from glob import glob from os.path import join as pjoin import numpy as np import pandas as pd import plotly.express as px import plotly.graph_objects as go import torch import umap.umap_ as umap from PIL import Image from scipy.cluster.hierarchy import dendrogram, linkage from scipy.spatial.distance import squareform from sklearn.preprocessing import normalize from tqdm import tqdm ### # Get text embeddings from sentence-transformers model ### def sentence_mean_pooling(model_output, attention_mask): token_embeddings = model_output[ 0 ] # First element of model_output contains all token embeddings input_mask_expanded = ( attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float() ) return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp( input_mask_expanded.sum(1), min=1e-9 ) def compute_text_embeddings(sentences, text_tokenizer, text_model): batch = text_tokenizer( sentences, padding=True, truncation=True, return_tensors="pt" ) with torch.no_grad(): model_output = text_model(**batch) sentence_embeds = sentence_mean_pooling(model_output, batch["attention_mask"]) sentence_embeds /= sentence_embeds.norm(dim=-1, keepdim=True) return sentence_embeds ### # Get image embeddings from BLIP VQA models ### # returns the average pixel embedding from the last layer of the image encoder def get_compute_image_embedding_blip_vqa_pixels( img, blip_processor, blip_model, device="cpu" ): pixel_values = blip_processor(img, "", return_tensors="pt")["pixel_values"].to( device ) with torch.no_grad(): vision_outputs = blip_model.vision_model( pixel_values=pixel_values, output_hidden_states=True, ) image_embeds = vision_outputs[0].sum(dim=1).squeeze() image_embeds /= image_embeds.norm() return image_embeds.detach().cpu().numpy() # returns the average token embedding from the question encoder (conditioned on the image) along with the generated answer # adapted from: # https://github.com/huggingface/transformers/blob/2411f0e465e761790879e605a4256f3d4afb7f82/src/transformers/models/blip/modeling_blip.py#L1225 def get_compute_image_embedding_blip_vqa_question( img, blip_processor, blip_model, question=None, device="cpu" ): question = "what word best describes this person?" if question is None else question inputs = blip_processor(img, question, return_tensors="pt") with torch.no_grad(): # make question embeddings vision_outputs = blip_model.vision_model( pixel_values=inputs["pixel_values"].to(device), output_hidden_states=True, ) image_embeds = vision_outputs[0] image_attention_mask = torch.ones(image_embeds.size()[:-1], dtype=torch.long) question_embeds = blip_model.text_encoder( input_ids=inputs["input_ids"].to(device), attention_mask=inputs["attention_mask"].to(device), encoder_hidden_states=image_embeds, encoder_attention_mask=image_attention_mask, return_dict=False, ) question_embeds = question_embeds[0] # generate outputs question_attention_mask = torch.ones( question_embeds.size()[:-1], dtype=torch.long ).to(question_embeds.device) bos_ids = torch.full( (question_embeds.size(0), 1), fill_value=blip_model.decoder_bos_token_id, device=question_embeds.device, ) outputs = blip_model.text_decoder.generate( input_ids=bos_ids, eos_token_id=blip_model.config.text_config.sep_token_id, pad_token_id=blip_model.config.text_config.pad_token_id, encoder_hidden_states=question_embeds, encoder_attention_mask=question_attention_mask, # **generate_kwargs, ) answer = blip_processor.decode(outputs[0], skip_special_tokens=True) # average and normalize question embeddings res_question_embeds = question_embeds.sum(dim=1).squeeze() res_question_embeds /= res_question_embeds.norm() res_question_embeds = res_question_embeds.detach().cpu().numpy() return (res_question_embeds, answer) ### # Plotting utilities: 2D and 3D projection + scatter plots ### def make_2d_plot(embeds, text_list, color_list=None, shape_list=None, umap_spread=10): # default color and shape color_list = [0 for _ in text_list] if color_list is None else color_list shape_list = ["circle" for _ in text_list] if shape_list is None else shape_list # project to 2D fit = umap.UMAP( metric="cosine", n_neighbors=len(embeds) - 1, min_dist=1, n_components=2, spread=umap_spread, ) u = fit.fit_transform(embeds) fig = go.Figure() fig.add_trace( go.Scatter( x=u[:, 0].tolist(), y=u[:, 1].tolist(), mode="markers", name="nodes", marker=dict( symbol=shape_list, color=color_list, ), text=text_list, hoverinfo="text", marker_line_color="midnightblue", marker_line_width=2, marker_size=10, opacity=0.8, ) ) fig.update_yaxes( scaleanchor="x", scaleratio=1, ) fig.update_layout( autosize=False, width=800, height=800, ) fig.layout.showlegend = False return fig def make_3d_plot(embeds, text_list, color_list=None, shape_list=None, umap_spread=10): # default color and shape color_list = [0 for _ in text_list] if color_list is None else color_list shape_list = ["circle" for _ in text_list] if shape_list is None else shape_list # project to 3D fit = umap.UMAP( metric="cosine", n_neighbors=len(embeds) - 1, min_dist=1, n_components=3, spread=umap_spread, ) u = fit.fit_transform(embeds) # make nodes df = pd.DataFrame( { "x": u[:, 0].tolist(), "y": u[:, 1].tolist(), "z": u[:, 2].tolist(), "color": color_list, "hover": text_list, "symbol": shape_list, "size": [5 for _ in text_list], } ) fig = px.scatter_3d( df, x="x", y="y", z="z", color="color", symbol="symbol", size="size", hover_data={ "hover": True, "x": False, "y": False, "z": False, "color": False, "symbol": False, "size": False, }, ) fig.layout.showlegend = False return fig ### # Plotting utilities: cluster re-ordering and heatmaps ### ### Some utility functions to get the similarities between two lists of arrays # average pairwise similarity def sim_pairwise_avg(vecs_1, vecs_2): res = np.matmul(np.array(vecs_1), np.array(vecs_2).transpose()).mean() return res # distance between (normalized) centroids def sim_centroids(vecs_1, vecs_2): res = np.dot( normalize(np.array(vecs_1).mean(axis=0, keepdims=True), norm="l2")[0], normalize(np.array(vecs_2).mean(axis=0, keepdims=True), norm="l2")[0], ) return res # distance to nearest neighbot/examplar def sim_pairwise_examplar(vecs_1, vecs_2): res = np.matmul(np.array(vecs_1), np.array(vecs_2).transpose()).max() return res # To make pretty heatmaps, similar rows need to be close to each other # we achieve that by computing a hierarchical clustering of the points # then ordering the items as the leaves of a dendrogram def get_cluster_order(similarity_matrix, label_names=None): label_names = ( ["" for _ in range(similarity_matrix.shape[0])] if label_names is None else label_names ) dissimilarity = 1 - similarity_matrix np.fill_diagonal(dissimilarity, 0.0) # checks = False because similarity checks can fail for torch to numpy conversion Z = linkage(squareform(dissimilarity, checks=False), "average") # no_plot when inside a function call required because of jupyter/matplotlib issue ddgr = dendrogram( Z, labels=label_names, orientation="top", leaf_rotation=90, no_plot=True ) cluster_order = ddgr["leaves"] return cluster_order # then make heat map from similarity matrix def make_heat_map(sim_matrix, labels_x, labels_y, scale=25): fig = go.Figure( data=go.Heatmap(z=sim_matrix, x=labels_x, y=labels_y, hoverongaps=False) ) fig.update_yaxes( scaleanchor="x", scaleratio=1, ) fig.update_layout( autosize=False, width=scale * len(labels_x), height=scale * len(labels_y), ) fig.layout.showlegend = False return fig # bring things together for a square heatmap def build_heat_map_square( img_list, embed_field, sim_fun, label_list, row_order=None, hm_scale=20 ): sim_mat = np.zeros((len(img_list), len(img_list))) for i, dct_i in enumerate(img_list): for j, dct_j in enumerate(img_list): sim_mat[i, j] = sim_fun(dct_i[embed_field], dct_j[embed_field]) # optionally reorder labels and similarity matrix to be prettier if row_order is None: row_order = get_cluster_order(sim_mat) labels_sorted = [label_list[i] for i in row_order] sim_mat_sorted = sim_mat[np.ix_(row_order, row_order)] # make heatmap from similarity matrix heatmap_fig = make_heat_map( sim_mat_sorted, labels_sorted, labels_sorted, scale=hm_scale ) return heatmap_fig # bring things together for a rectangle heatmap: across lists def build_heat_map_rect( img_list_rows, img_list_cols, label_list_rows, label_list_cols, embed_field, sim_fun, center=False, temperature=5, hm_scale=20, ): sim_mat = np.zeros((len(img_list_rows), len(img_list_cols))) for i, dct_i in enumerate(img_list_rows): for j, dct_j in enumerate(img_list_cols): sim_mat[i, j] = sim_fun(dct_i[embed_field], dct_j[embed_field]) # normalize and substract mean sim_mat_exp = np.exp(temperature * sim_mat) sim_mat_exp /= sim_mat_exp.sum(axis=1, keepdims=1) if center: sim_mat_exp_avg = sim_mat_exp.mean(axis=0, keepdims=1) sim_mat_exp -= sim_mat_exp_avg sim_mat_exp_avg = sim_mat_exp_avg * sim_mat_exp.max() / sim_mat_exp_avg.max() # rows are reordered by decreasing norm, sim_mat_norm = np.sum(sim_mat_exp * sim_mat_exp, axis=1) row_order = np.argsort(sim_mat_norm, axis=-1) row_labels_sorted = [label_list_rows[i] for i in row_order] if center: # columns are ordered by bias col_order = np.argsort(sim_mat_exp_avg.sum(axis=0), axis=-1) else: # columns sre reordered by similarity sim_mat_exp_norm = normalize(sim_mat_exp, axis=0, norm="l2") cluster_cols_sim_mat = np.matmul(sim_mat_exp_norm.transpose(), sim_mat_exp_norm) col_order = get_cluster_order(cluster_cols_sim_mat) col_labels_sorted = [label_list_cols[i] for i in col_order] # make heatmap from similarity matrix if center: row_order = list(row_order) + [len(row_order), len(row_order) + 1] row_labels_sorted = row_labels_sorted + ["_", "AVERAGE"] sim_mat_exp = np.concatenate( [sim_mat_exp, np.zeros_like(sim_mat_exp_avg), sim_mat_exp_avg], axis=0 ) sim_mat_exp_sorted = sim_mat_exp[np.ix_(row_order, col_order)] heatmap_fig = make_heat_map( sim_mat_exp_sorted, col_labels_sorted, row_labels_sorted, scale=hm_scale ) return heatmap_fig