def __init__(self, state_dim, action_dim, seq_len):
super(Actor, self).__init__()
# self.tcn = TCN(state_dim, state_dim, num_channels, kernel_size=kernel_size, dropout=dropout).to(device) # 预测所有状态
self.bn1 = nn.BatchNorm1d(state_dim)
self.tcn = TemporalConvNet(input_channels, num_channels, kernel_size=kernel_size, dropout=dropout)
self.fc1 = nn.Linear(num_channels[-1], state_dim)
self.fc2 = nn.Linear(num_channels[-1], action_dim)
def __init__(self,
input_size,
output_size,
num_channels,
kernel_size=2,
dropout=0.3,
emb_dropout=0.1,
tied_weights=False):
super(TCN, self).__init__()
self.encoder = nn.Embedding(output_size, input_size)
self.tcn = TemporalConvNet(input_size,
num_channels,
kernel_size,
dropout=dropout)
self.decoder = nn.Linear(num_channels[-1], output_size)
if tied_weights:
if num_channels[-1] != input_size:
raise ValueError(
'When using the tied flag, nhid must be equal to emsize')
self.decoder.weight = self.encoder.weight
print("Weight tied")
self.drop = nn.Dropout(emb_dropout)
self.emb_dropout = emb_dropout
self.init_weights()
def __init__(self,
input_size,
output_size,
num_channels,
kernel_size=2,
dropout=0.3,
emb_dropout=0.1,
tied_weights=False):
super(TCN, self).__init__()
# 将one-hot encoding 部分送入编码器作为一个批量的词嵌入向量
# output_size为词汇量，input_size是词向量的长度
self.encoder = nn.Embedding(output_size, input_size)
self.tcn = TemporalConvNet(input_size,
num_channels,
kernel_size,
dropout=dropout)
# 定义最后线性变换的维度，即最后一个卷积层的通道数到所有词汇的映射
self.decoder = nn.Linear(num_channels[-1], output_size)
if tied_weights:
# 是否共享编码器与解码器的权重，默认值为共享
# 共享时需要保持隐藏单元数等于词嵌入的长度
# 此时将预测的向量认为是词嵌入向量
if num_channels[-1] != input_size:
raise ValueError(
'When using the tied flag, nhid must be equal to emsize')
self.decoder.weight = self.encoder.weight
print("Weight tied")
self.drop = nn.Dropout(emb_dropout)
# 对输入词嵌入进行dropout
self.emb_dropout = emb_dropout
self.init_weights()
def __init__(self, input_size, output_size, num_channels, kernel_size, dropout):
super(TCN, self).__init__()
# 构建tcn网络
self.tcn = TemporalConvNet(input_size, num_channels, kernel_size=kernel_size, dropout=dropout)
# 将网络输出映射到实际输出，使用线性映射
self.linear = nn.Linear(num_channels[-1], output_size)
self.init_weights()
def __init__(self,
input_size,
output_size,
num_channels,
kernel_size=2,
dropout=0.3,
emb_dropout=0.1,
tied_weights=False):
#input_size = 600
#output_size = 10000 vocab_size 不重复单词个数
#num_channels = [600,600,600,600]
#kernel_size = 3
#dropout = 0.45
#emb_dropout = 0.25
#tied_weights= True
super(TCN, self).__init__()
self.encoder = nn.Embedding(output_size,
input_size) #vocab_10000 * 600 随机初始化
self.tcn = TemporalConvNet(input_size,
num_channels,
kernel_size,
dropout=dropout)
self.decoder = nn.Linear(num_channels[-1], output_size)
if tied_weights:
if num_channels[-1] != input_size:
raise ValueError(
'When using the tied flag, nhid must be equal to emsize')
self.decoder.weight = self.encoder.weight
print("Weight tied")
self.drop = nn.Dropout(emb_dropout)
self.emb_dropout = emb_dropout
self.init_weights()
def __init__(self, input_size, output_size, num_channels, kernel_size,
dropout):
super(TCN, self).__init__()
self.tcn = TemporalConvNet(input_size,
num_channels,
kernel_size=kernel_size,
dropout=dropout)
self.linear = nn.Linear(num_channels[-1], output_size)
def __init__(self, input_size, output_size, num_channels, kernel_size,
dropout, use_fixup_init):
super(TCN, self).__init__()
self.tcn = TemporalConvNet(input_size,
num_channels,
kernel_size,
dropout=dropout,
use_fixup_init=use_fixup_init)
self.linear = nn.Linear(num_channels[-1], output_size)
self.sig = nn.Sigmoid()
def __init__(self,
args,
num_inputs,
num_actions,
hidden_size,
action_range=1.,
init_w=3e-3,
log_std_min=-20,
log_std_max=2):
super(PolicyNetwork, self).__init__()
# # Example of using Sequential
# model = nn.Sequential(
# nn.Conv2d(1,20,5),
# nn.ReLU(),
# nn.Conv2d(20,64,5),
# nn.ReLU()
# )
self.log_std_min = log_std_min
self.log_std_max = log_std_max
self.num_inputs = num_inputs
# self.output_dim = num_actions
# self.hidden_dim = hidden_size
# self.bn1 = nn.BatchNorm1d(state_dim)
self.tcn = TemporalConvNet(input_channels,
[args['nhid']] * args['levels'],
kernel_size=kernel_size,
dropout=args['dropout'])
# self.fc1 = nn.Linear(num_channels[-1], hidden_size)
self.model = nn.Sequential(
nn.Linear(num_channels[-1], hidden_size), # 输入10维，隐层20维
# nn.BatchNorm1d(hidden_size, affine=True, track_running_stats=True), # BN层，参数为隐层的个数
nn.LayerNorm(hidden_size, elementwise_affine=True),
nn.ReLU(), # 激活函数
# nn.Linear(hidden_size, hidden_size), # 输入10维，隐层20维
# # nn.BatchNorm1d(hidden_size, affine=True, track_running_stats=True), # BN层，参数为隐层的个数
# nn.LayerNorm(hidden_size, elementwise_affine=True),
# nn.ReLU(),
)
self.mean_linear = nn.Linear(hidden_size, num_actions)
self.mean_linear.weight.data.uniform_(-init_w, init_w)
self.mean_linear.bias.data.uniform_(-init_w, init_w)
self.log_std_linear = nn.Linear(hidden_size, num_actions)
self.log_std_linear.weight.data.uniform_(-init_w, init_w)
self.log_std_linear.bias.data.uniform_(-init_w, init_w)
self.action_range = action_range
self.num_actions = num_actions
def __init__(self, input_size, output_size, num_channels, kernel_size,
dropout):
# input_size = 88, output_size = 88, num_channels = [150] * 4, kernel_size = 5
super(TCN, self).__init__()
# With Batch size as N, Seq len as L and Number of features as C, initial data is N x L x C (C = 88)
self.tcn = TemporalConvNet(input_size,
num_channels,
kernel_size,
dropout=dropout)
self.linear = nn.Linear(num_channels[-1], output_size)
self.sig = nn.Sigmoid()
def build_tcn(inputs, tcn_dropout, kernel_size, num_channels):
# inputs = placeholder
# self.dropout = tf.placeholder_with_default(0., shape=())
# num_channels = [hidden1, hidden2, ...., outputchannel]
# kernel_size
tcn = TemporalConvNet(num_channels,
stride=1,
kernel_size=kernel_size,
dropout=tcn_dropout)
outputs = tcn(inputs)
return outputs
class TCN(nn.Module):
def __init__(self, input_size=1, output_size=10, num_channels=None, kernel_size=7, dropout=0.05):
super(TCN, self).__init__()
num_channels = [25] * 8 if num_channels is None else num_channels
self.tcn = TemporalConvNet(input_size, num_channels, kernel_size=kernel_size, dropout=dropout)
self.linear = nn.Linear(num_channels[-1], output_size)
self.receptive_field = self.tcn.receptive_field
def forward(self, inputs):
"""Inputs have to have dimension (N, C_in, L_in)"""
y1 = self.tcn(inputs) # input should have dimension (N, C, L)
o = self.linear(y1[:, :, -1])
return F.log_softmax(o, dim=1)
def single_forward(self, input):
return F.log_softmax(self.linear(self.tcn.single_forward(input).squeeze(dim=2)), dim=1)
def fast_inference(self, batch_size):
self.tcn.fast_inference(batch_size)
def compare(self, inputs):
y1 = self.tcn(inputs)
y2 = torch.zeros(y1.size()).to(y1.device)
for i in range(inputs.size()[2]):
y2[:,:,i] = self.tcn.single_forward(inputs[:,:,i].view(inputs.size()[0], inputs.size()[1], 1)).squeeze()
return (y1 == y2).all().item()
def reset_cache(self):
self.tcn.reset_cache()
def __init__(self, input_size, output_size, num_channels, kernel_size,
dropout):
# input_size = 1, output_size = 10, num_channels = [nhid] * levels = [10] * 8, kernel_size = 8
super(TCN, self).__init__()
# With Batch size as N, Seq len as L and 1 feature, initial data is 32 x 1 x 1020
self.tcn = TemporalConvNet(input_size,
num_channels,
kernel_size=kernel_size,
dropout=dropout)
# input to linear is N x L x nhid = 32 x 1020 x 10
self.linear = nn.Linear(num_channels[-1], output_size)
# input of linear is N x L x nhid = 32 x 1020 x 10
self.init_weights()
def __init__(self, input_size, output_size, num_channels, kernel_size,
dropout):
# input_size = 2, output_size = 1, num_channels = [nhid] * levels = [30] * 4, kernel_size = 7
super(TCN, self).__init__()
# With Batch size as N, Seq len as L and 2 features, initial data is N x 2 x L
self.tcn = TemporalConvNet(input_size,
num_channels,
kernel_size=kernel_size,
dropout=dropout)
# input to linear is N x nhid x l = N x 30 x 1 where l is last element of sequence
self.linear = nn.Linear(num_channels[-1], output_size)
# output of linear is N x 1 x 1
self.init_weights()
def __init__(self, input_size, output_size, num_channels, window_size,
num_fingers, kernel_size, dropout):
super(TCN, self).__init__()
self.tcn = TemporalConvNet(input_size,
num_channels,
kernel_size=kernel_size,
dropout=dropout)
self.linear1 = nn.Linear(
(8 + (window_size - 1) * num_fingers) * num_channels[-1],
(8 + (window_size - 1) * num_fingers) * num_channels[-1])
self.linear2 = nn.Linear(
(8 + (window_size - 1) * num_fingers) * num_channels[-1],
output_size)
def __init__(self, output_size: int, seq_length: int, num_channels: list,
kernel_size: int, dropout: float, dt):
super(LowResolutionTCN, self).__init__()
self.tcn = TemporalConvNet(output_size,
num_channels,
kernel_size=kernel_size,
dropout=dropout)
self.linear1 = nn.Linear(num_channels[-1] + 2, output_size)
# self.linear2 = nn.Linear(output_size, output_size)
# self.fc = nn.Sequential(self.linear1, nn.Sigmoid(), self.linear2)
self.euler_clock = TimePassing(dt)
self.output_size = output_size
self.init_weights()
def __init__(self,
word2vec_model,
embedding_dim,
hidden_dim,
output_dim,
num_layers=1,
dropout=0.2,
use_gdelt=False,
use_TCN=False,
effective_history=91):
"""
:param word2vec_model: the actual model that would embed the tweets
:param embedding_dim:
:param hidden_dim:
:param num_layers:
:param dropout:
"""
super().__init__()
self.word2vec_model = word2vec_model
self.embedding_dim = embedding_dim
self.hidden_dim = hidden_dim
self.output_dim = output_dim
self.use_gdelt = use_gdelt
self.use_TCN = use_TCN
if use_TCN:
num_levels, kernel_size = get_TCN_params_from_effective_history(
effective_history)
num_channels = [hidden_dim] * num_levels
self.temporal_extractor = TemporalConvNet(embedding_dim,
num_channels,
kernel_size, dropout)
else:
self.temporal_extractor = nn.LSTM(embedding_dim,
hidden_dim,
num_layers,
batch_first=True,
dropout=dropout)
# at the moment this is without considering additional info about the tweets like the number of mentions, etc...
# also the structure is arbitrary at the moment
self.num_handmade_features = 4
self.feature_extractor = nn.Linear(
hidden_dim + self.num_handmade_features, output_dim)
self.means = torch.Tensor([1.0901, 0.0135, 0.8929, 2.8000])
self.stds = torch.Tensor([17.6291, 0.1156, 0.3092, 1.9421])
def __init__(self,
input_size,
output_size,
num_channels,
kernel_size=2,
dropout=0.2,
emb_dropout=0.2):
super(TCN, self).__init__()
self.encoder = nn.Embedding(output_size, input_size)
self.tcn = TemporalConvNet(input_size,
num_channels,
kernel_size=kernel_size,
dropout=dropout)
self.decoder = nn.Linear(input_size, output_size)
self.decoder.weight = self.encoder.weight
self.drop = nn.Dropout(emb_dropout)
self.init_weights()
def __init__(self, input_size=1, output_size=10, num_channels=None, kernel_size=7, dropout=0.05):
super(TCN, self).__init__()
num_channels = [25] * 8 if num_channels is None else num_channels
self.tcn = TemporalConvNet(input_size, num_channels, kernel_size=kernel_size, dropout=dropout)
self.linear = nn.Linear(num_channels[-1], output_size)
self.receptive_field = self.tcn.receptive_field