做外贸网站流程,网站建设有哪些板块,软文网站外包,整合营销传播方法包括Lag-Llama: Towards Foundation Models for Time Series Forecasting 文章内容#xff1a; 时间序列预测任务#xff0c;单变量预测单变量#xff0c;基于Llama大模型#xff0c;在zero-shot场景下模型表现优异。创新点#xff0c;引入滞后特征作为协变量来进行预测。 获得… Lag-Llama: Towards Foundation Models for Time Series Forecasting 文章内容 时间序列预测任务单变量预测单变量基于Llama大模型在zero-shot场景下模型表现优异。创新点引入滞后特征作为协变量来进行预测。 获得不同频率的lag来自glunoTS库里面的源码
def _make_lags(middle: int, delta: int) - np.ndarray:Create a set of lags around a middle point including /- delta.return np.arange(middle - delta, middle delta 1).tolist()def get_lags_for_frequency(freq_str: str,lag_ub: int 1200,num_lags: Optional[int] None,num_default_lags: int 7,
) - List[int]:Generates a list of lags that that are appropriate for the given frequencystring.By default all frequencies have the following lags: [1, 2, 3, 4, 5, 6, 7].Remaining lags correspond to the same season (/- delta) in previousk cycles. Here delta and k are chosen according to the existing code.Parameters----------freq_strFrequency string of the form [multiple][granularity] such as 12H,5min, 1D etc.lag_ubThe maximum value for a lag.num_lagsMaximum number of lags; by default all generated lags are returned.num_default_lagsThe number of default lags; by default it is 7.# Lags are target values at the same season (/- delta) but in the# previous cycle.def _make_lags_for_second(multiple, num_cycles3):# We use previous num_cycles hours to generate lagsreturn [_make_lags(k * 60 // multiple, 2) for k in range(1, num_cycles 1)]def _make_lags_for_minute(multiple, num_cycles3):# We use previous num_cycles hours to generate lagsreturn [_make_lags(k * 60 // multiple, 2) for k in range(1, num_cycles 1)]def _make_lags_for_hour(multiple, num_cycles7):# We use previous num_cycles days to generate lagsreturn [_make_lags(k * 24 // multiple, 1) for k in range(1, num_cycles 1)]def _make_lags_for_day(multiple, num_cycles4, days_in_week7, days_in_month30):# We use previous num_cycles weeks to generate lags# We use the last month (in addition to 4 weeks) to generate lag.return [_make_lags(k * days_in_week // multiple, 1)for k in range(1, num_cycles 1)] [_make_lags(days_in_month // multiple, 1)]def _make_lags_for_week(multiple, num_cycles3):# We use previous num_cycles years to generate lags# Additionally, we use previous 4, 8, 12 weeksreturn [_make_lags(k * 52 // multiple, 1) for k in range(1, num_cycles 1)] [[4 // multiple, 8 // multiple, 12 // multiple]]def _make_lags_for_month(multiple, num_cycles3):# We use previous num_cycles years to generate lagsreturn [_make_lags(k * 12 // multiple, 1) for k in range(1, num_cycles 1)]# multiple, granularity get_granularity(freq_str)offset to_offset(freq_str)# normalize offset name, so that both W and W-SUN refer to Woffset_name norm_freq_str(offset.name)if offset_name A:lags []elif offset_name Q:assert (offset.n 1), Only multiple 1 is supported for quarterly. Use x month instead.lags _make_lags_for_month(offset.n * 3.0)elif offset_name M:lags _make_lags_for_month(offset.n)elif offset_name W:lags _make_lags_for_week(offset.n)elif offset_name D:lags _make_lags_for_day(offset.n) _make_lags_for_week(offset.n / 7.0)elif offset_name B:lags _make_lags_for_day(offset.n, days_in_week5, days_in_month22) _make_lags_for_week(offset.n / 5.0)elif offset_name H:lags (_make_lags_for_hour(offset.n) _make_lags_for_day(offset.n / 24) _make_lags_for_week(offset.n / (24 * 7)))# minuteselif offset_name T:lags (_make_lags_for_minute(offset.n) _make_lags_for_hour(offset.n / 60) _make_lags_for_day(offset.n / (60 * 24)) _make_lags_for_week(offset.n / (60 * 24 * 7)))# secondelif offset_name S:lags (_make_lags_for_second(offset.n) _make_lags_for_minute(offset.n / 60) _make_lags_for_hour(offset.n / (60 * 60)))else:raise Exception(invalid frequency)# flatten lags list and filterlags [int(lag) for sub_list in lags for lag in sub_list if 7 lag lag_ub]lags list(range(1, num_default_lags 1)) sorted(list(set(lags)))return lags[:num_lags]第一部分生成以middle为中心以delta为半径的区间[middle-delta,middledelta] ,这很好理解比如一周的周期是7天周期大小在7天附近波动很正常。
第二部分对于年月日时分秒这些不同的采样频率采用不同的具体的函数来确定lags其中有一个参数num_cycle,进一步利用了周期性我们考虑间隔1、2、3、…num个周期的时间点之间的联系 原理类似于这张图这种周期性的重复性体现在邻近的多个周期上 lag的用途
计算各类窗口大小
计算采样窗口大小
window_size estimator.context_length max(estimator.lags_seq) estimator.prediction_length# Here we make a window slightly bigger so that instance sampler can sample from each window# An alternative is to have exact size and use different instance sampler (e.g. ValidationSplitSampler)
window_size 10 * window_size# We change ValidationSplitSampler to add min_pastestimator.validation_sampler ValidationSplitSampler(min_pastestimator.context_length max(estimator.lags_seq),min_futureestimator.prediction_length,)
构建静态特征
lags lagged_sequence_values(self.lags_seq, prior_input, input, dim-1)#构建一个包含给定序列的滞后值的数组static_feat torch.cat((loc.abs().log1p(), scale.log()), dim-1)
expanded_static_feat unsqueeze_expand(static_feat, dim-2, sizelags.shape[-2]
)return torch.cat((lags, expanded_static_feat, time_feat), dim-1), loc, scale数据集准备过程
对每个数据集采样window_size13500,也挺离谱的 train_data, val_data [], []for name in TRAIN_DATASET_NAMES:new_data create_sliding_window_dataset(name, window_size)train_data.append(new_data)new_data create_sliding_window_dataset(name, window_size, is_trainFalse)val_data.append(new_data)采样的具体过程这里有个问题样本数量很小的数据集实际采样窗口大小小于设定的window_size,后续会如何对齐呢
文章设置单变量预测单变量所以样本进行了通道分离同一样本的不同特征被采样为不同的样本
def create_sliding_window_dataset(name, window_size, is_trainTrue):#划分非重叠的滑动窗口数据集window_size是对数据集采样的数量对每个数据集只取前windowsize个样本# Splits each time series into non-overlapping sliding windowsglobal_id 0freq get_dataset(name, pathdataset_path).metadata.freq#从数据集中获取时间频率data ListDataset([], freqfreq)#创建空数据集dataset get_dataset(name, pathdataset_path).train if is_train else get_dataset(name, pathdataset_path).test#获取原始数据集for x in dataset:windows []#划分滑动窗口#target:滑动窗口的目标值#start:滑动窗口的起始位置#item_id,唯一标识符#feat_static_cat:静态特征数组for i in range(0, len(x[target]), window_size):windows.append({target: x[target][i:iwindow_size],start: x[start] i,item_id: str(global_id),feat_static_cat: np.array([0]),})global_id 1data ListDataset(windows, freqfreq)return data合并数据集
# Here weights are proportional to the number of time series (sliding windows)weights [len(x) for x in train_data]# Here weights are proportinal to the number of individual points in all time series# weights [sum([len(x[target]) for x in d]) for d in train_data]train_data CombinedDataset(train_data, weightsweights)val_data CombinedDataset(val_data, weightsweights)class CombinedDataset:def __init__(self, datasets, seedNone, weightsNone):self._seed seedself._datasets datasetsself._weights weightsn_datasets len(datasets)if weights is None:#如果未提供权重默认平均分配权重self._weights [1 / n_datasets] * n_datasetsdef __iter__(self):return CombinedDatasetIterator(self._datasets, self._seed, self._weights)def __len__(self):return sum([len(ds) for ds in self._datasets])网络结构
lagllama
class LagLlamaModel(nn.Module):def __init__(self,max_context_length: int,scaling: str,input_size: int,n_layer: int,n_embd: int,n_head: int,lags_seq: List[int],rope_scalingNone,distr_outputStudentTOutput(),num_parallel_samples: int 100,) - None:super().__init__()self.lags_seq lags_seqconfig LTSMConfig(n_layern_layer,n_embdn_embd,n_headn_head,block_sizemax_context_length,feature_sizeinput_size * (len(self.lags_seq)) 2 * input_size 6,rope_scalingrope_scaling,)self.num_parallel_samples num_parallel_samplesif scaling mean:self.scaler MeanScaler(keepdimTrue, dim1)elif scaling std:self.scaler StdScaler(keepdimTrue, dim1)else:self.scaler NOPScaler(keepdimTrue, dim1)self.distr_output distr_outputself.param_proj self.distr_output.get_args_proj(config.n_embd)self.transformer nn.ModuleDict(dict(wtenn.Linear(config.feature_size, config.n_embd),hnn.ModuleList([Block(config) for _ in range(config.n_layer)]),ln_fRMSNorm(config.n_embd),))主要是transformer里面首先是一个线性层然后加了n_layer个Block最后是RMSNorm,接下来解析Block的代码 Block
class Block(nn.Module):def __init__(self, config: LTSMConfig) - None:super().__init__()self.rms_1 RMSNorm(config.n_embd)self.attn CausalSelfAttention(config)self.rms_2 RMSNorm(config.n_embd)self.mlp MLP(config)self.y_cache Nonedef forward(self, x: torch.Tensor, is_test: bool) - torch.Tensor:if is_test and self.y_cache is not None:# Only use the most recent one, rest is in cachex x[:, -1:]x x self.attn(self.rms_1(x), is_test)y x self.mlp(self.rms_2(x))if is_test:if self.y_cache is None:self.y_cache y # Build cacheelse:self.y_cache torch.cat([self.y_cache, y], dim1)[:, 1:] # Update cachereturn y
代码看到这里不太想继续看了太多glunoTS库里面的函数了我完全不熟悉这个库看起来太痛苦了还有很多的困惑最大的困惑就是数据是怎么对齐的怎么输入到Llama里面的慢慢看吧
其他
来源
