def init_weights(self):
init.orthogonal_(self.rnn.weight_ih_l0)
init.uniform_(self.rnn.weight_hh_l0, a=-0.01, b=0.01)
embeddings = preprocessing.load_embeddings(self.pretrained_vectors)
embedding_weights = torch.FloatTensor(self.vocab_size, self.input_size)
init.uniform_(embedding_weights, a=-0.25, b=0.25)
for k, v in embeddings.items():
embedding_weights[k] = torch.FloatTensor(v)
embedding_weights[0] = torch.FloatTensor([0] * self.input_size)
del self.embedding.weight
self.embedding.weight = nn.Parameter(embedding_weights)
print('Loading networks', end='...')
decoder = DecoderWithAttention(ATTENTION_DIM, EMBBEDING_DIM, DECODER_DIM,
vocab_size, ENCODER_DIM, DROPOUT)
encoder = Encoder(output_size=12)
print('done')
if START_EPOCH != 0:
print('Loading last model', end='...')
decoder.load_state_dict(
torch.load('../models/image_captioning_{}.model'.format(START_EPOCH)))
print('done')
# Embedding
if EMBBEDING_DIM == 200:
print('Loading embeddings', end='...')
embedding, _ = load_embeddings(embedding_file, DATA_FOLDER)
decoder.load_pretrained_embeddings(embedding, fine_tune=True)
print('done')
# Loss function
criterion = nn.CrossEntropyLoss().to(DEVICE)
# Data loader
train_loader = torch.utils.data.DataLoader(CaptionDataset(
DATA_FOLDER, 'TRAIN'),
batch_size=BATCH_SIZE,
shuffle=True,
num_workers=1,
pin_memory=True)
valid_loader = torch.utils.data.DataLoader(CaptionDataset(DATA_FOLDER, 'VAL'),
# Batch parameters
batch_size = 64
context_length = 2
user_filter = "Exiatus"
# Model parameters
decoder_learning_ratio = 5.0
epochs = 5000
grad_clip = 50
hidden_size = 200
learning_rate = 0.0001
# Load vocab/embeddings
w2i, i2w = preprocessing.load_vocab(vocab_file)
embeddings = preprocessing.load_embeddings(embedding_file, w2i)
vocab_size = len(w2i)
input_size = len(embeddings[1])
print("Loaded %d words" % vocab_size)
# Load/numberize messages
messages = preprocessing.load_messages(messages_file)
numberized_messages = preprocessing.numberize_messages(messages, w2i)
print("Loaded %d messages" % len(messages))
# Create encoder
encoder = models.Encoder(
input_size=input_size,
hidden_size=hidden_size,