Initial commit

app.py

@@ -0,0 +1,68 @@
+from langchain.llms import HuggingFacePipeline
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline, AutoModelForSeq2SeqLM
+
+from components import caption_chain, tag_chain
+from components import pexels, utils
+import os, gc
+import gradio as gr
+
+model = AutoModelForSeq2SeqLM.from_pretrained("declare-lab/flan-alpaca-gpt4-xl")
+tokenizer = AutoTokenizer.from_pretrained("declare-lab/flan-alpaca-gpt4-xl")
+
+pipe = pipeline(
+    'text2text-generation',
+    model=model,
+    tokenizer= tokenizer,
+    max_length=120
+)
+
+local_llm = HuggingFacePipeline(pipeline=pipe)
+
+llm_chain = caption_chain.chain(llm=local_llm)
+sum_llm_chain = tag_chain.chain(llm=local_llm)
+
+pexels_api_key = os.getenv('pexels_api_key')
+
+def pred(product_name, orientation):
+    if orientation == "Shorts/Reels/TikTok (1080 x 1920)":
+        orientation = "potrait"
+        height = 1920
+        width = 1080
+    elif orientation == "Youtube Videos (1920 x 1080)":
+        orientation = "landscape"
+        height = 1080
+        width = 1920
+    else :
+        orientation = "square"
+        height = 1080
+        width = 1080
+    folder_name, sentences = pexels.generate_videos(product_name, pexels_api_key, orientation, height, width, llm_chain, sum_llm_chain)
+    gc.collect()
+    utils.combine_videos(folder_name)
+    return ["\n".join(sentences), os.path.join(folder_name, "Final_Ad_Video.mp4")]
+        #{'video':os.path.join(folder_name, "Final_Ad_Video.mp4"),
+       # 'captions':"\n".join(sentences)}
+    
+
+with gr.Blocks() as demo:
+    gr.Markdown(
+        """
+        # Ads Generator
+        Create video ads based on your product name using AI
+        ### Note : the video generation takes about 2-4 minutes 
+        """
+    )
+    dimension = gr.Dropdown(
+            ["Shorts/Reels/TikTok (1080 x 1920)", "Facebook/Youtube Videos (1920 x 1080)", "Square (1080 x 1080)"], 
+            label="Video Dimension", info="Choose dimension"
+        )
+    product_name = gr.Textbox(label="product name")
+    captions = gr.Textbox(label="captions")
+    video = gr.Video()
+    btn = gr.Button("Submit")
+    btn.click(pred, inputs=[product_name, dimension], outputs=[captions,video])
+    
+    
+
+demo.launch()

components/caption_chain.py

@@ -0,0 +1,12 @@
+from langchain import PromptTemplate
+from langchain.chains import LLMChain
+
+def chain(llm):
+    template = """Make 5 different advertisement captions about this product. 
+    {product}
+    """
+    
+    prompt = PromptTemplate(template=template, input_variables=["product"])
+    
+    llm_chain = LLMChain(prompt=prompt, llm=llm)
+    return llm_chain

components/pexels.py

@@ -0,0 +1,81 @@
+import requests
+import shutil,os,re
+
+# Searching for the videos
+def search_pexels(keyword, api_key, orientation='potrait', size='medium', endpoint='videos', num_pages=50):
+    
+  if orientation not in ['potrait', 'landscape', 'square']:
+    raise Exception("Error! orientation must be one of {'square', 'landscape', 'potrait'}")
+
+  if size not in ['medium', 'small', 'large']:
+    raise Exception("Error! size must be one of ['medium', 'small', 'large']")
+
+  base_url = 'https://api.pexels.com/'
+
+  headers = {
+      'Authorization': f'{api_key}'
+  }
+
+  url = f'{base_url}{endpoint}/search?query={keyword}&per_page={num_pages}&orientation={orientation}&size={size}'
+
+
+  response = requests.get(url, headers=headers)
+
+  # Check if request was successful (status code 200)
+  if response.status_code == 200:
+      data = response.json()
+      return data
+  else:
+      print(f'Error: {response.status_code}')
+
+
+# Video download function
+def download_video(data, parent_path, height, width, links, i):
+    for x in data['videos'] :
+        if x['id'] in links:
+            continue
+        
+        vid = x['video_files']
+        for v in vid: 
+            if v['height'] == height and v['width'] == width :
+                with open(f"{os.path.join(parent_path,str(i) + '_' + str(v['id']))}.mp4", 'bw') as f:
+                    f.write(requests.get(v['link']).content)
+                    print("Sucessfully saved video in", os.path.join(parent_path,str(i) + '_' + str(v['id'])) + '.mp4')
+                    return x['id']
+
+
+# Utilizing the LLMs to find the relevant videos 
+def generate_videos(product, api_key, orientation, height, width, llm_chain=None, sum_llm_chain=None):
+    prod = product.strip().replace(" ", "_")
+    links = []
+    try :
+        # Split the paragraph by sentences
+            
+        sentences = llm_chain.run(product.strip())
+        print('Sentence :', sentences)
+        
+#         sentences = sentences.split(".")[:-1]
+        sentences = [x.strip() for x in re.split(r'\d+\.', sentences) if len(x) > 6]
+    
+        
+        # Create directory with the product's name
+        if os.path.exists(prod):
+            shutil.rmtree(prod)
+        os.mkdir(prod)
+        
+        # Generate video for every sentence
+        print("Keyword :")
+        for i,s in enumerate(sentences):
+            keyword = sum_llm_chain.run(s)
+            print(i+1, ":", keyword)
+            data = search_pexels(keyword, api_key, orientation.lower())
+            link = download_video(data, prod, height, width, links,i)
+            links.append(link)
+            
+        print("Success! videos has been generated")
+    except Exception as e :
+        print("Error! Failed generating videos")
+        print(e)
+    
+    return prod, sentences
+        

components/tag_chain.py

@@ -0,0 +1,16 @@
+from langchain.prompts import PromptTemplate
+from langchain.chains import LLMChain
+
+def chain(llm):
+  sum_template = """what is the most significant actions or places or things, say it in at most 5 words :
+  
+  {sentence}
+  
+  
+  """
+  
+  sum_prompt = PromptTemplate(template=sum_template, input_variables=["sentence"])
+  
+  sum_llm_chain = LLMChain(prompt=sum_prompt, llm=llm)
+
+  return sum_llm_chain

components/utils.py

@@ -0,0 +1,23 @@
+from moviepy.editor import VideoFileClip, concatenate_videoclips,vfx
+import os 
+
+def combine_videos(folder_name):
+
+    length = len(os.listdir(folder_name))
+
+    vids = [VideoFileClip(os.path.join(folder_name,x)) for x in sorted(os.listdir(folder_name))]
+    
+    combined = concatenate_videoclips(
+        [vid.subclip(0,7).fx(vfx.fadein,0.5).fx(vfx.fadeout,0.5) if vid.duration > 7 else vid.fx(vfx.fadein,0.5).fx(vfx.fadeout,0.5) for vid in vids],
+        # [VideoFileClip(os.path.join(folder_name,x)).subclip(0,7).fx(vfx.fadein,0.5).fx(vfx.fadeout,0.5) for x in sorted(os.listdir(folder_name))], 
+        "compose",bg_color=None, padding=0
+    )
+    combined.write_videofile(os.path.join(folder_name, "Final_Ad_Video.mp4"))
+    
+#     if length <= 6:
+#         combined = concatenate_videoclips([VideoFileClip(os.path.join(folder_name,x)).subclip(0,10) for x in os.listdir(folder_name)])
+#         combined.write_videofile(os.path.join(folder_name, "Ad_Video.mp4"))
+#     else :
+#         combined = concatenate_videoclips([VideoFileClip(os.path.join(folder_name,x)).subclip(0,60//length) for x in os.listdir(folder_name)])
+#         combined.write_videofile(os.path.join(folder_name, "Ad_Video.mp4"))
+    print("Done! Your ads video has been created")

requirements.txt

@@ -0,0 +1,4 @@
+langchain
+moviepy
+transformers
+torch