mirror of
https://github.com/microsoft/qlib.git
synced 2026-07-02 10:31:00 +08:00
32c3070b73
* Run ddg-da successfully * Support include valid; More parameters * Support L2 reg & visualization * Blackformat * Enable fill_method * Support specify handler & optim dataset * Fix Pylint
114 lines
3.9 KiB
Python
114 lines
3.9 KiB
Python
# Copyright (c) Microsoft Corporation.
|
|
# Licensed under the MIT License.
|
|
|
|
import numpy as np
|
|
import torch
|
|
from torch import nn
|
|
|
|
from qlib.constant import EPS
|
|
from qlib.log import get_module_logger
|
|
|
|
|
|
class ICLoss(nn.Module):
|
|
def forward(self, pred, y, idx, skip_size=50):
|
|
"""forward.
|
|
FIXME:
|
|
- Some times it will be a slightly different from the result from `pandas.corr()`
|
|
- It may be caused by the precision problem of model;
|
|
|
|
:param pred:
|
|
:param y:
|
|
:param idx: Assume the level of the idx is (date, inst), and it is sorted
|
|
"""
|
|
prev = None
|
|
diff_point = []
|
|
for i, (date, inst) in enumerate(idx):
|
|
if date != prev:
|
|
diff_point.append(i)
|
|
prev = date
|
|
diff_point.append(None)
|
|
# The lengths of diff_point will be one more larger then diff_point
|
|
|
|
ic_all = 0.0
|
|
skip_n = 0
|
|
for start_i, end_i in zip(diff_point, diff_point[1:]):
|
|
pred_focus = pred[start_i:end_i] # TODO: just for fake
|
|
if pred_focus.shape[0] < skip_size:
|
|
# skip some days which have very small amount of stock.
|
|
skip_n += 1
|
|
continue
|
|
y_focus = y[start_i:end_i]
|
|
if pred_focus.std() < EPS or y_focus.std() < EPS:
|
|
# These cases often happend at the end of test data.
|
|
# Usually caused by fillna(0.)
|
|
skip_n += 1
|
|
continue
|
|
|
|
ic_day = torch.dot(
|
|
(pred_focus - pred_focus.mean()) / np.sqrt(pred_focus.shape[0]) / pred_focus.std(),
|
|
(y_focus - y_focus.mean()) / np.sqrt(y_focus.shape[0]) / y_focus.std(),
|
|
)
|
|
ic_all += ic_day
|
|
if len(diff_point) - 1 - skip_n <= 0:
|
|
raise ValueError("No enough data for calculating IC")
|
|
if skip_n > 0:
|
|
get_module_logger("ICLoss").info(
|
|
f"{skip_n} days are skipped due to zero std or small scale of valid samples."
|
|
)
|
|
ic_mean = ic_all / (len(diff_point) - 1 - skip_n)
|
|
return -ic_mean # ic loss
|
|
|
|
|
|
def preds_to_weight_with_clamp(preds, clip_weight=None, clip_method="tanh"):
|
|
"""
|
|
Clip the weights.
|
|
|
|
Parameters
|
|
----------
|
|
clip_weight: float
|
|
The clip threshold.
|
|
clip_method: str
|
|
The clip method. Current available: "clamp", "tanh", and "sigmoid".
|
|
"""
|
|
if clip_weight is not None:
|
|
if clip_method == "clamp":
|
|
weights = torch.exp(preds)
|
|
weights = weights.clamp(1.0 / clip_weight, clip_weight)
|
|
elif clip_method == "tanh":
|
|
weights = torch.exp(torch.tanh(preds) * np.log(clip_weight))
|
|
elif clip_method == "sigmoid":
|
|
# intuitively assume its sum is 1
|
|
if clip_weight == 0.0:
|
|
weights = torch.ones_like(preds)
|
|
else:
|
|
sm = nn.Sigmoid()
|
|
weights = sm(preds) * clip_weight # TODO: The clip_weight is useless here.
|
|
weights = weights / torch.sum(weights) * weights.numel()
|
|
else:
|
|
raise ValueError("Unknown clip_method")
|
|
else:
|
|
weights = torch.exp(preds)
|
|
return weights
|
|
|
|
|
|
class SingleMetaBase(nn.Module):
|
|
def __init__(self, hist_n, clip_weight=None, clip_method="clamp"):
|
|
# method can be tanh or clamp
|
|
super().__init__()
|
|
self.clip_weight = clip_weight
|
|
if clip_method in ["tanh", "clamp"]:
|
|
if self.clip_weight is not None and self.clip_weight < 1.0:
|
|
self.clip_weight = 1 / self.clip_weight
|
|
self.clip_method = clip_method
|
|
|
|
def is_enabled(self):
|
|
if self.clip_weight is None:
|
|
return True
|
|
if self.clip_method == "sigmoid":
|
|
if self.clip_weight > 0.0:
|
|
return True
|
|
else:
|
|
if self.clip_weight > 1.0:
|
|
return True
|
|
return False
|