merge with new gpu stuff

1 files changed, 55 insertions, 81 deletions

diff --git a/ot/gpu/da.py b/ot/gpu/da.py
index ac3f8d7..f5e7daa 100644
--- a/ot/gpu/da.py
+++ b/ot/gpu/da.py
@@ -11,79 +11,30 @@ Domain adaptation with optimal transport with GPU implementation
 # License: MIT License
 
 
-import numpy as np
-#from ..utils import unif
+import cupy as np  # np used for matrix computation
+import cupy as cp  # cp used for cupy specific operations
+import numpy as npp
+from . import utils
+
 from .bregman import sinkhorn
-import cudamat
 
 
-def pairwiseEuclideanGPU(a, b, returnAsGPU=False, squared=False):
+def sinkhorn_lpl1_mm(a, labels_a, b, M, reg, eta=0.1, numItermax=10,
+                     numInnerItermax=200, stopInnerThr=1e-9, verbose=False,
+                     log=False, to_numpy=True):
     """
-    Compute the pairwise euclidean distance between matrices a and b.
-
-
-    Parameters
-    ----------
-    a : np.ndarray (n, f)
-        first matrice
-    b : np.ndarray (m, f)
-        second matrice
-    returnAsGPU : boolean, optional (default False)
-        if True, returns cudamat matrix still on GPU, else return np.ndarray
-    squared : boolean, optional (default False)
-        if True, return squared euclidean distance matrice
+    Solve the entropic regularization optimal transport problem with nonconvex
+    group lasso regularization on GPU
 
+    If the input matrix are in numpy format, they will be uploaded to the
+    GPU first which can incur significant time overhead.
 
-    Returns
-    -------
-    c : (n x m) np.ndarray or cudamat.CUDAMatrix
-        pairwise euclidean distance distance matrix
-    """
-    # a is shape (n, f) and b shape (m, f). Return matrix c of shape (n, m).
-    # First compute in c_GPU the squared euclidean distance. And return its
-    # square root. At each cell [i,j] of c, we want to have
-    # sum{k in range(f)} ( (a[i,k] - b[j,k])^2 ). We know that
-    # (a-b)^2 = a^2 -2ab +b^2. Thus we want to have in each cell of c:
-    # sum{k in range(f)} ( a[i,k]^2 -2a[i,k]b[j,k] +b[j,k]^2).
-
-    a_GPU = cudamat.CUDAMatrix(a)
-    b_GPU = cudamat.CUDAMatrix(b)
-
-    # Multiply a by b transpose to obtain in each cell [i,j] of c the
-    # value sum{k in range(f)} ( a[i,k]b[j,k] )
-    c_GPU = cudamat.dot(a_GPU, b_GPU.transpose())
-    # multiply by -2 to have sum{k in range(f)} ( -2a[i,k]b[j,k] )
-    c_GPU.mult(-2)
-
-    # Compute the vectors of the sum of squared elements.
-    a_GPU = cudamat.pow(a_GPU, 2).sum(axis=1)
-    b_GPU = cudamat.pow(b_GPU, 2).sum(axis=1)
-
-    # Add the vectors in each columns (respectivly rows) of c.
-    # sum{k in range(f)} ( a[i,k]^2 -2a[i,k]b[j,k] )
-    c_GPU.add_col_vec(a_GPU)
-    # sum{k in range(f)} ( a[i,k]^2 -2a[i,k]b[j,k] +b[j,k]^2)
-    c_GPU.add_row_vec(b_GPU.transpose())
-
-    if not squared:
-        c_GPU = cudamat.sqrt(c_GPU)
-
-    if returnAsGPU:
-        return c_GPU
-    else:
-        return c_GPU.asarray()
-
-
-def sinkhorn_lpl1_mm(a, labels_a, b, M_GPU, reg, eta=0.1, numItermax=10,
-                     numInnerItermax=200, stopInnerThr=1e-9,
-                     verbose=False, log=False):
-    """
-    Solve the entropic regularization optimal transport problem with nonconvex group lasso regularization
 
     The function solves the following optimization problem:
 
     .. math::
-        \gamma = arg\min_\gamma <\gamma,M>_F + reg\cdot\Omega_e(\gamma)+ \eta \Omega_g(\gamma)
+        \gamma = arg\min_\gamma <\gamma,M>_F + reg\cdot\Omega_e(\gamma)
+        + \eta \Omega_g(\gamma)
 
         s.t. \gamma 1 = a
 
@@ -93,11 +44,16 @@ def sinkhorn_lpl1_mm(a, labels_a, b, M_GPU, reg, eta=0.1, numItermax=10,
     where :
 
     - M is the (ns,nt) metric cost matrix
-    - :math:`\Omega_e` is the entropic regularization term :math:`\Omega_e(\gamma)=\sum_{i,j} \gamma_{i,j}\log(\gamma_{i,j})`
-    - :math:`\Omega_g` is the group lasso  regulaization term :math:`\Omega_g(\gamma)=\sum_{i,c} \|\gamma_{i,\mathcal{I}_c}\|^{1/2}_1`   where  :math:`\mathcal{I}_c` are the index of samples from class c in the source domain.
+    - :math:`\Omega_e` is the entropic regularization term
+        :math:`\Omega_e(\gamma)=\sum_{i,j} \gamma_{i,j}\log(\gamma_{i,j})`
+    - :math:`\Omega_g` is the group lasso  regulaization term
+      :math:`\Omega_g(\gamma)=\sum_{i,c} \|\gamma_{i,\mathcal{I}_c}\|^{1/2}_1`
+      where  :math:`\mathcal{I}_c` are the index of samples from class c
+      in the source domain.
     - a and b are source and target weights (sum to 1)
 
-    The algorithm used for solving the problem is the generalised conditional gradient as proposed in  [5]_ [7]_
+    The algorithm used for solving the problem is the generalised conditional
+    gradient as proposed in  [5]_ [7]_
 
 
     Parameters
@@ -108,7 +64,7 @@ def sinkhorn_lpl1_mm(a, labels_a, b, M_GPU, reg, eta=0.1, numItermax=10,
         labels of samples in the source domain
     b : np.ndarray (nt,)
         samples weights in the target domain
-    M_GPU : cudamat.CUDAMatrix (ns,nt)
+    M : np.ndarray (ns,nt)
         loss matrix
     reg : float
         Regularization term for entropic regularization >0
@@ -124,6 +80,8 @@ def sinkhorn_lpl1_mm(a, labels_a, b, M_GPU, reg, eta=0.1, numItermax=10,
         Print information along iterations
     log : bool, optional
         record log if True
+    to_numpy : boolean, optional (default True)
+        If true convert back the GPU array result to numpy format.
 
 
     Returns
@@ -137,8 +95,13 @@ def sinkhorn_lpl1_mm(a, labels_a, b, M_GPU, reg, eta=0.1, numItermax=10,
     References
     ----------
 
-    .. [5] N. Courty; R. Flamary; D. Tuia; A. Rakotomamonjy, "Optimal Transport for Domain Adaptation," in IEEE Transactions on Pattern Analysis and Machine Intelligence , vol.PP, no.99, pp.1-1
-    .. [7] Rakotomamonjy, A., Flamary, R., & Courty, N. (2015). Generalized conditional gradient: analysis of convergence and applications. arXiv preprint arXiv:1510.06567.
+    .. [5] N. Courty; R. Flamary; D. Tuia; A. Rakotomamonjy,
+       "Optimal Transport for Domain Adaptation," in IEEE
+       Transactions on Pattern Analysis and Machine Intelligence ,
+       vol.PP, no.99, pp.1-1
+    .. [7] Rakotomamonjy, A., Flamary, R., & Courty, N. (2015).
+       Generalized conditional gradient: analysis of convergence
+       and applications. arXiv preprint arXiv:1510.06567.
 
     See Also
     --------
@@ -147,29 +110,30 @@ def sinkhorn_lpl1_mm(a, labels_a, b, M_GPU, reg, eta=0.1, numItermax=10,
     ot.optim.cg : General regularized OT
 
     """
+
+    a, labels_a, b, M = utils.to_gpu(a, labels_a, b, M)
+
     p = 0.5
     epsilon = 1e-3
-    Nfin = len(b)
 
     indices_labels = []
-    classes = np.unique(labels_a)
+    labels_a2 = cp.asnumpy(labels_a)
+    classes = npp.unique(labels_a2)
     for c in classes:
-        idxc, = np.where(labels_a == c)
-        indices_labels.append(cudamat.CUDAMatrix(idxc.reshape(1, -1)))
+        idxc, = utils.to_gpu(npp.where(labels_a2 == c))
+        indices_labels.append(idxc)
 
-    Mreg_GPU = cudamat.empty(M_GPU.shape)
-    W_GPU = cudamat.empty(M_GPU.shape).assign(0)
+    W = np.zeros(M.shape)
 
     for cpt in range(numItermax):
-        Mreg_GPU.assign(M_GPU)
-        Mreg_GPU.add_mult(W_GPU, eta)
-        transp_GPU = sinkhorn(a, b, Mreg_GPU, reg, numItermax=numInnerItermax,
-                              stopThr=stopInnerThr, returnAsGPU=True)
+        Mreg = M + eta * W
+        transp = sinkhorn(a, b, Mreg, reg, numItermax=numInnerItermax,
+                          stopThr=stopInnerThr, to_numpy=False)
         # the transport has been computed. Check if classes are really
         # separated
-        W_GPU.assign(1)
-        W_GPU = W_GPU.transpose()
+        W = np.ones(M.shape)
         for (i, c) in enumerate(classes):
+<<<<<<< HEAD
             (_, nbRow) = indices_labels[i].shape
             tmpC_GPU = cudamat.empty((Nfin, nbRow)).assign(0)
             transp_GPU.transpose().select_columns(indices_labels[i], tmpC_GPU)
@@ -184,3 +148,13 @@ def sinkhorn_lpl1_mm(a, labels_a, b, M_GPU, reg, eta=0.1, numItermax=10,
         W_GPU = W_GPU.transpose()
 
     return transp_GPU.asarray()
+=======
+            majs = np.sum(transp[indices_labels[i]], axis=0)
+            majs = p * ((majs + epsilon)**(p - 1))
+            W[indices_labels[i]] = majs
+
+    if to_numpy:
+        return utils.to_np(transp)
+    else:
+        return transp
+>>>>>>> master