def test_take(self):
idx = gpuarray.arange(0, 10000, 2, dtype=np.uint32)
for dtype in [np.float32, np.complex64]:
a = gpuarray.arange(0, 600000, dtype=np.uint32).astype(dtype)
a_host = a.get()
result = gpuarray.take(a, idx)
assert (a_host[idx.get()] == result.get()).all()
def test_take(self):
idx = gpuarray.arange(0, 10000, 2, dtype=np.uint32)
for dtype in [np.float32, np.complex64]:
a = gpuarray.arange(0, 600000, dtype=np.uint32).astype(dtype)
a_host = a.get()
result = gpuarray.take(a, idx)
assert (a_host[idx.get()] == result.get()).all()
def unpermute(self, x):
return gpuarray.take(x, self.old2new_fetch_indices)
def permute(self, x):
return gpuarray.take(x, self.new2old_fetch_indices)
scores = gpuarray.to_gpu((5 * np.random.randn(B, K)).astype(np.float32))
probs = gpuarray.to_gpu(np.random.rand(B, K).astype(np.float32))
maxscores = gpuarray.empty((B,), dtype=np.float32)
maxscoreids = gpuarray.empty((B,), dtype=np.uint32)
deltas = gpuarray.empty_like(scores)
sumdeltas = gpuarray.empty((B,), dtype=np.float32)
cpu_probs = np.empty((B, K), dtype=np.float32)
indices = np.random.randint(0, K, size=(N, B)).astype(np.uint32)
gpu_ind = gpuarray.empty((B,), dtype=np.uint32)
selected_probs = gpuarray.empty((B,), dtype=np.float32)
for i in range(10):
gpu_ind.set(indices[i])
gpuarray.take(probs, gpu_ind, out=selected_probs)
utils.scalar_sub(selected_probs, 1.0, selected_probs)
gpuarray.multi_put([selected_probs], gpu_ind, out=[probs])
#print probs
t1 = time.clock()
for i in range(N):
# get the softmax probs first
utils.max(scores, 1, maxscores, maxscoreids)
utils.sub_matvec(scores, maxscores, 0, deltas)
cumath.exp(deltas, out=deltas)
scm.sum(deltas, 1, sumdeltas)
utils.div_matvec(deltas, sumdeltas, 0, probs)
# probs.get(cpu_probs)
def test_take(self):
idx = gpuarray.arange(0, 200000, 2, dtype=np.uint32)
a = gpuarray.arange(0, 600000, 3, dtype=np.float32)
result = gpuarray.take(a, idx)
assert ((3 * idx).get() == result.get()).all()
scores = gpuarray.to_gpu((5 * np.random.randn(B, K)).astype(np.float32))
probs = gpuarray.to_gpu(np.random.rand(B, K).astype(np.float32))
maxscores = gpuarray.empty((B, ), dtype=np.float32)
maxscoreids = gpuarray.empty((B, ), dtype=np.uint32)
deltas = gpuarray.empty_like(scores)
sumdeltas = gpuarray.empty((B, ), dtype=np.float32)
cpu_probs = np.empty((B, K), dtype=np.float32)
indices = np.random.randint(0, K, size=(N, B)).astype(np.uint32)
gpu_ind = gpuarray.empty((B, ), dtype=np.uint32)
selected_probs = gpuarray.empty((B, ), dtype=np.float32)
for i in range(10):
gpu_ind.set(indices[i])
gpuarray.take(probs, gpu_ind, out=selected_probs)
utils.scalar_sub(selected_probs, 1.0, selected_probs)
gpuarray.multi_put([selected_probs], gpu_ind, out=[probs])
#print probs
t1 = time.clock()
for i in range(N):
# get the softmax probs first
utils.max(scores, 1, maxscores, maxscoreids)
utils.sub_matvec(scores, maxscores, 0, deltas)
cumath.exp(deltas, out=deltas)
scm.sum(deltas, 1, sumdeltas)
utils.div_matvec(deltas, sumdeltas, 0, probs)
# probs.get(cpu_probs)
def unpermute(self, x):
return gpuarray.take(x, self.old2new_fetch_indices)
def permute(self, x):
return gpuarray.take(x, self.new2old_fetch_indices)
def test_take(self):
idx = gpuarray.arange(0, 200000, 2, dtype=np.uint32)
a = gpuarray.arange(0, 600000, 3, dtype=np.float32)
result = gpuarray.take(a, idx)
assert ((3*idx).get() == result.get()).all()
