# Copyright 2021 The HuggingFace Team. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import math from os.path import exists from os.path import join as pjoin import plotly.graph_objects as go import torch import transformers from datasets import load_from_disk from plotly.io import read_json from tqdm import tqdm from .dataset_utils import EMBEDDING_FIELD def sentence_mean_pooling(model_output, attention_mask): """Mean pooling of token embeddings for a sentence.""" 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 ) class Embeddings: def __init__( self, dstats=None, text_dset=None, text_field_name="text", cache_path="", use_cache=False, ): """Item embeddings and clustering""" self.device = "cuda:0" if torch.cuda.is_available() else "cpu" self.model_name = "sentence-transformers/all-mpnet-base-v2" self.tokenizer = transformers.AutoTokenizer.from_pretrained(self.model_name) self.model = transformers.AutoModel.from_pretrained(self.model_name).to( self.device ) self.text_dset = text_dset if dstats is None else dstats.text_dset self.text_field_name = ( text_field_name if dstats is None else dstats.our_text_field ) self.cache_path = cache_path if dstats is None else dstats.cache_path self.embeddings_dset_fid = pjoin(self.cache_path, "embeddings_dset") self.embeddings_dset = None self.node_list_fid = pjoin(self.cache_path, "node_list.th") self.node_list = None self.nid_map = None self.fig_tree_fid = pjoin(self.cache_path, "node_figure.json") self.fig_tree = None self.cached_clusters = {} self.use_cache = use_cache def compute_sentence_embeddings(self, sentences): """ Takes a list of sentences and computes their embeddings using self.tokenizer and self.model (with output dimension D) followed by mean pooling of the token representations and normalization Args: sentences ([string]): list of N input sentences Returns: torch.Tensor: sentence embeddings, dimension NxD """ batch = self.tokenizer( sentences, padding=True, truncation=True, return_tensors="pt" ) batch = {k: v.to(self.device) for k, v in batch.items()} with torch.no_grad(): model_output = self.model(**batch) sentence_embeds = sentence_mean_pooling( model_output, batch["attention_mask"] ) sentence_embeds /= sentence_embeds.norm(dim=-1, keepdim=True) return sentence_embeds def make_embeddings(self): """ Batch computes the embeddings of the Dataset self.text_dset, using the field self.text_field_name as input. Returns: Dataset: HF dataset object with a single EMBEDDING_FIELD field corresponding to the embeddings (list of floats) """ def batch_embed_sentences(sentences): return { EMBEDDING_FIELD: [ embed.tolist() for embed in self.compute_sentence_embeddings( sentences[self.text_field_name] ) ] } self.embeddings_dset = self.text_dset.map( batch_embed_sentences, batched=True, batch_size=32, remove_columns=[self.text_field_name], ) return self.embeddings_dset def make_text_embeddings(self): """Load embeddings dataset from cache or compute it.""" if self.use_cache and exists(self.embeddings_dset_fid): self.embeddings_dset = load_from_disk(self.embeddings_dset_fid) else: self.embeddings_dset = self.make_embeddings() self.embeddings_dset.save_to_disk(self.embeddings_dset_fid) def make_hierarchical_clustering( self, batch_size=1000, approx_neighbors=1000, min_cluster_size=10, ): if self.use_cache and exists(self.node_list_fid): self.node_list, self.nid_map = torch.load(self.node_list_fid) else: self.make_text_embeddings() embeddings = torch.Tensor(self.embeddings_dset[EMBEDDING_FIELD]) self.node_list = fast_cluster( embeddings, batch_size, approx_neighbors, min_cluster_size ) self.nid_map = dict( [(node["nid"], nid) for nid, node in enumerate(self.node_list)] ) torch.save((self.node_list, self.nid_map), self.node_list_fid) print(exists(self.fig_tree_fid), self.fig_tree_fid) if self.use_cache and exists(self.fig_tree_fid): self.fig_tree = read_json(self.fig_tree_fid) else: self.fig_tree = make_tree_plot( self.node_list, self.nid_map, self.text_dset, self.text_field_name ) self.fig_tree.write_json(self.fig_tree_fid) def find_cluster_beam(self, sentence, beam_size=20): """ This function finds the `beam_size` leaf clusters that are closest to the proposed sentence and returns the full path from the root to the cluster along with the dot product between the sentence embedding and the cluster centroid Args: sentence (string): input sentence for which to find clusters beam_size (int): this is a beam size algorithm to explore the tree Returns: [([int], float)]: list of (path_from_root, score) sorted by score """ embed = self.compute_sentence_embeddings([sentence])[0].to("cpu") active_paths = [([0], torch.dot(embed, self.node_list[0]["centroid"]).item())] finished_paths = [] children_ids_list = [ [ self.nid_map[nid] for nid in self.node_list[path[-1]]["children_ids"] if nid in self.nid_map ] for path, score in active_paths ] while len(active_paths) > 0: next_ids = sorted( [ ( beam_id, nid, torch.dot(embed, self.node_list[nid]["centroid"]).item(), ) for beam_id, children_ids in enumerate(children_ids_list) for nid in children_ids ], key=lambda x: x[2], reverse=True, )[:beam_size] paths = [ (active_paths[beam_id][0] + [next_id], score) for beam_id, next_id, score in next_ids ] active_paths = [] for path, score in paths: if ( len( [ nid for nid in self.node_list[path[-1]]["children_ids"] if nid in self.nid_map ] ) > 0 ): active_paths += [(path, score)] else: finished_paths += [(path, score)] children_ids_list = [ [ self.nid_map[nid] for nid in self.node_list[path[-1]]["children_ids"] if nid in self.nid_map ] for path, score in active_paths ] return sorted( finished_paths, key=lambda x: x[-1], reverse=True, )[:beam_size] def prepare_merges(embeddings, batch_size=1000, approx_neighbors=1000, low_thres=0.5): """ Prepares an initial list of merges for hierarchical clustering. First compute the `approx_neighbors` nearest neighbors, then propose a merge for any two points that are closer than `low_thres` Note that if a point has more than `approx_neighbors` neighbors closer than `low_thres`, this approach will miss some of those merges Args: embeddings (toch.Tensor): Tensor of sentence embeddings - dimension NxD batch_size (int): compute nearest neighbors of `batch_size` points at a time approx_neighbors (int): only keep `approx_neighbors` nearest neighbors of a point low_thres (float): only return merges where the dot product is greater than `low_thres` Returns: torch.LongTensor: proposed merges ([i, j] with i>j) - dimension: Mx2 torch.Tensor: merge scores - dimension M """ top_idx_pre = torch.cat( [torch.LongTensor(range(embeddings.shape[0]))[:, None]] * batch_size, dim=1 ) top_val_all = torch.Tensor(0, approx_neighbors) top_idx_all = torch.LongTensor(0, approx_neighbors) n_batches = math.ceil(len(embeddings) / batch_size) for b in tqdm(range(n_batches)): # TODO: batch across second dimension cos_scores = torch.mm( embeddings[b * batch_size : (b + 1) * batch_size], embeddings.t() ) for i in range(cos_scores.shape[0]): cos_scores[i, (b * batch_size) + i :] = -1 top_val_large, top_idx_large = cos_scores.topk( k=approx_neighbors, dim=-1, largest=True ) top_val_all = torch.cat([top_val_all, top_val_large], dim=0) top_idx_all = torch.cat([top_idx_all, top_idx_large], dim=0) max_neighbor_dist = top_val_large[:, -1].max().item() if max_neighbor_dist > low_thres: print( f"WARNING: with the current set of neireast neighbor, the farthest is {max_neighbor_dist}" ) all_merges = torch.cat( [ top_idx_pre[top_val_all > low_thres][:, None], top_idx_all[top_val_all > low_thres][:, None], ], dim=1, ) all_merge_scores = top_val_all[top_val_all > low_thres] return (all_merges, all_merge_scores) def merge_nodes(nodes, current_thres, previous_thres, all_merges, all_merge_scores): """ Merge all nodes if the max dot product between any of their descendants is greater than current_thres. Args: nodes ([dict]): list of dicts representing the current set of nodes current_thres (float): merge all nodes closer than current_thres previous_thres (float): nodes closer than previous_thres are already merged all_merges (torch.LongTensor): proposed merges ([i, j] with i>j) - dimension: Mx2 all_merge_scores (torch.Tensor): merge scores - dimension M Returns: [dict]: extended list with the newly created internal nodes """ merge_ids = (all_merge_scores <= previous_thres) * ( all_merge_scores > current_thres ) if merge_ids.sum().item() > 0: merges = all_merges[merge_ids] for a, b in merges.tolist(): node_a = nodes[a] while node_a["parent_id"] != -1: node_a = nodes[node_a["parent_id"]] node_b = nodes[b] while node_b["parent_id"] != -1: node_b = nodes[node_b["parent_id"]] if node_a["nid"] == node_b["nid"]: continue else: # merge if threshold allows if (node_a["depth"] + node_b["depth"]) > 0 and min( node_a["merge_threshold"], node_b["merge_threshold"] ) == current_thres: merge_to = None merge_from = None if node_a["nid"] < node_b["nid"]: merge_from = node_a merge_to = node_b if node_a["nid"] > node_b["nid"]: merge_from = node_b merge_to = node_a merge_to["depth"] = max(merge_to["depth"], merge_from["depth"]) merge_to["weight"] += merge_from["weight"] merge_to["children_ids"] += ( merge_from["children_ids"] if merge_from["depth"] > 0 else [merge_from["nid"]] ) for cid in merge_from["children_ids"]: nodes[cid]["parent_id"] = merge_to["nid"] merge_from["parent_id"] = merge_to["nid"] # else new node else: new_nid = len(nodes) new_node = { "nid": new_nid, "parent_id": -1, "depth": max(node_a["depth"], node_b["depth"]) + 1, "weight": node_a["weight"] + node_b["weight"], "children": [], "children_ids": [node_a["nid"], node_b["nid"]], "example_ids": [], "merge_threshold": current_thres, } node_a["parent_id"] = new_nid node_b["parent_id"] = new_nid nodes += [new_node] return nodes def finalize_node(node, nodes, min_cluster_size): """Post-process nodes to sort children by descending weight, get full list of leaves in the sub-tree, and direct links to the cildren nodes, then recurses to all children. Nodes with fewer than `min_cluster_size` descendants are collapsed into a single leaf. """ node["children"] = sorted( [ finalize_node(nodes[cid], nodes, min_cluster_size) for cid in node["children_ids"] ], key=lambda x: x["weight"], reverse=True, ) if node["depth"] > 0: node["example_ids"] = [ eid for child in node["children"] for eid in child["example_ids"] ] node["children"] = [ child for child in node["children"] if child["weight"] >= min_cluster_size ] assert node["weight"] == len(node["example_ids"]), print(node) return node def fast_cluster( embeddings, batch_size=1000, approx_neighbors=1000, min_cluster_size=10, low_thres=0.5, ): """ Computes an approximate hierarchical clustering based on example embeddings. The join criterion is min clustering, i.e. two clusters are joined if any pair of their descendants are closer than a threshold The approximate comes from the fact that only the `approx_neighbors` nearest neighbors of an example are considered for merges """ batch_size = min(embeddings.shape[0], batch_size) all_merges, all_merge_scores = prepare_merges( embeddings, batch_size, approx_neighbors, low_thres ) # prepare leaves nodes = [ { "nid": nid, "parent_id": -1, "depth": 0, "weight": 1, "children": [], "children_ids": [], "example_ids": [nid], "merge_threshold": 1.0, } for nid in range(embeddings.shape[0]) ] # one level per threshold range for i in range(10): p_thres = 1 - i * 0.05 c_thres = 0.95 - i * 0.05 nodes = merge_nodes(nodes, c_thres, p_thres, all_merges, all_merge_scores) # make root root_children = [ node for node in nodes if node["parent_id"] == -1 and node["weight"] >= min_cluster_size ] root = { "nid": len(nodes), "parent_id": -1, "depth": max([node["depth"] for node in root_children]) + 1, "weight": sum([node["weight"] for node in root_children]), "children": [], "children_ids": [node["nid"] for node in root_children], "example_ids": [], "merge_threshold": -1.0, } nodes += [root] for node in root_children: node["parent_id"] = root["nid"] # finalize tree tree = finalize_node(root, nodes, min_cluster_size) node_list = [] def rec_map_nodes(node, node_list): node_list += [node] for child in node["children"]: rec_map_nodes(child, node_list) rec_map_nodes(tree, node_list) # get centroids and distances for node in node_list: node_embeds = embeddings[node["example_ids"]] node["centroid"] = node_embeds.sum(dim=0) node["centroid"] /= node["centroid"].norm() node["centroid_dot_prods"] = torch.mv(node_embeds, node["centroid"]) node["sorted_examples_centroid"] = sorted( [ (eid, edp.item()) for eid, edp in zip(node["example_ids"], node["centroid_dot_prods"]) ], key=lambda x: x[1], reverse=True, ) return node_list def make_tree_plot(node_list, nid_map, text_dset, text_field_name): """ Makes a graphical representation of the tree encoded in node-list. The hover label for each node shows the number of descendants and the 5 examples that are closest to the centroid """ for nid, node in enumerate(node_list): # get list of node_examples = {} for sid, score in node["sorted_examples_centroid"]: node_examples[text_dset[sid][text_field_name]] = score if len(node_examples) >= 5: break node["label"] = node.get( "label", f"{nid:2d} - {node['weight']:5d} items
" + "
".join( [ f" {score:.2f} > {txt[:64]}" + ("..." if len(txt) >= 63 else "") for txt, score in node_examples.items() ] ), ) # make plot nodes labels = [node["label"] for node in node_list] root = node_list[0] root["X"] = 0 root["Y"] = 0 def rec_make_coordinates(node): total_weight = 0 add_weight = len(node["example_ids"]) - sum( [child["weight"] for child in node["children"]] ) for child in node["children"]: child["X"] = node["X"] + total_weight child["Y"] = node["Y"] - 1 total_weight += child["weight"] + add_weight / len(node["children"]) rec_make_coordinates(child) rec_make_coordinates(root) E = [] # list of edges Xn = [] Yn = [] Xe = [] Ye = [] for nid, node in enumerate(node_list): Xn += [node["X"]] Yn += [node["Y"]] for child in node["children"]: E += [(nid, nid_map[child["nid"]])] Xe += [node["X"], child["X"], None] Ye += [node["Y"], child["Y"], None] # make figure fig = go.Figure() fig.add_trace( go.Scatter( x=Xe, y=Ye, mode="lines", line=dict(color="rgb(210,210,210)", width=1), hoverinfo="none", ) ) fig.add_trace( go.Scatter( x=Xn, y=Yn, mode="markers", name="nodes", marker=dict( symbol="circle-dot", size=18, color="#6175c1", line=dict(color="rgb(50,50,50)", width=1) # '#DB4551', ), text=labels, hoverinfo="text", opacity=0.8, ) ) return fig