update impl

alphamind/benchmarks/benchmarks.py

@@ -20,7 +20,11 @@ if __name__ == '__main__':
     benchmark_standardize_with_group(3000, 10, 1000, 30)
     benchmark_standardize(30, 10, 50000)
     benchmark_standardize_with_group(30, 10, 5000, 5)
+    benchmark_standardize(50000, 50, 20)
+    benchmark_standardize_with_group(50000, 50, 20, 50)
     benchmark_winsorize_normal(3000, 10, 1000)
     benchmark_winsorize_normal_with_group(3000, 10, 1000, 30)
     benchmark_winsorize_normal(30, 10, 50000)
     benchmark_winsorize_normal_with_group(30, 10, 5000, 5)
+    benchmark_winsorize_normal(50000, 50, 20)
+    benchmark_winsorize_normal_with_group(50000, 50, 20, 50)

alphamind/data/impl.pyx

@@ -13,9 +13,9 @@ from libc.math cimport sqrt
 
 @cython.boundscheck(False)
 @cython.wraparound(False)
-cdef int max_groups(long[:] groups, long length) nogil:
+cdef int max_groups(long[:] groups, size_t length) nogil:
     cdef long curr_max = 0
-    cdef long i
+    cdef size_t i
     cdef long curr
 
     for i in range(length):
@@ -27,54 +27,76 @@ cdef int max_groups(long[:] groups, long length) nogil:
 @cython.boundscheck(False)
 @cython.wraparound(False)
 @cython.cdivision(True)
-cpdef np.ndarray[double, ndim=2] agg_mean(long[:] groups, double[:, :] x):
-    cdef long length = groups.shape[0]
-    cdef long width = x.shape[1]
+cdef double[:, :] agg_mean(long[:] groups, double[:, :] x, size_t length, size_t width):
     cdef long max_g = max_groups(groups, length)
     cdef double[:, :] res = np.zeros((max_g+1, width))
     cdef long[:] bin_count = np.zeros(max_g+1, dtype=int)
-    cdef long i
-    cdef long j
+    cdef size_t i
+    cdef size_t j
     cdef long curr
 
-    for i in range(length):
-        for j in range(width):
-            res[groups[i], j] += x[i, j]
-        bin_count[groups[i]] += 1
+    with nogil:
+        for i in range(length):
+            for j in range(width):
+                res[groups[i], j] += x[i, j]
+            bin_count[groups[i]] += 1
+
+        for i in range(res.shape[0]):
+            curr = bin_count[i]
+            if curr != 0:
+                for j in range(width):
+                    res[i, j] /= curr
+    return res
+
+
+@cython.boundscheck(False)
+@cython.wraparound(False)
+cpdef np.ndarray[double, ndim=2] transform(long[:] groups, double[:, :] x, str func):
+
+    cdef size_t length = x.shape[0]
+    cdef size_t width = x.shape[1]
+    cdef double[:, :] res_data = np.zeros((length, width))
+    cdef double[:, :] value_data = np.zeros((length, width))
+    cdef size_t i
+    cdef size_t j
+
+    if func == 'mean':
+        value_data = agg_mean(groups, x, length, width)
+    elif func == 'std':
+        value_data = agg_std(groups, x, length, width, ddof=1)
 
-    for i in range(res.shape[0]):
-        curr = bin_count[i]
-        if curr != 0:
+    with nogil:
+        for i in range(length):
             for j in range(width):
-                res[i, j] /= curr
-    return np.asarray(res)
+                res_data[i, j] = value_data[groups[i], j]
+
+    return np.asarray(res_data)
 
 
 @cython.boundscheck(False)
 @cython.wraparound(False)
 @cython.cdivision(True)
-cpdef np.ndarray[double, ndim=2] agg_std(long[:] groups, double[:, :] x, long ddof=1):
-    cdef long length = groups.shape[0]
-    cdef long width = x.shape[1]
+cdef double[:, :] agg_std(long[:] groups, double[:, :] x, size_t length, size_t width, long ddof=1):
     cdef long max_g = max_groups(groups, length)
     cdef double[:, :] running_sum_square = np.zeros((max_g+1, width))
     cdef double[:, :] running_sum = np.zeros((max_g+1, width))
     cdef long[:] bin_count = np.zeros(max_g+1, dtype=int)
-    cdef long i
-    cdef long j
+    cdef size_t i
+    cdef size_t j
     cdef long curr
     cdef double raw_value
 
-    for i in range(length):
-        for j in range(width):
-            raw_value = x[i, j]
-            running_sum[groups[i], j] += raw_value
-            running_sum_square[groups[i], j] += raw_value * raw_value
-        bin_count[groups[i]] += 1
-
-    for i in range(running_sum_square.shape[0]):
-        curr = bin_count[i]
-        if curr > ddof:
+    with nogil:
+        for i in range(length):
             for j in range(width):
-                running_sum_square[i, j] = sqrt((running_sum_square[i, j] - running_sum[i, j] * running_sum[i, j] / curr) / (curr - ddof))
-    return np.asarray(running_sum_square)
\ No newline at end of file
+                raw_value = x[i, j]
+                running_sum[groups[i], j] += raw_value
+                running_sum_square[groups[i], j] += raw_value * raw_value
+            bin_count[groups[i]] += 1
+
+        for i in range(running_sum_square.shape[0]):
+            curr = bin_count[i]
+            if curr > ddof:
+                for j in range(width):
+                    running_sum_square[i, j] = sqrt((running_sum_square[i, j] - running_sum[i, j] * running_sum[i, j] / curr) / (curr - ddof))
+    return running_sum_square
\ No newline at end of file

alphamind/data/standardize.py

@@ -6,20 +6,14 @@ Created on 2017-4-25
 """
 
 import numpy as np
-from alphamind.data.impl import agg_mean
-from alphamind.data.impl import agg_std
+from alphamind.data.impl import transform
 
 
 def standardize(x: np.ndarray, groups: np.ndarray=None) -> np.ndarray:
 
     if groups is not None:
-        mean_values = agg_mean(groups, x)
-        std_values = agg_std(groups, x, ddof=1)
-
-        value_index = np.searchsorted(range(len(mean_values)), groups)
-
-        mean_values = mean_values[value_index]
-        std_values = std_values[value_index]
+        mean_values = transform(groups, x, 'mean')
+        std_values = transform(groups, x, 'std')
 
         return (x - mean_values) / std_values
     else:

alphamind/data/winsorize.py

@@ -6,29 +6,20 @@ Created on 2017-4-25
 """
 
 import numpy as np
-from alphamind.data.impl import agg_mean
-from alphamind.data.impl import agg_std
+from alphamind.data.impl import transform
 
 
 def winsorize_normal(x: np.ndarray, num_stds: int=3, groups: np.ndarray=None) -> np.ndarray:
 
     if groups is not None:
-        mean_values = agg_mean(groups, x)
-        std_values = agg_std(groups, x, ddof=1)
-
-        value_index = np.searchsorted(range(len(mean_values)), groups)
-
-        ubound = mean_values + num_stds * std_values
-        lbound = mean_values - num_stds * std_values
-
-        ubound = ubound[value_index]
-        lbound = lbound[value_index]
+        mean_values = transform(groups, x, 'mean')
+        std_values = transform(groups, x, 'std')
     else:
         std_values = x.std(axis=0)
         mean_values = x.mean(axis=0)
 
-        ubound = mean_values + num_stds * std_values
-        lbound = mean_values - num_stds * std_values
+    ubound = mean_values + num_stds * std_values
+    lbound = mean_values - num_stds * std_values
 
     res = np.where(x > ubound, ubound, np.where(x < lbound, lbound, x))