import gradio as gr import torch from PIL import Image import torch.nn.functional as F import numpy as np import pickle import json import requests from transformers import CLIPProcessor, AutoModelForSemanticSegmentation, AutoFeatureExtractor, CLIPModel from torch import nn import io # Initialize the models using huggingface device = "cuda" if torch.cuda.is_available() else "cpu" # Load the CLIP model from hugging face clip_hg = CLIPModel.from_pretrained("openai/clip-vit-base-patch32").to(device).eval() processor_hg = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32") seg_hg = AutoModelForSemanticSegmentation.from_pretrained('mattmdjaga/segformer_b2_clothes').to(device).eval() extractor_hg = AutoFeatureExtractor.from_pretrained('mattmdjaga/segformer_b2_clothes', reduce_labels=False) # Load the data and normalize the embeddings just in case. features = torch.load('features.pt').to(device) features_main = F.normalize(features) item_embeddings = torch.load('item_embeds.pt').to(device) item_embeddings = F.normalize(item_embeddings) url_list_main = pickle.load(open('new_url_list.pt','rb')) clothes_tree = json.load(open('clothes_tree_new_data.json')) rec_dic = json.load(open('top5_mini_new.json')) # URL for an image if no image is selected url = 'https://bitsofco.de/content/images/2018/12/Screenshot-2018-12-16-at-21.06.29.png' # Set up all the variables label = ['Background', 'Hat', 'Hair', 'Sunglasses', 'Upper-clothes', 'Skirt', 'Pants', 'Dress', 'Belt', 'Left-shoe', 'Right-shoe', 'Face', 'Left-leg', 'Right-leg', 'Left-arm', 'Right-arm', 'Bag', 'Scarf'] clothing_type = ['top', 'bottom', 'dress'] top_type = ['t-shirt', 'tank top', 'blouse', 'sweater', 'hoodie', 'cardigan','turtleneck','blazer','polo','collar shirt','knitwear', 'tuxedo', 'Compression top','duffle coat', 'peacoat', 'long coat', 'trench coat', 'biker jacket', 'blazer', 'bomber jacket', 'hooded jacket', 'leather jacket', 'military jacket', 'down jacket', 'shirt jacket', 'suit jacket', 'dinner jacket', 'gillet', 'track jacket' ] bottom_type = ['skirt', 'leggings', 'sweatpants', 'skinny pants', 'tailored pants', 'track pants', 'wide-leg pants' , 'cargo shorts', 'denim shorts', 'track shorts', 'compression shorts', 'cycling shorts','denim pants', 'cargo pants', 'chino pants', 'chino shorts' ] dress_type = ['casual dress', 'cocktail dress', 'evening dress', 'maxi dress', 'mini dress', 'party dress', 'sundress'] styles = ['plain','polka dot','striped','floral','checkered','zebra print','leopard print','plaid','paisley'] colors = ['blue','red','pink','orange','yellow','purple','gold','white','off white','black','grey','green','brown','beige','cream','navy','maroon'] top_list = [f"{t}, {color}, {style}" for t in top_type for style in styles for color in colors] bottom_list = [f"{t}, {color}, {style}" for t in bottom_type for style in styles for color in colors] dress_list = [f"{t}, {color}, {style}" for t in dress_type for style in styles for color in colors] all_items = top_list + bottom_list + dress_list clothing_type = ['top', 'bottom', 'dress'] all_types = {'top' :top_type, 'bottom' : bottom_type, 'dress':dress_type} patterns_list = styles.copy() colors_list = colors.copy() clicks = 0 c_types = [] types = [] colors = [] patterns = [] new_files = [] out = [] clothes_click = 0 global_mask = None mask_choice = 'Clothes' # Define all needed functions def find_closest(target_feature, features): ''' Purpose: Find the closest embedding to the given image embedding Inputs: target_feature (tenosr): embedding of our search item features (tensor): embedding of all the items in the dataset Outputs: group_sorted_indices (list): indicies of the closest items in a sorted order ''' cos_similarity = features.to(torch.float32) @ target_feature.to(torch.float32).T group_sorted_indices = torch.argsort(cos_similarity, descending=True,dim=0).squeeze(1).cpu().tolist() return group_sorted_indices def filter_function(choices): ''' Purpose: Find a list of items that match the given filters Inputs: choices (list): list of filters Outputs: Upating the choices of filters ''' # Import the global variables global clicks global c_types global types global colors global patterns global new_files new_choices = [] # Clicks is just a reference to how far we are in the filter tree # We keep going down and saving the selected options until we reach the end # Then we add items which had the desired filters to the new_choices list # This is then used to filter out the items that don't match the filters in search if clicks == 0: temp_choices = [choice for choice in choices if choice in clothing_type] if len(temp_choices) == 0: temp_choices = clothing_type for choice in temp_choices: c_types.append(choice) new_choices.extend(list(clothes_tree[choice].keys())) if clicks == 1: temp_choices = [choice for c_type in c_types for choice in choices if choice in all_types[c_type]] if len(temp_choices) == 0: types = [] for c_type in c_types: types.extend([(t,c_type) for t in clothes_tree[c_type].keys()]) for choice in temp_choices: if choice in clothes_tree['top']: types.append((choice,'top')) elif choice in clothes_tree['bottom']: types.append((choice,'bottom')) else : types.append((choice,'dress')) new_choices = list(clothes_tree['top']['t-shirt'].keys()) if clicks == 2: temp_choices = [choice for choice in choices if choice in colors_list] if len(temp_choices) == 0: colors = colors_list.copy() for choice in temp_choices: colors.append(choice) new_choices = list(clothes_tree['top']['t-shirt']['red'].keys()) if clicks == 3: temp_choices = [choice for choice in choices if choice in patterns_list] if len(temp_choices) == 0: patterns = patterns_list.copy() for choice in temp_choices: patterns.append(choice) for type_,c_type in types: for color in colors: for pattern in patterns: new_files.extend(clothes_tree[c_type][type_][color][pattern]) clicks += 1 new_choices = ['Press Search to use the set filter. Dont press this button'] return gr.update(choices=new_choices, label='Press Search to use the filter or press filter to reset the filter') if clicks == 4: c_types.clear() types.clear() colors.clear() patterns.clear() new_files.clear() clicks = 0 new_choices = ['top','bottom','dress'] return gr.update(choices=new_choices,label='Select the type of clothing you want to search for') clicks += 1 return gr.update(choices=new_choices) def set_theme(theme): ''' Purpose: Set the theme using filters Inputs: theme (string): theme to be set Outputs: Upadting to show the chosen theme ''' global new_files new_files.clear() # Here we just manually set the filters to the desired theme # Then we just find images with the desired filters if theme == 'Red carpet': types = [('evening dress','dress'), ('tuxedo','top'), ('suit jacket','top'), ('dinner jacket','top'),('maxi dress','dress')] colors = ['red','purple','gold','white','off white','black','beige','cream','navy','maroon'] patterns = ['plain'] elif theme == 'Sports': types = [ ('track shorts','bottom'), ('track pants','bottom'), ('track jacket','top'), ('Compression top','top'),('cycling shorts','bottom'),('compression shorts','bottom'),('tank top','top')] colors = colors_list.copy() patterns = patterns_list.copy()# elif theme =='My preference': types = [('evening dress','dress'), ('tuxedo','top'), ('suit jacket','top'), ('dinner jacket','top'),('maxi dress','dress')] colors = ['red','purple','gold'] patterns = ['plain','zebra print'] else: return gr.update(label='Chosen theme: None') for type_,c_type in types: for color in colors: for pattern in patterns: new_files.extend(clothes_tree[c_type][type_][color][pattern]) return gr.update(label='Chosen theme: '+theme) def segment(img): ''' Purpose: Segment the image to get the mask Inputs: img(pil image): image to be segmented Outputs: img(pil image): original image arr(numpy array): array of image pred_seg(tensor): mask ''' # Get the segmentation mask then umsample it to the original size encoding = extractor_hg(img.convert('RGB'), return_tensors="pt") pixel_values = encoding.pixel_values.to(device) outputs = seg_hg(pixel_values=pixel_values) logits = outputs.logits.cpu() upsampled_logits = nn.functional.interpolate( logits, size=img.size[::-1], mode="bilinear", align_corners=False, ) pred_seg = upsampled_logits.argmax(dim=1)[0] arr_img = np.array(img) return img, arr_img, pred_seg def clean_img(img): ''' Purpose: Clean the image to remove the chosen items Inputs: img(numpy array): image to be cleaned Outputs: img(numpy array): cleaned image ''' # Here we remove pixels whihc are not in our desired class global global_mask global mask_choice bad = [] mask_size = global_mask.shape img_size = img.shape[:2] if img_size != mask_size: return img if mask_choice=='Person': bad.append(0) elif mask_choice=='Clothes': bad.extend([0,2,15,14,13,12,11]) elif mask_choice=='Upper Body/Dress': bad.extend([0,5,6,9,10,12,13,16]) elif mask_choice=='Lower Body': bad.extend([0,1,2,3,4,7,8,11,14,15,16]) elif mask_choice=='Upper Body/Dress, no person': bad.extend([0,1,2,15,11,14,5,6,9,10,12,13,16,3]) for i in bad: global_mask[global_mask==i] = 50 img[global_mask==50] = 255 return img def label_to_rec_lables (label): ''' Purpose: Use the label to get the corresponding reccomendation labels Inputs: label(string): label of the image Outputs: rec_labels(list): list of reccomendation labels ''' # This function is used to get the reccomendation labels to then # filter the reccomendation search to them labels = label.split(',') new_label = rec_dic[','.join(labels[:2])] print('Reccomendation label: ',new_label) n = 5 if len(new_label) >= 5 else len(new_label) labels = [] labels = [new_label[i][0].split(',') for i in range(n)] chosen = [] c_types = ['top','bottom','dress'] for item in labels: label_type = item[0] label_color = item[1].strip() for c_type in c_types: if label_type in all_types[c_type]: item_type = c_type chosen.append([item_type,label_type,label_color]) print('Chosen: ',chosen) return chosen def filter_features(labels, rec=False, rec_items=None): ''' Purpose: Filter the features to only contain the chosen label Inputs: labels(str): label string rec(bool): if the function is called from the recommendation function rec_items(list): list containing the label info Outputs: url_list(list): list of urls after filtering features(tensor): features after filtering ''' global url_list_main global features_main # Here we filter the features to only contain the desired labels and # also provide the new url list labels = labels.split(',') label_type = labels[0] label_color = labels[1].strip() c_types = ['top', 'bottom', 'dress'] for c_type in c_types: if label_type in all_types[c_type]: item_type = c_type new_list = set() if rec: item_type = rec_items[0] label_type = rec_items[1] label_color = rec_items[2] for pattern in patterns_list: new_list.update(clothes_tree[item_type][label_type][label_color][pattern]) else: #for color in colors_list: color = label_color for pattern in patterns_list: new_list.update(clothes_tree[item_type][label_type][color][pattern]) new_files = list(new_list) temp_url = [] temp_features = torch.zeros(len(new_files), 512).to(device) for c,i in enumerate(new_files): temp_url.append(url_list_main[i]) temp_features[c] = features_main[i] url_list = temp_url features = temp_features.to(torch.float32) return url_list, features def get_image_from_url(idx,url_list,items=5): ''' Purpose: Get a list of images from the url list using the indecies Inputs: idx(list): list of indecies url_list(list): list of urls items(int): number of images to return Outputs: images(list): list of images ''' # Looping until we have the desired number of images res = [] i = 0 n = 15 if len(idx) > 15 else len(idx) while len(res) != items and i != n: try: req = requests.get(url_list[idx[i]],stream=True,timeout=5) img = Image.open(req.raw).convert('RGB') img = np.array(img) res.append(img) i += 1 except: print('Error with: ' + url_list[i]) i += 1 continue return res def get_label(img): ''' Purpose: Get the label of the image Inputs: img(numpy array or pil image): image to get label of Outputs: label(string): label of the image ''' img_features = processor_hg(images=img, return_tensors="pt", padding=True).to(device) with torch.no_grad(): img_features = clip_hg.get_image_features(**img_features) idx = find_closest(img_features,item_embeddings)[0] label = all_items[idx] return label def resize_img(img,thresh=384): ''' Purpose: Resize the image to have the largest dimension be thresh Inputs: img(pil image): image to resize thresh(int): threshold for the largest dimension Outputs: img(pil image): resized image ''' size = img.size larger_dim = 0 if size[0] > size[1] else 1 if size[larger_dim] > thresh: size = (int(size[0] * thresh / size[larger_dim]), int(size[1] * thresh / size[larger_dim])) img = img.resize(size) return img def segment_function(choice): ''' Purpose: Set the mask choice so that it can be called during search Inputs: choice(string): mask choice Outputs: None ''' global mask_choice mask_choice = choice return gr.update(label =f'Selection: {choice}') def rec_function(option): ''' Purpose: using an image to get a reccomendation return that image and the reccomendations Inputs: option(int): option to use Outputs: rec_out(list): list of images temp_out(numpy array): choice image ''' global out global url_list_main global features_main # Here we get the items which should be reccomended based on the # chosen image. Then we find the closest items to the chosen image # out of the reccomended items. Finally we crop the images so that # we only see the reccomended items in the output if not out: req = requests.get(url,stream=True) img = np.array(Image.open(req.raw).convert('RGB')) rec_out = [img]*5 return rec_out img = Image.fromarray(out[option]) choice_img = resize_img(img) label = get_label(choice_img) target_labels = label_to_rec_lables(label) temp_out = [] img_features = processor_hg(images=choice_img, return_tensors="pt", padding=True).to(device) with torch.no_grad(): img_features = clip_hg.get_image_features(**img_features) n = len(target_labels) if n == 1: return_items = 5 elif n == 2: return_items = 3 elif n == 3: return_items = 2 else: return_items = 1 for item in target_labels: url_list, features = filter_features(label, rec=True, rec_items=item) idx = find_closest(img_features, features)[:5] temp_out.extend(get_image_from_url(idx,url_list,items=return_items)) rec_out = [] for temp_img in temp_out: temp_img = resize_img(Image.fromarray(temp_img)) img, seg_img, out_mask = segment(temp_img) label_type = label.split(',')[0].strip() bad = [] if label_type in top_type or label_type in dress_type: bad.extend([0,1,2,3,4,7,8,11,14,15,16]) elif label_type in bottom_type: bad.extend([0,5,6,9,10,12,13,16]) for i in bad: out_mask[out_mask==i] = 50 img = np.array(img) img[out_mask==50] = 255 h, w = img.shape[:2] # find the highest and lowest y-coordinates where the pixel is not white top = 0 bottom = h for i in range(h): if np.all(img[i] == 255): top = i else: break for i in range(h-1, 0, -1): if np.all(img[i] == 255): bottom = i else: break # find the highest and lowest x-coordinates where the pixel is not white left = 0 right = w for i in range(w): if np.all(img[:, i] == 255): left = i else: break for i in range(w-1, 0, -1): if np.all(img[:, i] == 255): right = i else: break # crop the image # add 10 pixels to the top and bottom if those are not the edges of the image if top - 10 > 0: top -= 10 if bottom + 10 < h: bottom += 10 # add 10 pixels to the left and right if those are not the edges of the image if left - 10 > 0: left -= 10 if right + 10 < w: right += 10 if top > bottom or right < left: rec_out.append(temp_img) else: temp_img = np.array(temp_img) img = temp_img[top:bottom, left:right] rec_out.append(img) temp_out = [choice_img] return rec_out, temp_out def reset_values(): ''' Purpose: reset the values of the global variables Inputs: None Outputs: None ''' global global_mask global out global mask_choice global clicks global c_types global types global colors global patterns global new_files global_mask = None out = None mask_choice = None clicks = 0 c_types.clear() types.clear() colors.clear() patterns.clear() new_files.clear() return [gr.update(choices=['top','bottom','dress'],value=[]),gr.update(choices=['Person','Clothes','Upper Body/Dress','Upper Body/Dress, no person','Lower Body'],value=None) ,gr.update(value=None), gr.update(value=[]),gr.update(value=[]),gr.update(value=0)] def search_function(img, text, use_choice,use_label): ''' Purpose: search for images based on the text input or image input Inputs: img(pil image): image input text(string): text input use_choice(boolean): Boolen to know if to use image or text use_label(boolean): whether to use the label Outputs: out(list): list of images ''' global new_files global global_mask global out use_img = False use_text = False if use_choice == 'Use Image': use_img = True elif use_choice == 'Use Text': use_text = True if new_files: global url_list_main global features_main temp_url = [] new_files = list(set(new_files)) temp_features = torch.zeros(len(new_files), 512).to(device) for c,i in enumerate(new_files): temp_url.append(url_list_main[i]) temp_features[c] = features_main[i] url_list = temp_url features = temp_features.to(torch.float32) else: features = features_main.clone() url_list = url_list_main.copy() if use_text and not use_img: text_features = processor_hg(text=text, return_tensors="pt", padding=True).to(device) with torch.no_grad(): text_features = clip_hg.get_text_features(**text_features) idx = find_closest(text_features, features)[:15] out = get_image_from_url(idx,url_list) else : if not isinstance(global_mask,type(None)): seg_img = clean_img(img) else: seg_img = img img = Image.fromarray(seg_img) label = get_label(img) print(label) if not new_files and use_label: url_list, features = filter_features(label) img_features = processor_hg(images=img, return_tensors="pt", padding=True).to(device) with torch.no_grad(): img_features = clip_hg.get_image_features(**img_features) idx = find_closest(img_features, features)[:15] out = get_image_from_url(idx,url_list) if use_img: out.pop() out.insert(0, seg_img) return out def search(img,text, choice,use_label,rotation): global global_mask try: img = Image.fromarray(img).convert('RGB') except: img = Image.open(requests.get(url, stream=True).raw).convert('RGB') img = img.rotate(rotation) img = resize_img(img) pil, img, out_mask = segment(img) global_mask = out_mask res = search_function(img, text, choice,use_label) return res # Define the app layout with gr.Blocks() as demo: gr.Markdown("Search using image segmentation") with gr.Tab("Search"): with gr.Row(): search_image = gr.Image() search_input = [search_image,gr.Textbox(lines=2, label="Search Text")] with gr.Column(): search_type = gr.Radio(choices=['Use Image','Use Text'],label='Select the type of search you want to perform',value='Use Image') use_label = gr.Checkbox(label="Use Label",value=True) image_output = [gr.Gallery(label='Outputs')] rec_out = [gr.Gallery(label='Recommendations',interactive=True)] with gr.Row(): rec_selector = gr.Radio(label='Select which item you want a recommendation for',choices = [1,2,3,4],value=1) rec_button = gr.Button("Get Recommendation") with gr.Row(): clothes_selector = gr.Radio(label='Choose a segmentation', choices=['Person','Clothes','Upper Body/Dress','Upper Body/Dress, no person','Lower Body'],interactive=True) theme_radio = gr.Radio(label='Choose a theme',choices=['None','Red carpet','Sports'],interactive=True) rotation_radio = gr.Radio(label='Choose a rotation',choices=[0,90,180,270],interactive=True,value=0) with gr.Row(): filter_checkbox = gr.CheckboxGroup(label='Choose the clothing types', choices=['top','bottom','dress'],interactive=True,value=['top']) filter_button = gr.Button("Filter Button") search_button = gr.Button("Search Button") clothes_selector.change(segment_function,inputs=[clothes_selector],outputs=clothes_selector) search_image.change(reset_values, inputs=None, outputs=[filter_checkbox,clothes_selector,theme_radio,image_output[0],rec_out[0],rotation_radio]) theme_radio.change(set_theme, inputs=theme_radio, outputs=theme_radio) rec_button.click(rec_function, inputs=rec_selector, outputs=[rec_out[0],image_output[0]]) filter_button.click(filter_function, inputs=filter_checkbox, outputs=filter_checkbox) search_button.click(search, inputs=search_input+[search_type,use_label,rotation_radio], outputs=image_output) demo.launch(share=False)