Great, Thanks. The multi line approach works with alphabet, but with not with Japanese. It does multi lines but ignoring the size constraint, and it crops. The text always appears the same regardless of what I select for the size (i.e. 0.8*clip width, 0.2* clip width etc..). Any thought?
@deleabdulqudus67922 ай бұрын
Nice video! What website did you get the api endpoint from?
@sheikhmusa84422 ай бұрын
Nice to see basic level concepts of JSON's in this advanced tutorial guide. You are doing great job
@mytravls3 ай бұрын
And flexing that. What nerve
@raghavkishore4773 ай бұрын
Sir, can you tell me why you havent used the tfidfvectorizer instead?
@Modbaiby4 ай бұрын
My phone doesn't have that trending effect how can I get it
@bent3bent4 ай бұрын
Master of smacking your lips when you speak 👌
@salmannabi36704 ай бұрын
Very Informative! Create more content like this.
@PracticalAIbyRamsri4 ай бұрын
Thanks, Salman! Sure!
@KunaalNaik5 ай бұрын
Ramsri, this is amazing! I am currently on projects on GenAI. With this you give a very good idea on how to launch my own product in market! Thank you :)
@PracticalAIbyRamsri5 ай бұрын
Glad that you found it useful, Kunaal :)
@RandomPositiveLive5 ай бұрын
Hallo Sir, can I Get your Contact phone. I want discuss and learn about Automatic distractor generation with syntagmatic relation on vocabulary question toefl with LLM.. I will thanksfull if your can help me. OR can i get your email? thankyou sir
@GAtamil145 ай бұрын
Great vid sir ❤❤
@parasmadan_in5 ай бұрын
Loved it
@atchutram98945 ай бұрын
How to you manage your billing? How do you make sure the money you spend for the API's are recovered?
@knowledgenidhi5 ай бұрын
Thank you Ramsri sir
@LuizChequini5 ай бұрын
Could you help regarding this mask_color effect is not working Thanks main_clip = VideoFileClip("videos/19990812-uhd_2560_1440_30fps.mp4").subclip(0, 10) effect_clip = VideoFileClip("effects/production ID_4514642.mp4").subclip(0, 10) # Create a mask for the effect, removing the black background effect_clip = effect_clip.fx(vfx.mask_color, color=[0, 0, 0] ,thr=0, s=1) # Adjust opacity to 50% #effect_clip = effect_clip.set_opacity(.5) # Combine the clips composite_clip = CompositeVideoClip([main_clip, effect_clip.set_position(("center", "center"))])
@developershub20245 ай бұрын
super cool, well explained, thank you.
@thepresistence59355 ай бұрын
Where can I get this code?
@jayaraghavendra90255 ай бұрын
Hi Ramsri, Could you provide slack link please.
@ms-ej4gd5 ай бұрын
Great efforts
@heisenbergwhite58455 ай бұрын
Please share the code, if possible What is your opinion on learning NLP from scratch, like shown in this course, versus just using openai APIs
@Hassan-b6l5 ай бұрын
❤❤❤ Great work, is right to left language be use for examples and also audio/speech processing. ❤❤❤❤❤
@ModuleAYX6 ай бұрын
Thanks your style and solution, both rocks. Json text, has excellent timings for each words. Can we convert to text, For example. Hi my name is RAM, comvet to Audio and merge to this video. Does this works?
@eso9076 ай бұрын
You are the best ! Thank you.
@zero-code1396 ай бұрын
Great video Ram. Do you deploy these as Google Cloud Functions? Is it possible?
@SantiYounger6 ай бұрын
great video thank you!
@clearmindstudios6 ай бұрын
Love your videos super helpful and friendly
@rhsundaram7 ай бұрын
Nice, can we do the same for input to save loading all the embedding token weights?
@jayanthnani80227 ай бұрын
I saved using model.save_preatrained(output_path) without lora and peft , only with base llm model. Still got the error. Anyways i tried the approach shown in the video for text classification wherein the output class is a combination of multiple tokens instead of a single token as shown in video. However , the output is always a single token which is the first token of the predicted class and not all the tokens that are part of that class even though max_new_parameters is 5 . Any reason for this behaviour?? @PracticalAIbyRamsri
@jayanthnani80227 ай бұрын
I saved the changed architecture model in a folder and then tried loading from that folder for inference, i am getting error because of shape mismatch in base model and our latest model
@PracticalAIbyRamsri7 ай бұрын
Are you saving PeFt(Lora) model or full model. Basically do a model and lora merge and save.
@jayanthnani80227 ай бұрын
I am just saving model using model.save_pretrained(output_path), no lora and peft , i am using only base model . I used the approach for a text classification job where in the output class comprises of multiple tokens instead of a single token as shown in video, when doing inference the output is only starting token of the class even though max_new_tokens parameter is 5, any idea on why this particular behaviour is happening?@@PracticalAIbyRamsri
@jayanthnani80227 ай бұрын
I saved using model.save_preatrained(output_path) without lora and peft , only with base llm model. Still got the error. Anyways i tried the approach shown in the video for text classification wherein the output class is a combination of multiple tokens instead of a single token as shown in video. However , the output is always a single token which is the first token of the predicted class and not all the tokens that are part of that class even though max_new_parameters is 5 . Any reason for this behaviour??@@PracticalAIbyRamsri
@jayanthnani80227 ай бұрын
I saved using model.save_preatrained(output_path) without lora and peft , only with base llm model. Still got the error. Anyways i tried the approach shown in the video for text classification wherein the output class is a combination of multiple tokens instead of a single token as shown in video. However , the output is always a single token which is the first token of the predicted class and not all the tokens that are part of that class even though max_new_parameters is 5 . Any reason for this behaviour?? @@PracticalAIbyRamsri
@no_name56967 ай бұрын
Why to strip this, you can prompt the model to output this. However, great demo. Keep it up?
@PracticalAIbyRamsri7 ай бұрын
As mentioned in the video to reduce hallucinations significantly, adapt the model to be monolingual (just respond in Telugu, Hindi, etc). Constrained generation essentially if your task necessiates that like with sentiment analysis, classification, mono language generation etc.
@indavarapuaneesh28717 ай бұрын
off topic: where do you generally deploy the finetuned AI model ? Is it on AWS or any other cloud service provider ?
@PracticalAIbyRamsri7 ай бұрын
Modal .com , replicate
@indavarapuaneesh28717 ай бұрын
Great video Anna, keep releasing the videos. Learned a lot about classification. Even GPT 4.0 is not very good with classification of neutral statements. But more often than not we would be very interested in only classifying negative or positive values which means those models can be trained in a single dimension.
@PracticalAIbyRamsri7 ай бұрын
True! Sure will keep posting!
@VedicVisionsToday7 ай бұрын
Insightful
@gunasekhar84407 ай бұрын
Thanks for making the video. But i have a doubt that how to make tokenization model with help of these encoding and decoding the sentence correctly ?
@PracticalAIbyRamsri7 ай бұрын
Not sure if I understood your question, Guna Shekhar! Can you rephrase it ?
@gunasekhar84407 ай бұрын
@@PracticalAIbyRamsri sure. I mean from the last two months I was stuck in building the own tokenization model for a particular telugu document. How can we use gemma for doing this? My main goal is to build a telugu text to image converter. For that tokenization model is very important for embedding the sentence right . As you mentioned in the video gemma used a byte pair encoding technique. But in the telugu for forming a word is a combination of multiple chara as you explained in the video.
@PracticalAIbyRamsri7 ай бұрын
@@gunasekhar8440 You can train a sentencepiece tokenizer from scratch giving it a lot of Telugu sentences. HuggingFace has tutorials on it
@gunasekhar84407 ай бұрын
@@PracticalAIbyRamsri thanks for that. I have tried to search but everything is there for English.
@AIGeddon7 ай бұрын
🎯 Key Takeaways for quick navigation: 💻 Generative AI creates new content, including text, code, images, speech, and music. indistinguishable from human creation. 🧠 It's not just one algorithm, but a collection of them focusing on different content types. 📈 Generative AI has become very sophisticated, making it almost indistinguishable from human creation. 🏢 Big players like OpenAI have been leading the recent generative AI advancements. 🤔 Generative AI progress raises questions about the future of human-like artificial intelligence (AGI). 🎨 Generative AI has real-world applications, like creating memes or translating videos. 💼 Opportunities exist for building innovative products and businesses on top of generative AI. 🏢 Several companies are leaders in generative AI: OpenAI, Cohere, A121 Labs, Anthropic, Hugging Face. 🗣️ OpenAI has made advancements in both text generation (GPT-3, ChatGPT) and speech recognition (Whisper). 👂Other companies are pushing text-to-speech boundaries (Much.ai, 11Labs.ai). 🎨 Image & video generation is also a key area (Midjourney, Stable Diffusion, Dali, Synthesia.io, Rephrase.ai). 💡 Products can be built by combining different generative AI technologies (text, speech, image). 🎯 Focusing on a specific vertical (like EdTech or healthcare) is helpful for product development. 🧘♂️ Example: Personalized guided meditation sessions can be created with existing generative AI tools. 🌐 Opportunities exist to create innovative SaaS products using generative AI. 🎥 KZbin videos can be repurposed for multiple formats using generative AI. 📖 Transcripts can be turned into blog posts or quizzes with the help of GPT-3. 🎨 Illustrate books (especially children's books) with image generation tools (DALL-E, Stable Diffusion, etc.). 🤖 Create "faceless" KZbin channels where everything (script, voice, visuals) is AI-generated. 🧒 Potential niche: AI-generated children's stories & rhymes tailored to different cultures and languages. 🌐 Summarize existing content with AI, then translate it for regional KZbin channels in multiple languages. 🗣️ Clone your own voice with tools like Murf.ai or Synthesia for unique KZbin content. 🔮 The future may hold fully automated KZbin channels run by AI avatars. 🎨 Maintaining character consistency in image generation is a challenge, but workarounds exist. 🧘♂️ GPT-3 can generate personalized guided meditation scripts based on simple prompts. 🔧 Code can be used to make scripts dynamic (changing names, professions, etc.) for easy customization. 🗣️ Combining text-to-speech tools with generated scripts allows for audio creation. 🤖 Current AI voices can sound robotic, a known limitation in the field. 🧪 Tools like 11 Labs offer more emotionally nuanced text-to-speech synthesis. 🧩 Combining the best specialized tools (like Synthesia for visuals, 11 Labs for voice) can improve results. 📈 The overall quality of generative AI output is continuously improving. 📝 ChatGPT output can be made more engaging using paraphrasing tools like Quillbot. 🤖 Tools exist to detect AI-generated text, underscoring the "formal" quality of tools like ChatGPT. 👑 GPT-3 remains a top choice for versatile text generation due to its speed and quality. 👂 OpenAI's Whisper is considered the leader in speech-to-text accuracy. 🎨 Midjourney is a high-quality image generator, but currently lacks an API for building apps. 💸 When choosing tools, consider the cost model (pay-per-use vs. building your own infrastructure). 🔍 GPT-3 and OpenAI are also good for vector embeddings (text search functionalities). 🔄 For parsing data from existing text (like emails), GPT-3 remains a strong option. 🤝 Building a team with diverse skills (AI, development, marketing) can reduce costs and accelerate development. ☁️ Serverless platforms offer a cost-effective, pay-per-use model for AI apps. 🛠️ No-code tools like Bubble.io can provide affordable solutions for building app interfaces. 🎯 Finding a niche (like quiz generation for educators) and targeted marketing can boost growth. 🌐 Tools like Typedream can help create SEO-optimized landing pages. ⚖️ Solo founders may benefit from no-code solutions, while teams can leverage coding for faster iteration and better UI. 💰 Operating costs can vary, but both no-code and coded solutions have the potential to be profitable. 🌟 Feedback highlights the speaker's calm delivery and deep knowledge in the field. 📈 The speaker prioritizes organic growth (SEO) but recognizes the potential of paid marketing and affiliate channels. 💰 Strategic ad spending on platforms like Facebook can significantly accelerate growth. 🤖 Specialized chatbot solutions are emerging for industries like travel. 🔧 Tools like Lang Chain and tutorials from OpenAI can help build custom chatbots. 🔎 Some chatbot solutions can integrate internet search for providing real-time information. 👨💻 The speaker is open to sharing more specific tools/resources for building chatbots if contacted on Twitter. 📣 The host emphasizes the session's value and encourages viewers to express gratitude. 🔗 The speaker's Twitter handle is provided for further inquiries.
@mai_hun_sultan7 ай бұрын
anna, are you working on these ideas or you are suggesting the viewers these ideas?
@avijitthawani10517 ай бұрын
Curious if you recommend Cursor over VSCode+Copilot, and if so why?
@gloriamessengatiok50177 ай бұрын
Thanks you. it's really helpful
@shivanshjayara63728 ай бұрын
anyone can tell me this: When I am using notebook in local or in aws, following command shows error command 1: !apt install imagemagick Error: W: Unable to read /etc/apt/sources.list.d/sdcss.list - open (13: Permission denied) E: Could not open lock file /var/lib/dpkg/lock-frontend - open (13: Permission denied) E: Unable to acquire the dpkg frontend lock (/var/lib/dpkg/lock-frontend), are you root? Command 2: !cat /etc/ImageMagick-6/policy.xml | sed 's/none/read,write/g'> /etc/ImageMagick-6/policy.xml Error: /bin/bash: /etc/ImageMagick-6/policy.xml: Permission denied cat: write error: Broken pipe but when I am using this on colab it executed succesfully. Can anyone explain this please. THis will be a great help.
@shortsfusion_8 ай бұрын
Please is possible to use in local pc?
@wpatterson828 ай бұрын
You rock!
@wpatterson828 ай бұрын
This is incredible! Thank yoy
@ayushbhati43128 ай бұрын
is segmind realvis api can be used for commercial use case
@ketanpatel10688 ай бұрын
Hi! Great video!! Is there a word limit for the app?
@BlueFxVideo8 ай бұрын
Ramsri, this was very useful! Thanks for the Bubble - Huggingface api integration example!
@Triunity3789 ай бұрын
How can i change my profile pic in ideogram
@srikanthporandla49239 ай бұрын
can you share the full code of MCQS generations?
@tshaylatte95029 ай бұрын
Thank you so much for this content
@mehmetyilmaz0019 ай бұрын
Good work thanks. Is it possible to create the video transparent for using on another clip?