def main():
""" Load data, train and evaluate a model """
args = get_args()
if args.classifier_type in ['cnn', 'ffn']: # PyTorch
run_pkg = 'pytorch'
else:
run_pkg = 'sklearn'
if not args.text_emb_type:
emb_type_name = args.post_emb_type
else:
emb_type_name = args.text_emb_type
name_parts = [
args.model_name,
args.features.replace(',', '+'), emb_type_name, args.classifier_type
]
name_parts = [n for n in name_parts if n is not None]
exp_name = '_'.join(name_parts).strip('_')
exp_output_dirpath = os.path.join(args.output_dirpath, exp_name)
model_outpath = f'/projects/tumblr_community_identity/models/{exp_name}.pkl'
# Load trained embedding models
# TODO: Move the loading of things elsewhere (load_embeddings)
if (args.text_emb_type == 'unigrams' or args.text_emb_type is None) and \
(args.post_emb_type == 'unigrams' or args.post_emb_type == 'tags' or \
args.post_emb_type is None):
word_embs = None
graph_embs = None
sent_embs = None
else:
print("Loading embeddings...")
emb_loader = EmbeddingLoader(args.post_emb_type, args.text_emb_type)
load_word_embs, load_graph_embs, graph_embs = False, False, None
load_sent_embs, sent_embs = False, None
if args.post_emb_type != 'unigrams':
load_word_embs = True
if 'graph' in args.features:
load_graph_embs = True
if 'text' in args.features and args.text_emb_type in [
'fasttext', 'bert'
]:
load_sent_embs = True
emb_loader.load(word_embs=load_word_embs,
graph_embs=load_graph_embs,
sent_embs=load_sent_embs)
if load_graph_embs:
graph_embs = emb_loader.graph_embs
if load_sent_embs:
sent_embs = emb_loader.sent_embs
word_embs = emb_loader.word_embs
# Load and filter dataset
print("Loading and filtering data...")
dataset = Dataset()
dataset.load(args.data_location, args.task)
if args.load_preprocessed:
id2token = load_pickle(args.load_preprocessed)
user_filter = set(list(id2token.keys()))
dataset.filter(user_ids=user_filter,
word_filter=word_embs.wv,
word_filter_min=args.word_filter_min,
preprocessed_descs=id2token)
else:
if args.text_emb_type and args.post_emb_type and \
args.text_emb_type != 'unigrams' and args.post_emb_type != 'unigrams':
dataset.filter(word_filter=word_embs.wv,
word_filter_min=args.word_filter_min)
elif 'comms' in args.features:
dataset.load_filter_communities()
# Extract features
print("Extracting features...")
post_ngrams, post_tags, text_ngrams = False, False, False
if run_pkg == 'pytorch':
extractor = FeatureExtractor(args.features,
word_embs=word_embs,
graph_embs=graph_embs,
sent_embs=sent_embs,
word_inds=True,
padding_size=30)
dataset = extractor.extract(dataset, run_pkg, dev=True)
else:
if args.post_emb_type == 'unigrams':
post_ngrams = True
elif args.post_emb_type == 'tags':
post_tags = True
if args.text_emb_type == 'unigrams':
text_ngrams = True
extractor = FeatureExtractor(args.features,
word_embs=word_embs,
graph_embs=graph_embs,
sent_embs=sent_embs,
post_ngrams=post_ngrams,
post_tags=post_tags,
text_ngrams=text_ngrams,
select_k=args.feature_selection_k,
post_tag_pca=args.post_tag_pca,
post_tag_lda=args.post_tag_lda)
dataset = extractor.extract(dataset, run_pkg, dev=True)
# Run model
print("Running model...")
data_outpath = f'../tmp/{exp_name}_data.pkl'
dataset.save(data_outpath)
print(f"\tSaved dataset folds to {data_outpath}")
experiment = Experiment(extractor,
dataset,
args.classifier_type,
args.use_cuda,
args.epochs,
sfs_k=args.forward_feature_selection_k)
experiment.run()
# Print output
if experiment.dev_score:
print(f'\tDev set score: {experiment.dev_score: .4f}')
print(f'\tTest set score: {experiment.test_score: .4f}')
# Save settings, output
if run_pkg == 'sklearn':
dataset.save_settings(exp_output_dirpath)
experiment.save_output(exp_output_dirpath)
experiment.save_model(model_outpath)
