【gcForest源碼解讀】win_utils.py

一.序列數據

樣本\序列	1	2	3	4	5	6	7	8	9	10	11
Sample_1	A	B	E	C	D	E	F	G	A	B	C
Sample_2	B	A	B	C	D	E	F	G	A	B	C
Sample_2	A	B	C	D	E	F	G	A	B	C	D
Sample_2	A	D	C	D	E	F	G	A	B	C	M
Sample_5	N	A	B	C	D	E	F	G	A	B	C

原維度：shape =（#sample,#feature）= (5,11)

reshape：shape = (#sample,#feature,seq_len,height_y) = (n,c,h,w) = (5，1，11，1)   #序列數據👉height_y = 1（固定）

二.代碼註釋

以下代碼只提取了需要理解的部分，且爲修改後的代碼。

def get_windows(X, win_x, win_y, stride_x=1, stride_y=1, pad_x=0, pad_y=0):
    """
    parallizing get_windows
    Arguments:
        X (ndarray): n x c x h x w
    Return:
        X_win (ndarray): n x nh x nw x nc
    """
    n, c, h, w = X.shape
    nc = win_y * win_x * c
    nh = (h - win_x) / stride_x + 1
    nw = (w - win_y) / stride_y + 1
    X_win = np.empty(( nc, n * nh * nw), dtype=np.float32)

    Parallel(n_jobs=-1)(delayed(get_windows_channel)
(X, X_win, des_id, nw, nh, win_x, win_y, stride_x, stride_y) 
for des_id in range(c * win_x * win_y))
    
    X_win = X_win.reshape((n, nh, nw, nc))
    return X_win

 X_win.shape = (nc,n*nh*nw) →  (n,nh,nw,nc)  即

#windows/per sample（見上圖） ：    

#在x軸方向上窗口移動後的樣本的向量長度 / #pooled_feature_x：

#在y軸方向上窗口移動後的樣本的向量長度 / #pooled_feature_y：（序列數據：w = 1 , stride_y = 1）

注意：其中的Parallel()(delayed(get_windows_channel)...)代碼，是指在此時要運行如下代碼：

def get_windows_channel(X, X_win, des_id, nw, nh, win_x, win_y, stride_x, stride_y):
    """
    X: N x C x H x W
    X_win: N x nc x nh x nw
    (k, di, dj) in range(X.channle, win_y, win_x)
    """
    #des_id = (k * win_y + di) * win_x + dj
    dj = des_id % win_x
    di = des_id / win_x % win_y
    k = des_id / win_x / win_y
    src = X[:, k, dj:dj+nh*stride_x:stride_x, di:di+nw*stride_y:stride_y].ravel()
    des = X_win[des_id, :]
    np.copyto(des, src)

實際上就是：

for des_id in range(c * win_x * win_y) = range(nc):
    for k in range(c) = range(#feature) = range(1) = 0:
    for di in range(win_y):
    for dj in range(win_x):

des.shape = (n*nh*nw)

src.shape = [n * (dj:dj+nh*stride_x:stride_x) * (di:di+nw*stride_y:stride_y)]  #雙冒號爲切片(start:end:step)

src和des的維度應該是一致的。