63 lines
2.1 KiB
Python
63 lines
2.1 KiB
Python
import torch
|
|
from torch import nn
|
|
from layers.Transformer_EncDec import Encoder, EncoderLayer
|
|
from layers.SelfAttention_Family import FullAttention, AttentionLayer
|
|
from layers.Embed import PatchEmbedding
|
|
import timesfm
|
|
|
|
|
|
class Model(nn.Module):
|
|
def __init__(self, configs):
|
|
"""
|
|
patch_len: int, patch len for patch_embedding
|
|
stride: int, stride for patch_embedding
|
|
"""
|
|
super().__init__()
|
|
|
|
self.model = timesfm.TimesFM_2p5_200M_torch.from_pretrained("google/timesfm-2.5-200m-pytorch")
|
|
self.model.compile(
|
|
timesfm.ForecastConfig(
|
|
max_context=configs.seq_len,
|
|
max_horizon=configs.pred_len,
|
|
normalize_inputs=True,
|
|
use_continuous_quantile_head=True,
|
|
force_flip_invariance=True,
|
|
infer_is_positive=True,
|
|
fix_quantile_crossing=True,
|
|
)
|
|
)
|
|
|
|
self.task_name = configs.task_name
|
|
self.seq_len = configs.seq_len
|
|
self.pred_len = configs.pred_len
|
|
|
|
def forecast(self, x_enc, x_mark_enc, x_dec, x_mark_dec):
|
|
means = x_enc.mean(1, keepdim=True).detach()
|
|
x_enc = x_enc.sub(means)
|
|
stdev = torch.sqrt(
|
|
torch.var(x_enc, dim=1, keepdim=True, unbiased=False) + 1e-5)
|
|
x_enc = x_enc.div(stdev)
|
|
|
|
B, L, C = x_enc.shape
|
|
device = x_enc.device
|
|
x_enc = torch.reshape(x_enc, (B*C, L))
|
|
|
|
output, _ = self.model.forecast(
|
|
horizon=self.pred_len,
|
|
inputs=x_enc.cpu().numpy()
|
|
)
|
|
output = torch.Tensor(output).to(device)
|
|
|
|
dec_out = torch.reshape(output, (B, output.shape[-1], C)).to(x_enc.device)
|
|
dec_out = dec_out * \
|
|
(stdev[:, 0, :].unsqueeze(1).repeat(1, self.pred_len, 1))
|
|
dec_out = dec_out + \
|
|
(means[:, 0, :].unsqueeze(1).repeat(1, self.pred_len, 1))
|
|
return dec_out
|
|
|
|
def forward(self, x_enc, x_mark_enc, x_dec, x_mark_dec, mask=None):
|
|
if self.task_name == 'zero_shot_forecast':
|
|
dec_out = self.forecast(x_enc, x_mark_enc, x_dec, x_mark_dec)
|
|
return dec_out
|
|
return None
|