def __init__(self, config, gpu_list, *args, **params):
super(ModelXS, self).__init__()
self.batch_size = config.getint("model", "hidden_size")
self.hidden_size = config.getint("model", "hidden_size")
self.word_num = 0
f = open(config.get("data", "word2id"), "r", encoding="utf8")
for line in f:
self.word_num += 1
self.context_len = config.getint("data", "max_option_len") * 4
self.question_len = config.getint("data", "max_question_len")
self.embedding = nn.Embedding(self.word_num, self.hidden_size)
self.context_encoder = GRUEncoder(config, gpu_list, *args, **params)
self.question_encoder = GRUEncoder(config, gpu_list, *args, **params)
self.attention = Attention(config, gpu_list, *args, **params)
self.dropout = nn.Dropout(config.getfloat("model", "dropout"))
self.dim_capsule = config.getint("model", "dim_capsule")
self.num_compressed_capsule = config.getint("model",
"num_compressed_capsule")
self.ngram_size = [2, 4, 8]
self.convs_doc = nn.ModuleList([
nn.Conv1d(self.hidden_size * 2,
config.getint("model", "capsule_size"),
K,
stride=2) for K in self.ngram_size
])
torch.nn.init.xavier_uniform_(self.convs_doc[0].weight)
torch.nn.init.xavier_uniform_(self.convs_doc[1].weight)
torch.nn.init.xavier_uniform_(self.convs_doc[2].weight)
self.primary_capsules_doc = PrimaryCaps(num_capsules=self.dim_capsule,
in_channels=config.getint(
"model", "capsule_size"),
out_channels=16,
kernel_size=1,
stride=1)
self.flatten_capsules = FlattenCaps()
self.W_doc = nn.Parameter(
torch.FloatTensor(12224, self.num_compressed_capsule))
torch.nn.init.xavier_uniform_(self.W_doc)
self.fc_size = config.getint("model", "fc_size")
self.fc_capsules_doc_child = FCCaps(
config,
output_capsule_num=self.fc_size,
input_capsule_num=self.num_compressed_capsule,
in_channels=self.dim_capsule,
out_channels=self.dim_capsule)
self.bce = nn.MultiLabelSoftMarginLoss(reduction='sum')
self.gelu = nn.GELU()
# self.fc_module_q = nn.Linear(self.question_len, 1)
self.fc_module = nn.Linear(self.fc_size, 4)
self.accuracy_function = multi_label_top1_accuracy
def __init__(self, config, gpu_list, *args, **params):
super(BertCAPSModel, self).__init__()
self.hidden_size = config.getint("model", "hidden_size")
self.word_num = 0
f = open(config.get("data", "word2id"), "r", encoding="utf8")
for line in f:
self.word_num += 1
self.embedding = nn.Embedding(self.word_num, self.hidden_size)
self.context_encoder = BertEncoder(config, gpu_list, *args, **params)
self.question_encoder = LSTMEncoder(config, gpu_list, *args, **params)
self.attention = Attention(config, gpu_list, *args, **params)
self.num_classes = 4
# self.conv_channel = config.getint("data", "max_question_len") + config.getint("data", "max_option_len")
self.question_len = config.getint("data", "max_question_len")
self.context_len = config.getint("data", "max_option_len") * 4
self.dim_capsule = config.getint("model", "dim_capsule")
self.num_compressed_capsule = config.getint("model",
"num_compressed_capsule")
self.ngram_size = [2, 4, 8]
self.convs_doc = nn.ModuleList([
nn.Conv1d(self.hidden_size, 32, K, stride=2)
for K in self.ngram_size
])
torch.nn.init.xavier_uniform_(self.convs_doc[0].weight)
torch.nn.init.xavier_uniform_(self.convs_doc[1].weight)
torch.nn.init.xavier_uniform_(self.convs_doc[2].weight)
self.primary_capsules_doc = PrimaryCaps(num_capsules=self.dim_capsule,
in_channels=32,
out_channels=32,
kernel_size=1,
stride=1)
self.flatten_capsules = FlattenCaps()
self.W_doc = nn.Parameter(
torch.FloatTensor(49024, self.num_compressed_capsule))
torch.nn.init.xavier_uniform_(self.W_doc)
self.fc_capsules_doc_child = FCCaps(
config,
output_capsule_num=self.num_classes,
input_capsule_num=self.num_compressed_capsule,
in_channels=self.dim_capsule,
out_channels=self.dim_capsule)
# self.rank_module = nn.Linear(hidden_size, 1)
self.criterion = nn.CrossEntropyLoss()
self.bce = nn.BCEWithLogitsLoss(reduction='mean')
self.fc_module = nn.Linear(self.dim_capsule, self.num_classes)
self.accuracy_function = multi_label_top1_accuracy
def __init__(self, config, gpu_list, *args, **params):
super(ModelLC, self).__init__()
self.hidden_size = config.getint("model", "hidden_size")
self.context_len = config.getint("data", "max_option_len")
self.question_len = config.getint("data", "max_question_len")
self.context_encoder = XLNetEncoder(config, gpu_list, *args, **params)
#self.question_encoder = BertEncoder(config, gpu_list, *args, **params)
# self.attention = Attention(config, gpu_list, *args, **params)
self.dropout = nn.Dropout(config.getfloat("model", "dropout"))
self.dim_capsule = 128 #config.dim_capsule
self.num_compressed_capsule = 16 #config.num_compressed_capsule
self.ngram_size = [2, 4, 8]
self.convs_doc = nn.ModuleList([
nn.Conv1d(self.hidden_size, 32, K, stride=2)
for K in self.ngram_size
])
torch.nn.init.xavier_uniform_(self.convs_doc[0].weight)
torch.nn.init.xavier_uniform_(self.convs_doc[1].weight)
torch.nn.init.xavier_uniform_(self.convs_doc[2].weight)
self.primary_capsules_doc = PrimaryCaps(num_capsules=self.dim_capsule,
in_channels=32,
out_channels=32,
kernel_size=1,
stride=1)
self.flatten_capsules = FlattenCaps()
self.W_doc = nn.Parameter(
torch.FloatTensor(config.getint("model", "magic_param"),
self.num_compressed_capsule))
torch.nn.init.xavier_uniform_(self.W_doc)
self.fc_capsules_doc_child = FCCaps(
config,
output_capsule_num=4,
input_capsule_num=self.num_compressed_capsule,
in_channels=self.dim_capsule,
out_channels=self.dim_capsule)
self.bce = nn.CrossEntropyLoss(reduction='mean')
# self.gelu = nn.GELU()
# self.softmax = nn.Softmax(dim=1)
self.rank_module = nn.Linear(self.hidden_size * 2, 4)
self.accuracy_function = single_label_top1_accuracy