def __init__(self, config, gpu_list, *args, **params):
super(AttentionRNN, self).__init__()
self.input_dim = 768
self.hidden_dim = config.getint('model', 'hidden_dim')
self.dropout_rnn = config.getfloat('model', 'dropout_rnn')
self.dropout_fc = config.getfloat('model', 'dropout_fc')
self.bidirectional = config.getboolean('model', 'bidirectional')
if self.bidirectional:
self.direction = 2
else:
self.direction = 1
self.num_layers = config.getint("model", 'num_layers')
self.output_dim = config.getint("model", "output_dim")
self.max_para_q = config.getint('model', 'max_para_q')
if config.get('model', 'rnn') == 'lstm':
self.rnn = nn.LSTM(self.input_dim, self.hidden_dim, batch_first=True, num_layers=self.num_layers,
bidirectional=self.bidirectional, dropout=self.dropout_rnn)
else:
self.rnn = nn.GRU(self.input_dim, self.hidden_dim, batch_first=True, num_layers=self.num_layers,
bidirectional=self.bidirectional, dropout=self.dropout_rnn)
self.max_pool = nn.MaxPool1d(kernel_size=self.max_para_q)
self.fc_a = nn.Linear(self.hidden_dim*self.direction, self.hidden_dim*self.direction)
self.attention = Attention(config)
self.fc_f = nn.Linear(self.hidden_dim*self.direction, self.output_dim)
self.dropout = nn.Dropout(self.dropout_fc)
self.weight = self.init_weight(config, gpu_list)
#self.criterion = nn.CrossEntropyLoss(weight=self.weight)
self.criterion = nn.MSELoss()
self.accuracy_function = init_accuracy_function(config, *args, **params)
def __init__(self, config, gpu_list, *args, **params):
super(BaselineYSBert, self).__init__()
self.output_dim = config.getint("model", "output_dim")
self.bert = BertModel.from_pretrained(config.get("model", "bert_path"))
self.fc = nn.Linear(768, self.output_dim * 2)
self.criterion = MultiLabelSoftmaxLoss(config)
self.accuracy_function = init_accuracy_function(config, *args, **params)
def __init__(self, config, gpu_list, *args, **params):
super(BaselineZMBert, self).__init__()
self.output_dim = config.getint("model", "output_dim")
self.bert = BertModel.from_pretrained(config.get("model", "bert_path"))
self.fc = nn.Linear(768, self.output_dim)
self.criterion = cross_entropy_loss
self.accuracy_function = init_accuracy_function(
config, *args, **params)
def __init__(self, config, gpu_list, *args, **params):
super(BertPoint, self).__init__()
self.output_dim = config.getint("model", "output_dim")
self.output_mode = config.get('model', 'output_mode')
self.bert = BertModel.from_pretrained(config.get("model", "bert_path"))
self.fc = nn.Linear(768, self.output_dim)
self.criterion = nn.CrossEntropyLoss()
self.accuracy_function = init_accuracy_function(
config, *args, **params)
def __init__(self, config, gpu_list, *args, **params):
super(BasicBert, self).__init__()
self.output_dim = config.getint("model", "output_dim")
self.bert = BertEncoder(config, gpu_list, *args, **params)
self.fc = nn.Linear(768, self.output_dim)
self.seq = nn.Sequential(self.bert, self.fc)
self.criterion = nn.CrossEntropyLoss()
self.accuracy_function = init_accuracy_function(
config, *args, **params)
def __init__(self, config, gpu_list, *args, **params):
super(BaselineYSDPCNN, self).__init__()
self.model_name = "DPCNN"
self.emb_dim = 300
self.mem_dim = 150
self.output_dim = config.getint("model", "output_dim") * 2
self.word_num = 0
f = open(os.path.join(config.get("model", "bert_path"), "vocab.txt"), "r")
for line in f:
self.word_num += 1
self.embedding = nn.Embedding(self.word_num, self.emb_dim)
# region embedding
self.region_embedding = nn.Sequential(
nn.Conv1d(self.emb_dim, self.mem_dim,
kernel_size=3, padding=1),
nn.BatchNorm1d(num_features=self.mem_dim),
nn.ReLU(),
)
self.conv_block = nn.Sequential(
nn.BatchNorm1d(num_features=self.mem_dim),
nn.ReLU(),
nn.Conv1d(self.mem_dim, self.mem_dim,
kernel_size=3, padding=1),
nn.BatchNorm1d(num_features=self.mem_dim),
nn.ReLU(),
nn.Conv1d(self.mem_dim, self.mem_dim,
kernel_size=3, padding=1),
)
self.num_seq = config.getint("data", "max_seq_length")
resnet_block_list = []
while (self.num_seq > 2):
resnet_block_list.append(ResnetBlock(self.mem_dim))
self.num_seq = self.num_seq // 2
self.resnet_layer = nn.Sequential(*resnet_block_list)
self.fc = nn.Sequential(
nn.Linear(self.mem_dim * self.num_seq, self.output_dim),
nn.BatchNorm1d(self.output_dim),
nn.ReLU(inplace=True),
nn.Linear(self.output_dim, self.output_dim)
)
self.criterion = MultiLabelSoftmaxLoss(config)
self.accuracy_function = init_accuracy_function(config, *args, **params)
def __init__(self, config, gpu_list, *args, **params):
super(BaselineYSGRU, self).__init__()
self.emb_dim = 256
self.hidden_size = 256
self.output_dim = config.getint("model", "output_dim") * 2
self.word_num = 0
f = open(os.path.join(config.get("model", "bert_path"), "vocab.txt"), "r")
for line in f:
self.word_num += 1
self.embedding = nn.Embedding(self.word_num, self.emb_dim)
self.gru = nn.GRU(self.emb_dim, self.hidden_size, 4, batch_first=True, bidirectional=True)
self.fc = nn.Linear(2 * self.hidden_size, self.output_dim)
self.criterion = MultiLabelSoftmaxLoss(config)
self.accuracy_function = init_accuracy_function(config, *args, **params)
def __init__(self, config, gpu_list, *args, **params):
super(ZMDqn, self).__init__()
try:
self.save_details = config.get("eval",
"print_details").find("yes") != -1
except:
self.save_details = False
self.accuracy_function = init_accuracy_function(
config, *args, **params)
self.learning_rate = config.getfloat("train", "learning_rate")
self.batch_size = config.getint("rl", "batch_size")
self.epsilone = config.getfloat("rl", "epsilone")
self.gamma = config.getfloat("rl", "gamma")
self.target_update = config.getint("rl", "target_update")
self.memory_capacity = config.getint("rl", "memory_capacity")
self.bert = BertModel.from_pretrained(config.get("model", "bert_path"))
self.n_actions = config.getint("rl", "n_actions")
self.n_question = config.getint("rl", "n_questions")
self.output_dim = config.getint("model", "output_dim")
self.n_states = self.n_actions
self.user = UserSimulator()
self.agent_initialized = False
self.memory = ReplayMemory(self.memory_capacity)
self.policy_net = nn.Linear(self.n_states + 768, self.n_actions)
self.target_net = nn.Linear(self.n_states + 768, self.n_actions)
self.i_episode = 0
self.optimizer = torch.optim.Adam(self.policy_net.parameters(),
lr=self.learning_rate)
self.loss_function = nn.MSELoss()
self.criterion_out = MultiLabelSoftmaxLoss(config)
lgb_path = config.get("ml", "lgb_path")
self.lgb = pickle.load(
open(os.path.join(lgb_path, "predict_net.pkl"), "rb"))