def load_glove_embedding(self, freeze=False):
if not cfg.multi_gpu:
initial_arr = self.m.embedding.weight.data.cpu().numpy()
emb = torch.from_numpy(
utils.get_glove_matrix(cfg.glove_path, self.reader.vocab,
initial_arr))
self.m.embedding.weight.data.copy_(emb)
else:
initial_arr = self.m.module.embedding.weight.data.cpu().numpy()
emb = torch.from_numpy(
utils.get_glove_matrix(cfg.glove_path, self.reader.vocab,
initial_arr))
self.m.module.embedding.weight.data.copy_(emb)
def __init__(self, env, results_path, tok, episode_len=20):
super(Seq2SeqAgent, self).__init__(env, results_path)
self.tok = tok
self.episode_len = episode_len
self.feature_size = self.env.feature_size
# Models
self.glove_dim = 300
with open('img_features/objects/object_vocab.txt', 'r') as f_ov:
self.obj_vocab = [k.strip() for k in f_ov.readlines()]
glove_matrix = get_glove_matrix(self.obj_vocab, self.glove_dim)
self.objencoder = ObjEncoder(glove_matrix.size(0), glove_matrix.size(1), glove_matrix).cuda()
enc_hidden_size = args.rnn_dim//2 if args.bidir else args.rnn_dim
self.encoder = model.EncoderLSTM(tok.vocab_size(), args.wemb, enc_hidden_size, padding_idx,
args.dropout, bidirectional=args.bidir).cuda()
self.decoder = model.AttnDecoderLSTM_Graph(args.aemb, args.rnn_dim, args.dropout, feature_size=self.feature_size + args.angle_feat_size).cuda()
self.critic = model.Critic().cuda()
self.models = (self.encoder, self.decoder, self.critic)
# Optimizers
self.encoder_optimizer = args.optimizer(self.encoder.parameters(), lr=args.lr)
self.decoder_optimizer = args.optimizer(self.decoder.parameters(), lr=args.lr)
self.critic_optimizer = args.optimizer(self.critic.parameters(), lr=args.lr)
self.optimizers = (self.encoder_optimizer, self.decoder_optimizer, self.critic_optimizer)
# Evaluations
self.losses = []
self.criterion = nn.CrossEntropyLoss(ignore_index=args.ignoreid, size_average=False)
# Logs
sys.stdout.flush()
self.logs = defaultdict(list)
def get_matrix_glove(self):
model = Glove.load(some_parameters['glove_model'])
vectors = model.word_vectors
dictionary = model.dictionary
corpus_matrix = dict()
for i in self.corpus.items():
doc_name = i[0]
original_sentences = i[1][0]
preprocesed_sentences = i[1][1]
ori_sents = extract_sentences(original_sentences, False)
pp_sents = extract_sentences(preprocesed_sentences, True)
if self.proccessing:
matrix = get_glove_matrix(pp_sents, dictionary, vectors)
else:
matrix = get_glove_matrix(ori_sents, dictionary, vectors)
corpus_matrix[doc_name] = matrix
return corpus_matrix
def __init__(self, env, listener, tok):
self.env = env
self.feature_size = self.env.feature_size
self.tok = tok
self.tok.finalize()
self.listener = listener
# Model
self.glove_dim = 300
with open(
'/students/u5399302/MatterportData/data/obj_stats2/objectvocab_filtered_dyno4.txt',
'r') as f_ov:
self.obj_vocab = [k.strip() for k in f_ov.readlines()]
glove_matrix = utils.get_glove_matrix(self.obj_vocab, self.glove_dim)
self.objencoder = ObjEncoder(glove_matrix.size(0),
glove_matrix.size(1),
glove_matrix).cuda()
print("VOCAB_SIZE", self.tok.vocab_size())
self.encoder = model.SpeakerEncoder(self.feature_size +
args.angle_feat_size,
args.rnn_dim,
args.dropout,
bidirectional=args.bidir).cuda()
self.decoder = model.SpeakerDecoder(self.tok.vocab_size(), args.wemb,
self.tok.word_to_index['<PAD>'],
args.rnn_dim, args.dropout).cuda()
self.encoder_optimizer = args.optimizer(self.encoder.parameters(),
lr=args.lr)
self.decoder_optimizer = args.optimizer(self.decoder.parameters(),
lr=args.lr)
# Evaluation
self.softmax_loss = torch.nn.CrossEntropyLoss(
ignore_index=self.tok.word_to_index['<PAD>'])
# Will be used in beam search
self.nonreduced_softmax_loss = torch.nn.CrossEntropyLoss(
ignore_index=self.tok.word_to_index['<PAD>'],
size_average=False,
reduce=False)