import gradio as gr from sklearn.metrics.pairwise import cosine_similarity from scipy.sparse import csr_matrix import numpy as np from joblib import load import h5py from io import BytesIO import csv import re import random import compress_fasttext from collections import OrderedDict from lark import Lark from lark import Token from lark.exceptions import ParseError faq_content=""" # Questions: ## What is the purpose of this tool? When you enter a txt2img prompt and press the "submit" button, the Tagset Completer parses your prompt and checks that all your tags are valid e621 tags. If it finds any that are not, it recommends some valid e621 tags you can use to replace them in the "Unseen Tags" table. Additionally, in the "Top Artists" text box, it lists the artists who would most likely draw an image having the set of tags you provided, in case you want to look them up to get more ideas. This is useful to align your prompt with the expected input to an e621-trained model. ## Does input order matter? No ## Should I use underscores or spaces in the input tags? Spaces are preferred, but it will still work if you use underscores. The Unseen Tags table will just complain at you. ## Can I use parentheses or weights as in the Stable Diffusion Automatic1111 WebUI? Yes, but only '(' and ')' and numerical weights, and all of these things are ignored in all calculations. The main benefit of this is that you can copy/paste prompts from one program to another with minimal editing. An example that illustrates acceptable parentheses and weight formatting is: ((sunset over the mountains)), (clear sky:1.5), ((eagle flying high:2.0)), river, (fish swimming in the river:1.2), (campfire, (marshmallows:2.1):1.3), stars in the sky, ((full moon:1.8)), (wolf howling:1.7) ## Why are some valid tags marked as "unseen", and why don't some artists ever get returned? Some data is excluded from consideration if it did not occur frequently enough in the sample from which the application makes its calculations. If an artist or tag is too infrequent, we might not think we have enough data to make predictions about it. ## Are there any special tags? Yes. We normalized the favorite counts of each image to a range of 0-9, with 0 being the lowest favcount, and 9 being the highest. You can include any of these special tags: "score:0", "score:1", "score:2", "score:3", "score:4", "score:5", "score:6", "score:7", "score:8", "score:9" in your list to bias the output toward artists with higher or lower scoring images. Since they are not real tags, the Unseen Tags section will complain, but you can ignore that. ## Are there any other special tricks? Yes. If you want to more strongly bias the artist output toward a specific tag, you can just list it multiple times. So for example, the query "red fox, red fox, red fox, score:7" will yield a list of artists who are more strongly associated with the tag "red fox" than the query "red fox, score:7". ## Why is this space tagged "not-for-all-audience" The "not-for-all-audience" tag informs users that this tool's text output is derived from e621.net data for tag prediction and completion. The app will try not to display nsfw tags unless the "Allow NSFW Tags" is checked, but the filter is not perfect. ## How is the artist list calculated? Each artist is represented by a "pseudo-document" composed of all the tags from their uploaded images, treating these tags similarly to words in a text document. Similarly, when you input a set of tags, the system creates a pseudo-document for your query out of all the tags. It then uses a technique called cosine similarity to compare your tags against each artist's collection, essentially finding which artist's tags are most "similar" to yours. This method helps identify artists whose work is closely aligned with the themes or elements you're interested in. For those curious about the underlying mechanics of comparing text-like data, we employ the TF-IDF (Term Frequency-Inverse Document Frequency) method, a standard approach in information retrieval. You can read more about TF-IDF on its [Wikipedia page](https://en.wikipedia.org/wiki/Tf%E2%80%93idf). ## How does the tag corrector work? We collect the tag sets from over 4 million e621 posts, treating the tag set from each image as an individual document. We then randomly replace about 10% of the tags in each document with a randomly selected alias from e621's list of aliases for the tag (e.g. "canine" gets replaced with one of {k9,canines,mongrel,cannine,cnaine,feral_canine,anthro_canine}). We then train a FastText (https://fasttext.cc/) model on the documents. The result of this training is a function that maps arbitrary words to vectors such that the vector for a tag and the vectors for its aliases are all close together (because the model has seen them in similar contexts). Since the lists of aliases contain misspellings and rephrasings of tags, the model should be robust to these kinds of problems as long as they are not too dissimilar from the alias lists. To enhance the tag corrector further, we leverage conditional probabilities to refine our predictions. Using the same 4 million post dataset, we calculate the conditional probability of each tag given the context of other tags appearing within the same document. This is done by creating a co-occurrence matrix from our dataset, which records how frequently each pair of tags appears together across all documents. By considering the context in which tags are used, we can now not only correct misspellings and rephrasings but also make more contextually relevant suggestions. The "similarity weight" slider controls how much weight these conditional probabilities are given vs how much weight the FastText similarity model is given when suggesting replacements for invalid tags. A similarity weight slider value of 0 means that only the FastText model's predictions will be used to calculate similarity scores, and a value of 1 means only the conditional probabilities are used (although the FastText model is still used to trim the list of candidates). """ nsfw_threshold = 0.95 # Assuming the threshold value is defined here grammar=r""" !start: (prompt | /[][():]/+)* prompt: (emphasized | plain | comma | WHITESPACE)* !emphasized: "(" prompt ")" | "(" prompt ":" [WHITESPACE] NUMBER [WHITESPACE] ")" comma: "," WHITESPACE: /\s+/ plain: /([^,\\\[\]():|]|\\.)+/ %import common.SIGNED_NUMBER -> NUMBER """ # Initialize the parser parser = Lark(grammar, start='start') # Function to extract tags def extract_tags(tree): tags = [] def _traverse(node): if isinstance(node, Token) and node.type == '__ANON_1': tags.append(node.value.strip()) elif not isinstance(node, Token): for child in node.children: _traverse(child) _traverse(tree) return tags special_tags = ["score:0", "score:1", "score:2", "score:3", "score:4", "score:5", "score:6", "score:7", "score:8", "score:9"] def remove_special_tags(original_string): tags = [tag.strip() for tag in original_string.split(",")] remaining_tags = [tag for tag in tags if tag not in special_tags] removed_tags = [tag for tag in tags if tag in special_tags] return ", ".join(remaining_tags), removed_tags # Load the model and data once at startup with h5py.File('complete_artist_data.hdf5', 'r') as f: # Deserialize the vectorizer vectorizer_bytes = f['vectorizer'][()].tobytes() # Use io.BytesIO to convert bytes back to a file-like object for joblib to load vectorizer_buffer = BytesIO(vectorizer_bytes) vectorizer = load(vectorizer_buffer) # Load X_artist X_artist = f['X_artist'][:] # Load artist names and decode to strings artist_names = [name.decode() for name in f['artist_names'][:]] with h5py.File('conditional_tag_probabilities_matrix.h5', 'r') as f: # Reconstruct the sparse co-occurrence matrix conditional_co_occurrence_matrix = csr_matrix( (f['co_occurrence_data'][:], f['co_occurrence_indices'][:], f['co_occurrence_indptr'][:]), shape=f['co_occurrence_shape'][:] ) # Reconstruct the vocabulary conditional_words = f['vocabulary_words'][:] conditional_indices = f['vocabulary_indices'][:] conditional_vocabulary = {key.decode('utf-8'): value for key, value in zip(conditional_words, conditional_indices)} # Load the document count conditional_doc_count = f['doc_count'][()] conditional_smoothing = 100. / conditional_doc_count nsfw_tags = set() # Initialize an empty set to store words meeting the threshold # Open and read the CSV file with open("word_rating_probabilities.csv", 'r', newline='', encoding='utf-8') as csvfile: reader = csv.reader(csvfile) next(reader, None) # Skip the header row for row in reader: word = row[0] # The word is in the first column probability_sum = float(row[1]) # The sum of probabilities is in the second column, convert to float for comparison # Check if the probability sum meets the threshold and add the word to the set if it does if probability_sum >= nsfw_threshold: nsfw_tags.add(word) def clean_tag(tag): return ''.join(char for char in tag if ord(char) < 128) #Normally returns tag to aliases, but when reverse=True, returns alias to tags def build_aliases_dict(filename, reverse=False): aliases_dict = {} with open(filename, 'r', newline='', encoding='utf-8') as csvfile: reader = csv.reader(csvfile) for row in reader: tag = clean_tag(row[0]) alias_list = [] if row[3] == "null" else [clean_tag(alias) for alias in row[3].split(',')] if reverse: for alias in alias_list: aliases_dict.setdefault(alias, []).append(tag) else: aliases_dict[tag] = alias_list return aliases_dict #Imagine we are adding smoothing_value to the number of times word_j occurs in each document for smoothing. #Note the intention is that sum_i(P(word_i|word_j)) =(approx) # of words in a document rather than 1. def conditional_probability(word_i, word_j, co_occurrence_matrix, vocabulary, doc_count, smoothing_value=0.01): word_i_index = vocabulary.get(word_i) word_j_index = vocabulary.get(word_j) if word_i_index is not None and word_j_index is not None: # Directly access the sparse matrix elements word_j_count = co_occurrence_matrix[word_j_index, word_j_index] smoothed_word_j_count = word_j_count + (smoothing_value * doc_count) word_i_count = co_occurrence_matrix[word_i_index, word_i_index] co_occurrence_count = co_occurrence_matrix[word_i_index, word_j_index] smoothed_co_occurrence_count = co_occurrence_count + (smoothing_value * word_i_count) # Calculate the conditional probability with smoothing conditional_prob = smoothed_co_occurrence_count / smoothed_word_j_count return conditional_prob elif word_i_index is None: return 0 else: return None #geometric_mean_given_words(target_word, context_words, conditional_co_occurrence_matrix, conditioanl_vocabulary, conditional_doc_count, smoothing_value=conditional_smoothing): def geometric_mean_given_words(target_word, context_words, co_occurrence_matrix, vocabulary, doc_count, smoothing_value=0.01): probabilities = [] # Collect the conditional probabilities of the target word given each context word, ignoring None values for context_word in context_words: prob = conditional_probability(target_word, context_word, co_occurrence_matrix, vocabulary, doc_count, smoothing_value) if prob is not None: probabilities.append(prob) # Compute the geometric mean of the probabilities, avoiding division by zero if probabilities: # Check if the list is not empty geometric_mean = np.prod(probabilities) ** (1.0 / len(probabilities)) else: geometric_mean = 0.5 # Or assign some default value if all probabilities are None return geometric_mean def find_similar_tags(test_tags, similarity_weight, allow_nsfw_tags): #Initialize stuff if not hasattr(find_similar_tags, "fasttext_small_model"): find_similar_tags.fasttext_small_model = compress_fasttext.models.CompressedFastTextKeyedVectors.load('e621FastTextModel010Replacement_small.bin') tag_aliases_file = 'fluffyrock_3m.csv' if not hasattr(find_similar_tags, "tag2aliases"): find_similar_tags.tag2aliases = build_aliases_dict(tag_aliases_file) if not hasattr(find_similar_tags, "alias2tags"): find_similar_tags.alias2tags = build_aliases_dict(tag_aliases_file, reverse=True) transformed_tags = [tag.replace(' ', '_') for tag in test_tags] # Find similar tags and prepare data for dataframe. results_data = [] for tag in test_tags: if tag in special_tags: continue modified_tag_for_search = tag.replace(' ','_') similar_words = find_similar_tags.fasttext_small_model.most_similar(modified_tag_for_search, topn = 100) result, seen = [], set(transformed_tags) if modified_tag_for_search in find_similar_tags.tag2aliases: if tag in find_similar_tags.tag2aliases and "_" in tag: #Implicitly tell the user that they should get rid of the underscore result.append(modified_tag_for_search.replace('_',' '), 1) seen.add(tag) else: #The user correctly did not put underscores in their tag continue else: for item in similar_words: similar_word, similarity = item if similar_word not in seen: if similar_word in find_similar_tags.tag2aliases: result.append((similar_word.replace('_', ' '), round(similarity, 3))) seen.add(similar_word) else: for similar_tag in find_similar_tags.alias2tags.get(similar_word, []): if similar_tag not in seen: result.append((similar_tag.replace('_', ' '), round(similarity, 3))) seen.add(similar_tag) #Remove NSFW tags if appropriate. if not allow_nsfw_tags: result = [(word, score) for word, score in result if word.replace(' ','_') not in nsfw_tags] #Adjust score based on context for i in range(len(result)): word, score = result[i] # Unpack the tuple geometric_mean = geometric_mean_given_words(word.replace(' ','_'), [context_tag for context_tag in transformed_tags if context_tag != word and context_tag != tag], conditional_co_occurrence_matrix, conditional_vocabulary, conditional_doc_count, smoothing_value=conditional_smoothing) adjusted_score = (similarity_weight * geometric_mean) + ((1-similarity_weight)*score) # Apply the adjustment function result[i] = (word, adjusted_score) # Update the tuple with the adjusted score # Append tag and formatted similar tags to results_data result = sorted(result, key=lambda x: x[1], reverse=True)[:10] first_entry_for_tag = True for word, sim in result: if first_entry_for_tag: results_data.append([tag, word, sim]) first_entry_for_tag = False else: results_data.append(["", word, sim]) results_data.append(["", "", ""]) # Adds a blank line after each group of tags if not results_data: results_data.append(["No Unknown Tags Found", "", ""]) return results_data # Return list of lists for Dataframe def find_similar_artists(new_tags_string, top_n, similarity_weight, allow_nsfw_tags): try: new_tags_string = new_tags_string.lower() new_tags_string, removed_tags = remove_special_tags(new_tags_string) # Parse the prompt parsed = parser.parse(new_tags_string) # Extract tags from the parsed tree new_image_tags = extract_tags(parsed) new_image_tags = [tag.replace('_', ' ').replace('\\(', '(').replace('\\)', ')').strip() for tag in new_image_tags] ###unseen_tags = list(set(OrderedDict.fromkeys(new_image_tags)) - set(vectorizer.vocabulary_.keys())) #We may want this line again later. These are the tags that were not used to calculate the artists list. unseen_tags_data = find_similar_tags(new_image_tags, similarity_weight, allow_nsfw_tags) X_new_image = vectorizer.transform([','.join(new_image_tags + removed_tags)]) similarities = cosine_similarity(X_new_image, X_artist)[0] top_artist_indices = np.argsort(similarities)[-top_n:][::-1] top_artists = [(artist_names[i], similarities[i]) for i in top_artist_indices] top_artists_str = "\n".join([f"{rank+1}. {artist[3:]} ({score:.4f})" for rank, (artist, score) in enumerate(top_artists)]) dynamic_prompts_formatted_artists = "{" + "|".join([artist for artist, _ in top_artists]) + "}" return unseen_tags_data, top_artists_str, dynamic_prompts_formatted_artists except ParseError as e: return [], "Parse Error: Check for mismatched parentheses or something", "" iface = gr.Interface( fn=find_similar_artists, inputs=[ gr.Textbox(label="Enter image tags", placeholder="e.g. fox, outside, detailed background, ..."), gr.Slider(minimum=1, maximum=100, value=10, step=1, label="Number of artists"), gr.Slider(minimum=0, maximum=1, value=0.5, step=0.1, label="Similarity weight"), gr.Checkbox(label="Allow NSFW Tags", value=False) ], outputs=[ gr.Dataframe(label="Unseen Tags", headers=["Tag", "Similar Tags", "Similarity"]), gr.Textbox(label="Top Artists", info="These are the artists most strongly associated with your tags. The number in parenthes is a similarity score between 0 and 1, with higher numbers indicating greater similarity."), gr.Textbox(label="Dynamic Prompts Format", info="For if you're using the Automatic1111 webui (https://github.com/AUTOMATIC1111/stable-diffusion-webui) with the Dynamic Prompts extension activated (https://github.com/adieyal/sd-dynamic-prompts) and want to try them all individually.") ], title="Tagset Completer", description="Enter a list of comma-separated e6 tags", article=faq_content ) iface.launch()