1
0
mirror of https://github.com/microsoft/qlib.git synced 2026-07-07 21:11:50 +08:00

fix pytorch memory amount error

This commit is contained in:
Young
2021-11-02 20:41:39 +08:00
parent 3943b7001f
commit 4f2d6b0d84

View File

@@ -320,6 +320,7 @@ class TSDataSampler:
self.flt_data = flt_data.values
self.idx_map = self.flt_idx_map(self.flt_data, self.idx_map)
self.data_index = self.data_index[np.where(self.flt_data == True)[0]]
self.idx_map = self.idx_map2arr(self.idx_map)
self.start_idx, self.end_idx = self.data_index.slice_locs(
start=time_to_slc_point(start), end=time_to_slc_point(end)
@@ -328,6 +329,25 @@ class TSDataSampler:
del self.data # save memory
@staticmethod
def idx_map2arr(idx_map):
# pytorch data sampler will have better memory control without large dict or list
# - https://github.com/pytorch/pytorch/issues/13243
# - https://github.com/airctic/icevision/issues/613
# So we convert the dict into int array.
# The arr_map is expected to behave the same as idx_map
dtype = np.int32
# set a index out of bound to indicate the none existing
no_existing_idx = (np.iinfo(dtype).max, np.iinfo(dtype).max)
max_idx = max(idx_map.keys())
arr_map = []
for i in range(max_idx + 1):
arr_map.append(idx_map.get(i, no_existing_idx))
arr_map = np.array(arr_map, dtype=dtype)
return arr_map
@staticmethod
def flt_idx_map(flt_data, idx_map):
idx = 0