#Project ref: 
#1. https://github.com/haltakov/natural-language-image-search by Vladimir Haltakov
#2. OpenAI's CLIP


import torch
import requests
import numpy as np
import pandas as pd
import gradio as gr
from io import BytesIO

device = 'cuda' if torch.cuda.is_available() else 'cpu'

from PIL import Image as pil
from transformers import  CLIPProcessor, CLIPModel,CLIPTokenizer
from sentence_transformers import SentenceTransformer, util

model = CLIPModel.from_pretrained('openai/clip-vit-base-patch32').to(device)
processor = CLIPProcessor.from_pretrained('openai/clip-vit-base-patch32')
tokenizer = CLIPTokenizer.from_pretrained('openai/clip-vit-base-patch32')

examples = [['flag.jpg'],['flowe.jpg'],['dance.jpg']]

data = pd.read_csv('./photos.tsv000', sep='\t', header=0)
photo_features= np.load('./features.npy')
ids = pd.read_csv('./photo_ids.csv')
ids = list(ids['photo_id'])

def encode_img(image):
  img = pil.fromarray(image.astype('uint8'),'RGB')
  with torch.no_grad():
    processed =  processor(text = None,images=image, return_tensors='pt',padding=True)['pixel_values']
    search_photo_feature = model.get_image_features(processed.to(device))
    search_photo_feature /= search_photo_feature.norm(dim=1,keepdim=True)
  img_encoded = search_photo_feature.cpu().numpy()
  return img_encoded

def encode_txt(text):
  with torch.no_grad():
    inp = tokenizer([text],padding=True, return_tensors='pt')
    inp = processor(text=[inp], images=None, return_tensors='pt',padding=True)
  text_encoded = model.gt_text_text_features(**inp).detach().numpy()
  return text_encoded

def similarity(feature, photo_features):
  similarities = list((feature @ photo_features.T).squeeze(0))
  return similarities

def find_best_matches(image, mode, text):
  if mode == 'Text2Image':
    text_features = encode_txt(text)
    similarities = similarity(text_features, photo_features)
  else:
    img_features = encode_img(image)
    similarities = similarity(img_features, photo_features)

  best_photos = sorted(zip(similarities, range(photo_features.shape[0])))

  matched_images = []

  for i in range(4):
    idx = best_photos[i][1]
    photo_id = ids[idx]

    photo_data = data[data['photo_id'] == photo_id].iloc[0]

    response = requests.get(photo_data['photo_image_url']+ '?w=640')
    img = pil.open(BytesIO(response.content))
    matched_images.append(img)

  return matched_images

demo = gr.Interface(fn=find_best_matches, inputs = [ gr.Image(label='Search image'),
                                              gr.Radio(['Text to Image', 'Image to Image']),
                                              gr.Textbox(lines=1, label = 'Text query')],
                                              examples = examples,
                                              outputs = [gr.Gallery(label= 'Generated images', show_label=False, elem_id='gallery', scale = 3)],
                                              title = 'CLIP tutorial',
                                              description = 'This application gives best 4 results for the query by the user. Input is either text or image and get image and text as results respectively.')
demo.launch()