def run_on_batch(all_pids, all_lbls, all_records, all_canopies, model,
encoding_model, canopy2predictions, canopy2tree, trees):
features = encoding_model.encode(all_records)
if len(all_pids) > 1:
grinch = Agglom(model, features, num_points=len(all_pids))
grinch.build_dendrogram_hac()
fc = grinch.flat_clustering(model.aux['threshold'])
tree_id = len(trees)
trees.append(grinch)
for i in range(len(all_pids)):
if all_canopies[i] not in canopy2predictions:
canopy2predictions[all_canopies[i]] = [[], []]
canopy2tree[all_canopies[i]] = tree_id
canopy2predictions[all_canopies[i]][0].append(all_pids[i])
canopy2predictions[all_canopies[i]][1].append(
'%s-%s' % (all_canopies[i], fc[i]))
return canopy2predictions
else:
fc = [0]
for i in range(len(all_pids)):
if all_canopies[i] not in canopy2predictions:
canopy2predictions[all_canopies[i]] = [[], []]
canopy2tree[all_canopies[i]] = None
canopy2predictions[all_canopies[i]][0].append(all_pids[i])
canopy2predictions[all_canopies[i]][1].append(
'%s-%s' % (all_canopies[i], fc[i]))
return canopy2predictions
def run_on_batch(all_pids, all_lbls, all_records, all_canopies, model,
encoding_model, canopy2predictions, canopy2tree, trees,
pids_list, canopy_list):
features = encoding_model.encode(all_records)
grinch = Agglom(model,
features,
num_points=len(all_pids),
min_allowable_sim=0)
grinch.build_dendrogram_hac()
fc = grinch.flat_clustering(model.aux['threshold'])
logging.info(
'run_on_batch - threshold %s, linkages: min %s, max %s, avg %s, std %s',
model.aux['threshold'], np.min(grinch.all_thresholds()),
np.max(grinch.all_thresholds()), np.mean(grinch.all_thresholds()),
np.std(grinch.all_thresholds()))
tree_id = len(trees)
trees.append(grinch)
pids_list.append(all_pids)
canopy_list.append(all_canopies)
for i in range(len(all_pids)):
if all_canopies[i] not in canopy2predictions:
canopy2predictions[all_canopies[i]] = [[], []]
canopy2tree[all_canopies[i]] = tree_id
canopy2predictions[all_canopies[i]][0].append(all_pids[i])
canopy2predictions[all_canopies[i]][1].append('%s-%s' %
(all_canopies[i], fc[i]))
return canopy2predictions
def run_on_batch(all_pids, all_lbls, all_records, all_canopies, model,
encoding_model, canopy2predictions, canopy2tree, trees,
pids_list):
"""
:param all_pids: Ids of the records we are running on
:param all_lbls: Labels (or blank labels if they are not available)
:param all_records: The records (i.e. features - python strings of the inventor name, assignee etc)
:param all_canopies: The canopies of each record (right now this is always all the same canopy)
:param model: linear scoring / similarity model which takes in two feature vectors and gives a similarity
:param encoding_model: mapping from all_records gives you something that is input to model
:param canopy2predictions: where results are stored
:param canopy2tree: where results are stored
:param trees: where results are stored
:param pids_list: where results are stored
:return:
"""
# extracting features
features = encoding_model.encode(all_records)
if len(all_pids) > 1:
# running clustering
grinch = Agglom(model, features, num_points=len(all_pids))
grinch.build_dendrogram_hac()
fc = grinch.flat_clustering(model.aux['threshold'])
# storing the state of the clustering for the intremental setting
tree_id = len(trees)
# store the tree that is build for this canopy
trees.append(grinch)
pids_list.append(all_pids)
for i in range(len(all_pids)):
# record mapping from canopy to the tree id
canopy2tree[all_canopies[i]] = tree_id
if all_canopies[i] not in canopy2predictions:
canopy2predictions[all_canopies[i]] = [[], []]
canopy2tree[all_canopies[i]] = tree_id
# save predictions (used in the non incremental setting)
canopy2predictions[all_canopies[i]][0].append(all_pids[i])
canopy2predictions[all_canopies[i]][1].append(
'%s-%s' % (all_canopies[i], fc[i]))
return canopy2predictions
else:
raise Exception('Must have non-singleton canopies')
fc = [0]
for i in range(len(all_pids)):
if all_canopies[i] not in canopy2predictions:
canopy2predictions[all_canopies[i]] = [[], []]
canopy2tree[all_canopies[i]] = None
canopy2predictions[all_canopies[i]][0].append(all_pids[i])
canopy2predictions[all_canopies[i]][1].append(
'%s-%s' % (all_canopies[i], fc[i]))
return canopy2predictions
def run_on_batch(all_pids, all_lbls, all_records, all_canopies, model,
encoding_model, canopy2predictions):
features = encoding_model.encode(all_records)
# grinch = WeightedMultiFeatureGrinch(model, features, num_points=len(all_pids), max_nodes=3 * len(all_pids))
grinch = Agglom(model, features, num_points=len(all_pids))
grinch.build_dendrogram_hac()
# grinch.get_score_batch(grinch.all_valid_internal_nodes())
fc = grinch.flat_clustering(model.aux['threshold'])
for i in range(len(all_pids)):
if all_canopies[i] not in canopy2predictions:
canopy2predictions[all_canopies[i]] = [[], []]
canopy2predictions[all_canopies[i]][0].append(all_pids[i])
canopy2predictions[all_canopies[i]][1].append('%s-%s' %
(all_canopies[i], fc[i]))
return canopy2predictions
def dev_eval(self, datasets):
logging.info('dev eval using %s datasets', len(datasets))
trees = []
gold_clustering = []
dataset_names = []
for idx, (dataset_name, dataset) in enumerate(datasets.items()):
pids, lbls, records, features = dataset[0], dataset[1], dataset[2], dataset[3]
logging.info('Running on dev dataset %s of %s | %s with %s points', idx, len(datasets), dataset_name,
len(pids))
if len(pids) > 0:
grinch = Agglom(self.grinch.model, features, num_points=len(pids))
grinch.build_dendrogram_hac()
trees.append(grinch)
gold_clustering.extend(lbls)
dataset_names.append(dataset_name)
eval_ids = [i for i in range(len(gold_clustering)) if gold_clustering[i] != '-1']
thresholds = np.sort(np.squeeze(self.get_thresholds(trees, self.num_thresholds)))
scores_per_threshold = []
os.makedirs(os.path.join(self.outdir, 'dev'), exist_ok=True)
dev_out_f = open(os.path.join(self.outdir, 'dev', 'dev_%s.tsv' % self.global_step), 'w')
for thres in thresholds:
pred_clustering = []
for idx, t in enumerate(trees):
fc = t.flat_clustering(thres)
pred_clustering.extend(['%s-%s' % (dataset_names[idx], x) for x in fc])
metrics = eval_micro_pw_f1([pred_clustering[x] for x in eval_ids], [gold_clustering[x] for x in eval_ids])
scores_per_threshold.append(metrics)
logging.info('[dev] threshold %s | %s', thres, "|".join(['%s=%s' % (k, v) for k, v in metrics.items()]))
arg_best_f1 = max([x for x in range(len(scores_per_threshold))],
key=lambda x: scores_per_threshold[x]['micro_pw_f1'])
for idx, t in enumerate(trees):
dataset = datasets[dataset_names[idx]]
pids, lbls, records, features = dataset[0], dataset[1], dataset[2], dataset[3]
fc = t.flat_clustering(thresholds[arg_best_f1])
for j in range(len(records)):
dev_out_f.write("%s\n" % records[j].pretty_tsv('%s-%s' % (dataset_names[idx], fc[j]), lbls[j]))
metrics = scores_per_threshold[arg_best_f1]
logging.info('[dev] best threshold %s | %s', thresholds[arg_best_f1],
"|".join(['%s=%s' % (k, v) for k, v in metrics.items()]))
dev_out_f.close()
if metrics['micro_pw_f1'] > self.best_f1:
logging.info('new best f1 %s > %s', metrics['micro_pw_f1'], self.best_f1)
self.model.aux['threshold'] = thresholds[arg_best_f1]
self.best_f1 = metrics['micro_pw_f1']
self.save('best')
return thresholds[arg_best_f1], metrics
def run_on_batch(all_pids, all_lbls, all_records, all_canopies, model,
encoding_model, canopy2predictions):
features = encoding_model.encode(all_records)
grinch = Agglom(model,
features,
num_points=len(all_pids),
min_allowable_sim=0)
grinch.build_dendrogram_hac()
fc = grinch.flat_clustering(model.aux['threshold'])
# import pdb
# pdb.set_trace()
for i in range(len(all_pids)):
if all_canopies[i] not in canopy2predictions:
canopy2predictions[all_canopies[i]] = [[], []]
canopy2predictions[all_canopies[i]][0].append(all_pids[i])
canopy2predictions[all_canopies[i]][1].append('%s-%s' %
(all_canopies[i], fc[i]))
return canopy2predictions
def main(argv):
logging.info('Running clustering with argv %s', str(argv))
wandb.init(project="%s-%s" % (FLAGS.exp_name, FLAGS.dataset_name))
wandb.config.update(flags.FLAGS)
outdir = os.path.join(FLAGS.outprefix, wandb.env.get_project(),
os.environ.get(wandb.env.SWEEP_ID, 'solo'),
wandb.env.get_run())
logging.info('outdir %s', outdir)
os.makedirs(outdir, exist_ok=True)
os.makedirs(FLAGS.cache_dir, exist_ok=True)
need_to_load_train = False
dev_cache_file = os.path.join(
FLAGS.cache_dir, 'dev-%s.pkl' % os.path.basename(FLAGS.dev_data))
need_to_load_dev = False
if os.path.exists(dev_cache_file):
logging.info('Using cached dev data! %s', dev_cache_file)
with open(dev_cache_file, 'rb') as fin:
dev_collection = pickle.load(fin)
dev_datasets = dev_collection[0]
dev_set_tmp = dict()
for canopy in dev_datasets.keys():
pids, lbls, records = dev_datasets[canopy]
if len(records) >= FLAGS.min_dev_size and len(
records) <= FLAGS.max_dev_size:
dev_set_tmp[canopy] = [pids, lbls, records]
dev_datasets = dev_set_tmp
else:
need_to_load_dev = True
if need_to_load_train or need_to_load_dev:
logging.info('need_to_load_train %s', need_to_load_train)
logging.info('need_to_load_dev %s', need_to_load_dev)
logging.info('Loading canopy dictionary...')
with open(FLAGS.canopy2record_dict, 'rb') as fin:
canopy2record = pickle.load(fin)
with open(FLAGS.record2canopy_dict, 'rb') as fin:
record2canopy = pickle.load(fin)
logging.info('Loading rawinventor mentions....')
rawinventor_cache_file = os.path.join(
FLAGS.cache_dir, '%s.pkl' % os.path.basename(FLAGS.rawinventor))
logging.info('Loading rawinventor mentions....')
if os.path.exists(rawinventor_cache_file):
logging.info('Using cache of rawinventor mentions %s',
rawinventor_cache_file)
with open('%s.pkl' % FLAGS.rawinventor, 'rb') as fin:
record_dict = pickle.load(fin)
else:
logging.info('No cache of rawinventor mentions %s',
'%s' % FLAGS.rawinventor)
record_dict = dict()
for m in load_inventor_mentions(FLAGS.rawinventor):
record_dict[m.mention_id] = m
with open('%s.pkl' % FLAGS.rawinventor, 'wb') as fout:
pickle.dump(record_dict, fout)
if need_to_load_dev:
logging.info('Loading dev data ids and labels')
canopies = set()
point2label = dict()
with open(FLAGS.dev_data, 'r') as fin:
for line in fin:
splt = line.strip().split('\t')
# check to see we have the point id in the record dictionary
if splt[0] not in record_dict:
logging.info("[dev] missing record %s", splt[0])
continue
point2label[splt[0]] = splt[1]
pt_canopy = record2canopy[splt[0]]
canopies.add(pt_canopy)
dev_set_tmp = dict()
for canopy in canopies:
pids = canopy2record[canopy]
lbls = [
point2label[x] if x in point2label else '-1' for x in pids
]
records = [record_dict[x] for x in pids]
if len(records) >= FLAGS.min_dev_size and len(
records) <= FLAGS.max_dev_size:
dev_set_tmp[canopy] = [pids, lbls, records]
dev_datasets = dev_set_tmp
dev_collection = [dev_datasets]
with open(dev_cache_file, 'wb') as fin:
pickle.dump(dev_collection, fin)
logging.info('Number of dev canopies: %s', len(dev_datasets))
encoding_model = InventorModel.from_flags(FLAGS)
model = torch.load(FLAGS.model_path)
logging.info('setting up features for dev datasets...')
dev_sets = dict()
for canopy, dataset in tqdm(dev_datasets.items(), 'dev data build'):
pids, lbls, records = dataset[0], dataset[1], dataset[2]
# logging.info('canopy %s - len(dataset) %s', canopy, len(records))
dev_sets[canopy] = [
pids, lbls, records,
encoding_model.encode(records)
]
logging.info('dev eval using %s datasets', len(dev_sets))
trees = []
gold_clustering = []
dataset_names = []
for idx, (dataset_name, dataset) in enumerate(dev_sets.items()):
pids, lbls, records, features = dataset[0], dataset[1], dataset[
2], dataset[3]
logging.info('Running on dev dataset %s of %s | %s with %s points',
idx, len(dev_sets), dataset_name, len(pids))
if len(pids) > 0:
grinch = Agglom(model, features, num_points=len(pids))
grinch.build_dendrogram_hac()
trees.append(grinch)
gold_clustering.extend(lbls)
dataset_names.append(dataset_name)
eval_ids = [
i for i in range(len(gold_clustering)) if gold_clustering[i] != '-1'
]
thresholds = [model.aux['threshold']]
scores_per_threshold = []
os.makedirs(os.path.join(outdir, 'dev'), exist_ok=True)
dev_out_f = open(os.path.join(outdir, 'dev', 'dev.tsv'), 'w')
pred_clustering = []
for idx, t in enumerate(trees):
fc = t.flat_clustering(model.aux['threshold'])
pred_clustering.extend(['%s-%s' % (dataset_names[idx], x) for x in fc])
metrics = eval_micro_pw_f1([pred_clustering[x] for x in eval_ids],
[gold_clustering[x] for x in eval_ids])
scores_per_threshold.append(metrics)
logging.info('[dev] threshold %s | %s', model.aux['threshold'],
"|".join(['%s=%s' % (k, v) for k, v in metrics.items()]))
for idx, t in enumerate(trees):
dataset = dev_sets[dataset_names[idx]]
pids, lbls, records, features = dataset[0], dataset[1], dataset[
2], dataset[3]
fc = t.flat_clustering(model.aux['threshold'])
for j in range(len(records)):
dev_out_f.write(
"%s\n" %
records[j].pretty_tsv('%s-%s' %
(dataset_names[idx], fc[j]), lbls[j]))
dev_out_f.close()