model = whisper.load_model("base") self.img_cap = Caption() def setup_mistral_space_model(self): # if not self.hf_key: # raise ValueError("Hugging Face API key is not set or invalid.") self.client = InferenceClient("mistralai/Mistral-7B-Instruct-v0.2") # self.client = InferenceClient("mistralai/Mistral-7B-Instruct-v0.1") # self.client = InferenceClient("meta-llama/Meta-Llama-3-8B-Instruct") self.whisper_model = whisper.load_model("base") self.img_cap = Caption() def setup_mistral_model(self): self.model_id = "mistralai/Mistral-7B-Instruct-v0.2" self.device = f'cuda:{cuda.current_device()}' if cuda.is_available() else 'cpu' # self.device_name = torch.cuda.get_device_name() # print(f"Using device: {self.device} ({self.device_name})") bnb_config = transformers.BitsAndBytesConfig( load_in_4bit=True, bnb_4bit_quant_type='nf4', bnb_4bit_use_double_quant=True, bnb_4bit_compute_dtype=bfloat16, ) hf_auth = self.hf_key print(hf_auth) model_config = transformers.AutoConfig.from_pretrained( self.model_id, # use_auth_token=hf_auth ) self.model = transformers.AutoModelForCausalLM.from_pretrained( self.model_id, trust_remote_code=True, config=model_config, quantization_config=bnb_config, # use_auth_token=hf_auth ) self.model.eval() self.tokenizer = transformers.AutoTokenizer.from_pretrained( self.model_id, # use_auth_token=hf_auth ) self.generate_text = transformers.pipeline( model=self.model, tokenizer=self.tokenizer, return_full_text=True, task='text-generation', temperature=0.01, max_new_tokens=32 ) self.whisper_model = whisper.load_model("base") self.img_cap = Caption() self.llm = HuggingFacePipeline(pipeline=self.generate_text) def audio_extraction(self,video_input): """When running on local we use this library approach which consumes 3 seconds of gpu inference""" global audio_time start_time_audio = time.time() print(f"Processing video: {video_input} with {self.no_of_frames} frames.") mp4_file = video_input video_name = mp4_file.split("/")[-1] wav_file = "results/audiotrack.wav" video_clip = VideoFileClip(mp4_file) audioclip = video_clip.audio wav_file = audioclip.write_audiofile(wav_file) audioclip.close() video_clip.close() audiotrack = "results/audiotrack.wav" result = self.whisper_model.transcribe(audiotrack, fp16=False) transcript = result["text"] print("TRANSCRIPT",transcript) end_time_audio = time.time() audio_time=end_time_audio-start_time_audio # print("TIME TAKEN FOR AUDIO CONVERSION (WHISPER)",audio_time) return transcript def audio_extraction_space(self,video_input): """When running the project in space we use model directly from huggingface to beat the inference time""" MODEL_NAME = "openai/whisper-large-v3" BATCH_SIZE = 8 device = "cuda" if torch.cuda.is_available() else "cpu" global audio_time start_time_audio = time.time() print(f"Processing video: {video_input} with {self.no_of_frames} frames.") mp4_file = video_input video_name = mp4_file.split("/")[-1] wav_file = "results/audiotrack.wav" video_clip = VideoFileClip(mp4_file) audioclip = video_clip.audio wav_file = audioclip.write_audiofile(wav_file) audioclip.close() video_clip.close() audiotrack = "results/audiotrack.wav" pipe = pipeline( "automatic-speech-recognition", model=MODEL_NAME, device=device ) # if audio_file is None: # return "No audio file submitted! Please upload or record an audio file before submitting your request." # if not os.path.exists(audio_file): # return "File does not exist. Please check the file path." task="transcribe" result = pipe(audiotrack, batch_size=BATCH_SIZE, generate_kwargs={"task": task}, return_timestamps=True) return result["text"] def audio_extraction_chatgptapi(self,video_input): """For cpu inference , we use this function for faster api calling inference""" global audio_time start_time_audio = time.time() print(f"Processing video: {video_input} with {self.no_of_frames} frames.") mp4_file = video_input video_name = mp4_file.split("/")[-1] wav_file = "results/audiotrack.wav" video_clip = VideoFileClip(mp4_file) with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as temp_audio: video_clip.audio.write_audiofile(temp_audio.name, codec='pcm_s16le', nbytes=2, fps=16000) video_clip.close() with open(temp_audio.name, 'rb') as audio_file: transcription = self.chatgpt_client.audio.transcriptions.create( model="whisper-1", file=audio_file ) print(transcription.text) os.remove(temp_audio.name) # audioclip = video_clip.audio # wav_file = audioclip.write_audiofile(wav_file) # audioclip.close() # video_clip.close() # audiotrack = "results/audiotrack.wav" # # client = OpenAI(api_key="sk-proj-KY1qI7zTpsUiJhMUHuNdT3BlbkFJLOjVnTUSpYJi87yUtSEI") # # audiotrack= open("audiotrack.wav", "rb") # transcription = self.client.audio.transcriptions.create( # model="whisper-1", # file=audioclip # ) # print(transcription.text) return transcription.text def generate_text(self, inputs, parameters=None): if parameters is None: parameters = { "temperature": 0.7, "max_new_tokens": 50, "top_p": 0.9, "repetition_penalty": 1.2 } return self.client(inputs, parameters) default_checkbox = [] def classify_video(self,video_input,checkbox=default_checkbox): global classification_time , caption_time print("checkbox",checkbox) # transcript=self.audio_extraction_space(video_input) try: transcript=self.audio_extraction(video_input) except: transcript=self.audio_extraction_space(video_input) # try: # transcript=self.audio_extraction_chatgptapi(video_input) # except : # print("Chatgpt Key expired , inferencing using whisper library") # try: # transcript=self.audio_extraction(video_input) # except: # transcript=self.audio_extraction_space(video_input) start_time_caption = time.time() captions = "" if checkbox==["Image Captions and Audio for Classification"]: video = cv2.VideoCapture(video_input) length = int(video.get(cv2.CAP_PROP_FRAME_COUNT)) no_of_frame = int(self.no_of_frames) temp_div = length // no_of_frame currentframe = 50 caption_text = [] for i in range(no_of_frame): video.set(cv2.CAP_PROP_POS_FRAMES, currentframe) ret, frame = video.read() if ret: frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB) image = Image.fromarray(frame) content = self.img_cap.predict_image_caption_gemini(image) print("content", content) caption_text.append(content) currentframe += temp_div - 1 else: break captions = ", ".join(caption_text) print("CAPTIONS", captions) video.release() cv2.destroyAllWindows() # print("TIME TAKEN FOR IMAGE CAPTIONING", end_time_caption-start_time_caption) end_time_caption = time.time() caption_time=end_time_caption-start_time_caption start_time_generation = time.time() main_categories = Path("main_classes.txt").read_text() main_categories_list = ['Automotive', 'Books and Literature', 'Business and Finance', 'Careers', 'Education','Family and Relationships', 'Fine Art', 'Food & Drink', 'Healthy Living', 'Hobbies & Interests', 'Home & Garden','Medical Health', 'Movies', 'Music and Audio', 'News and Politics', 'Personal Finance', 'Pets', 'Pop Culture','Real Estate', 'Religion & Spirituality', 'Science', 'Shopping', 'Sports', 'Style & Fashion','Technology & Computing', 'Television', 'Travel', 'Video Gaming'] generate_kwargs = { "temperature": 0.9, "max_new_tokens": 256, "top_p": 0.95, "repetition_penalty": 1.0, "do_sample": True, "seed": 42, "return_full_text": False } template1 = '''Given below are the different type of main video classes {main_categories} You are a text classifier that catergorises the transcript and captions into one main class whose context match with one main class and only generate main class name no need of sub classe or explanation. Give more importance to Transcript while classifying . Transcript: {transcript} Captions: {captions} Return only the answer chosen from list and nothing else Main-class => ''' prompt1 = PromptTemplate(template=template1, input_variables=['main_categories', 'transcript', 'captions']) print("PROMPT 1",prompt1) # print(self.model) # print(f"Current model in use: {self.model}") if self.model_name=='mistral': try: print("Entering mistral chain approach") chain1 = LLMChain(llm=self.llm, prompt=prompt1) main_class = chain1.predict(main_categories=main_categories, transcript=transcript, captions=captions) except: print("Entering mistral template approach") prompt1 = template1.format(main_categories=main_categories, transcript=transcript, captions=captions) messages = [{"role": "user", "content": prompt1}] stream = self.client.chat_completion(messages, max_tokens=100) main_class = stream.choices[0].message.content.strip() # output = "" # for response in stream: # output += response['token'].text # print("Streaming output:", output) # main_class = output.strip() print(main_class) print("#######################################################") try: pattern = r"Main-class =>\s*(.+)" match = re.search(pattern, main_class) if match: main_class = match.group(1).strip() except: main_class=main_class else: prompt_text = template1.format(main_categories=main_categories, transcript=transcript, captions=captions) response = self.genai_model.generate_content(contents=prompt_text) main_class = response.text print(main_class) print("#######################################################") print("MAIN CLASS: ",main_class) def category_class(class_name,categories_list): def similar(str1, str2): return SequenceMatcher(None, str1, str2).ratio() index_no = 0 sim = 0 for sub in categories_list: res = similar(class_name, sub) if res>sim: sim = res index_no = categories_list.index(sub) class_name = categories_list[index_no] return class_name if main_class not in main_categories_list: main_class = category_class(main_class,main_categories_list) print("POST PROCESSED MAIN CLASS : ",main_class) tier_1_index_no = main_categories_list.index(main_class) + 1 with open('categories_json.txt') as f: data = json.load(f) sub_categories_list = data[main_class] print("SUB CATEGORIES LIST",sub_categories_list) with open("sub_categories.txt", "w") as f: no = 1 # print(data[main_class]) for i in data[main_class]: f.write(str(no)+')'+str(i) + '\n') no = no+1 sub_categories = Path("sub_categories.txt").read_text() template2 = '''Given below are the sub classes of {main_class}. {sub_categories} You are a text classifier that catergorises the transcript and captions into one sub class whose context match with one sub class and only generate sub class name, Don't give explanation . Give more importance to Transcript while classifying . Transcript: {transcript} Captions: {captions} Return only the Sub-class answer chosen from list and nothing else Answer in the format: Main-class => {main_class} Sub-class => ''' prompt2 = PromptTemplate(template=template2, input_variables=['sub_categories', 'transcript', 'captions','main_class']) if self.model_name=='mistral': try: chain2 = LLMChain(llm=self.llm, prompt=prompt2) sub_class = chain2.predict(sub_categories=sub_categories, transcript=transcript, captions=captions,main_class=main_class) except: prompt2 = template2.format(sub_categories=sub_categories, transcript=transcript, captions=captions,main_class=main_class) messages = [{"role": "user", "content": prompt2}] stream = self.client.chat_completion(messages, max_tokens=100) sub_class = stream.choices[0].message.content.strip() print("Preprocess Answer",sub_class) try: pattern = r"Sub-class =>\s*(.+)" match = re.search(pattern, sub_class) if match: sub_class = match.group(1).strip() except: subclass=sub_class else: prompt_text2 = template1.format(main_categories=main_categories, transcript=transcript, captions=captions) response = self.genai_model.generate_content(contents=prompt_text2) sub_class = response.text print("Preprocess Answer",sub_class) print("SUB CLASS",sub_class) if sub_class not in sub_categories_list: sub_class = category_class(sub_class,sub_categories_list) print("POST PROCESSED SUB CLASS",sub_class) tier_2_index_no = sub_categories_list.index(sub_class) + 1 print("ANSWER:",sub_class) final_answer = (f"Tier 1 category : IAB{tier_1_index_no} : {main_class}\nTier 2 category : IAB{tier_1_index_no}-{tier_2_index_no} : {sub_class}") first_video = os.path.join(os.path.dirname(__file__), "American_football_heads_to_India_clip.mp4") second_video = os.path.join(os.path.dirname(__file__), "PersonalFinance_clip.mp4") # return final_answer, first_video, second_video end_time_generation = time.time() classification_time = end_time_generation-start_time_generation print ("MODEL USED :",self.model_name) print("MODEL SETUP TIME :",setup_time) print("TIME TAKEN FOR AUDIO CONVERSION (WHISPER) :",audio_time) print("TIME TAKEN FOR IMAGE CAPTIONING :", caption_time) print("TIME TAKEN FOR CLASS GENERATION :",classification_time) print("TOTAL INFERENCE TIME :",audio_time+caption_time+classification_time) return final_answer def save_model_choice(self,model_name): global setup_time start_time_setup = time.time() self.model_name = model_name if self.model_name=='mistral': print("Setting up Mistral model for Class Selection") self.setup_mistral_space_model() else : print("Setting up Gemini model for Class Selection") self.setup_gemini_model() end_time_setup = time.time() setup_time=end_time_setup-start_time_setup # print("MODEL SETUP TIME",setup_time) return "Model selected: " + model_name def launch_interface(): css_code = """ .gradio-container { background-color: #d6cbd6; } /* Button styling for all buttons */ button { background-color: #d6cbd6; /* Default color for all other buttons */ color: black; border: 1px solid black; padding: 10px; margin-right: 10px; font-size: 16px; /* Increase font size */ font-weight: bold; /* Make text bold */ } /* Style for the second button */ button:nth-child(2) { background-color: #927fc7; /* Custom color for the second button */ } """ # First interface for model selection interface_1 = gr.Interface( fn=save_model_choice, inputs=gr.Dropdown( choices=['gemini', 'mistral'], label="Select Model", info="Default model: Gemini", ), outputs="text", title="Model Selection", ) # Second interface for video classification video_examples = [ [os.path.join(os.path.dirname(__file__), "American_football_heads_to_India_clip.mp4")], [os.path.dirname(__file__) + "/PersonalFinance_clip.mp4"], [os.path.dirname(__file__) + "/Motorcycle_clip.mp4"], [os.path.dirname(__file__) + "/Spirituality_1_clip.mp4"], [os.path.dirname(__file__) + "/Science_clip.mp4"], ] checkbox = gr.CheckboxGroup( ["Image Captions and Audio for Classification"], label="Features", info="default: Audio for classification", ) interface_2 = gr.Interface( fn=classify_video, inputs=[gr.PlayableVideo(), checkbox], outputs="text", examples=video_examples, title="Video Classification", css=css_code, ) # Create a tabbed interface gr.TabbedInterface( [interface_1, interface_2], ["Model Selection", "Video Classification"], css=css_code, ).launch(debug=True, share=True) launch_interface() # css_code = """ # .gradio-container {background-color: #FFFFFF;color:#000000;background-size: 200px; background-image:url(https://gitlab.ignitarium.in/saran/logo/-/raw/aab7c77b4816b8a4bbdc5588eb57ce8b6c15c72d/ign_logo_white.png);background-repeat:no-repeat; position:relative; top:1px; left:5px; padding: 50px;text-align: right;background-position: right top;} # """ # css_code += """ # :root { # --body-background-fill: #FFFFFF; /* New value */ # } # """ # css_code += """ # :root { # --body-background-fill: #000000; /* New value */ # } # """ # interface_1 = gr.Interface( # self.save_model_choice, # inputs=gr.Dropdown(choices=['gemini', 'mistral'], label="Select Model", info="Default model: Gemini"), # # outputs=interface_1_output, # outputs="text" # ) # video_examples = [ # [os.path.join(os.path.dirname(__file__), "American_football_heads_to_India_clip.mp4")], # [os.path.join(os.path.dirname(__file__), "PersonalFinance_clip.mp4")], # [os.path.join(os.path.dirname(__file__), "Motorcycle_clip.mp4")], # [os.path.join(os.path.dirname(__file__), "Spirituality_1_clip.mp4")], # [os.path.join(os.path.dirname(__file__), "Science_clip.mp4")] # ] # # Define the checkbox for additional feature control # checkbox = gr.CheckboxGroup( # ["Image Captions and Audio for Classification"], # label="Features", # info="default : Audio for classification", # ) # default_checkbox = [] # demo = gr.Interface(fn=self.classify_video, inputs=["playablevideo",checkbox],allow_flagging='never', examples=video_examples, # cache_examples=False, outputs=["text"], # css=css_code, title="Interactive Advertising Bureau (IAB) compliant Video-Ad classification") # # demo.launch(debug=True) # gr.TabbedInterface([interface_1, demo], ["Model Selection", "Video Classification"]).launch(debug=True) def run_inference(self, video_path,model): result = self.classify_video(video_path) print(result) if __name__ == "__main__": parser = argparse.ArgumentParser(description='Process some videos.') parser.add_argument("video_path", nargs='?', default=None, help="Path to the video file") parser.add_argument("-n", "--no_of_frames", type=int, default=3, help="Number of frames for image captioning") parser.add_argument("--mode", choices=['interface', 'inference'], default='interface', help="Mode of operation: interface or inference") parser.add_argument("--model", choices=['gemini','mistral'],default='gemini',help="Model for inference") args = parser.parse_args() vc = VideoClassifier(no_of_frames=args.no_of_frames, mode=args.mode , model=args.model) if args.mode == 'interface': vc.launch_interface() elif args.mode == 'inference' and args.video_path and args.model: vc.run_inference(args.video_path,args.model) else: print("Error: No video path/model provided for inference mode.")