Ravish Chawla ravishchawla

Model Configuration	Number of Epochs	Training F1 Score	Validation F1 Score
2 LSTM Layers with 64 Hidden Units	5 epochs	0.637	0.540
2 LSTM Layers with 128 Hidden Units	5 epochs	0.625	0.567
2 LSTM Layers with 128 Hidden Units	15 epochs	0.661	0.558
1 LSTM Layer and a Self Attention Layer	5 epochs	0.771	0.601

Sources

	nearest_syns = NearestNeighbors(n_neighbors=total_syns+1).fit(embeddings_matrix);
	neighbours_mat = nearest_syns.kneighbors(embeddings_matrix[1:top_k])[1];
	synonyms = {x[0]: x[1:] for x in neighbours_mat};

	def augment_sentence(encoded_sentence, prob = 0.5):
	for posit in range(len(encoded_sentence)):
	if random.random() > prob:
	try:
	syns = synonyms[encoded_sentence[posit]];
	rand_syn = np.random.choice(syns);

	class Model(nn.Module):
	def __init__(self, embedding_matrix, hidden_unit = 64):
	super(Model, self).__init__();
	vocab_size = embeddings_tensor.shape[0];
	embedding_dim = embeddings_tensor.shape[1];

	self.embedding_layer = nn.Embedding(vocab_size, embedding_dim);
	self.embedding_layer.weight = nn.Parameter(embeddings_tensor);
	self.embedding_layer.weight.requires_grad = True;

	class Model(nn.Module):
	def __init__(self, embedding_matrix, hidden_unit = 64):
	super(Model, self).__init__();
	vocab_size = embeddings_tensor.shape[0];
	embedding_dim = embeddings_tensor.shape[1];

	self.embedding_layer = nn.Embedding(vocab_size, embedding_dim);
	self.embedding_layer.weight = nn.Parameter(embeddings_tensor);
	self.embedding_layer.weight.requires_grad = True;

	def train(nn_model, nn_optimizer, nn_criterion, data_loader, val_loader = None, num_epochs = 5, print_ratio = 0.1, verbose=True):
	for epoch in range(num_epochs):
	# Enable Training for the model
	nn_model.train()
	running_loss = 0;

	for ite, (x, y, l) in enumerate(data_loader):
	init_time = time.time();

	# Convert our tensors to GPU tensors