def test_bprop(self):
r = []
for i in xrange(self.N):
repeats = self.rng.random_integers(42)
axis = self.rng.randint(2)
input_dim, output_dim = self.rng.random_integers(2000, size=2)
x = self.get_normal_matrix(input_dim, output_dim)
input_dim = input_dim if axis else input_dim * repeats
true_labels = self.rng.randint(output_dim, size=(input_dim, 1)).astype(np.int32)
device_id = 0
output = {}
for processor_type in ['gpu', 'cpu']:
quagga.processor_type = processor_type
qx = Connector(Matrix.from_npa(x), device_id)
qtrue_labels = Connector(Matrix.from_npa(true_labels))
repeat_block = RepeatBlock(qx, repeats, axis)
sce_block = SoftmaxCeBlock(repeat_block.output, qtrue_labels)
qx.fprop()
qtrue_labels.fprop()
repeat_block.fprop()
sce_block.fprop()
sce_block.bprop()
repeat_block.bprop()
output[processor_type] = qx.backward_matrix.to_host()
r.append(np.allclose(output['gpu'], output['cpu']))
self.assertEqual(sum(r), len(r))
def test_bprop_vector(self):
r = []
for _ in xrange(self.N):
embd_dim = self.rng.random_integers(10000)
batch_size, output_dim = self.rng.random_integers(2000, size=2)
W = self.get_orthogonal_matrix(embd_dim, output_dim)
row_idxs = self.rng.randint(embd_dim, size=(batch_size, 1)).astype(np.int32)
true_labels = self.rng.randint(output_dim, size=(batch_size, 1)).astype(np.int32)
device_id = 0
output = {}
for processor_type in ['gpu', 'cpu']:
quagga.processor_type = processor_type
qrow_idxs = Connector(Matrix.from_npa(row_idxs))
qtrue_labels = Connector(Matrix.from_npa(true_labels))
qW = Connector(Matrix.from_npa(W), device_id)
row_slicing_block = RowSlicingBlock(qW, qrow_idxs)
sce_block = SoftmaxCeBlock(row_slicing_block.output, qtrue_labels)
qW.fprop()
qrow_idxs.fprop()
row_slicing_block.fprop()
sce_block.fprop()
sce_block.bprop()
row_slicing_block.bprop()
qW.add(Context(), qW.backward_matrix)
output[processor_type] = qW.to_host()
r.append(np.allclose(output['gpu'], output['cpu']))
self.assertEqual(sum(r), len(r))
def test_bprop(self):
r = []
for i in xrange(self.N):
matrices = []
ncols = self.rng.random_integers(1, 3000)
nrows = [0]
row_slices = []
device_ids = []
for _ in xrange(self.rng.random_integers(1, 10)):
_nrows = self.rng.random_integers(1, 2000)
nrows.append(nrows[-1] + _nrows)
if self.rng.choice([True, False]):
device_ids.append(0)
row_slices.append((nrows[-2], nrows[-1]))
else:
device_ids.append(None)
matrices.append(
self.rng.rand(_nrows, ncols).astype(np.float32))
true_labels = self.rng.randint(ncols, size=(nrows[-1],
1)).astype(np.int32)
if not row_slices:
r.append(True)
continue
output = {}
for processor_type in ['gpu', 'cpu']:
quagga.processor_type = processor_type
qmatrices = [
Connector(Matrix.from_npa(m), d_id)
for m, d_id in izip(matrices, device_ids)
]
qtrue_labels = Connector(Matrix.from_npa(true_labels))
vstack_block = VerticalStackBlock(*qmatrices)
sce_block = SoftmaxCeBlock(vstack_block.output, qtrue_labels)
for m in qmatrices:
m.fprop()
qtrue_labels.fprop()
vstack_block.fprop()
sce_block.fprop()
sce_block.bprop()
vstack_block.bprop()
output[processor_type] = [
m.backward_matrix.to_host() for m in qmatrices
if m.bpropagable
]
for dL_dm_gpu, dL_dm_cpu in izip(output['gpu'], output['cpu']):
if not np.allclose(dL_dm_gpu, dL_dm_cpu):
r.append(False)
break
else:
r.append(True)
self.assertEqual(sum(r), self.N)
def test_theano_grad(self):
quagga.processor_type = 'gpu'
r = []
for i in xrange(self.N):
for sparse in [True, False]:
batch_size, dim = self.rng.random_integers(2000, size=2)
if sparse:
true_labels = np.zeros((batch_size, dim), np.float32)
for k, j in enumerate(self.rng.randint(dim, size=batch_size)):
true_labels[k, j] = 1.0
else:
true_labels = self.rng.randint(dim, size=(batch_size, 1)).astype(np.int32)
x = self.rng.randn(batch_size, dim).astype(np.float32)
mask = (self.rng.rand(batch_size, 1) < 0.8).astype(np.float32)
device_id = 0
for with_mask in [False, True]:
# Theano model
th_x = T.fmatrix()
th_mask = T.fcol()
th_true_labels = T.fmatrix() if sparse else T.ivector()
if with_mask:
probs = T.nnet.softmax(th_mask * th_x)
else:
probs = T.nnet.softmax(th_x)
loss = T.mean(T.nnet.categorical_crossentropy(probs, th_true_labels))
if with_mask:
get_theano_grads = theano.function([th_x, th_true_labels, th_mask], T.grad(loss, wrt=th_x))
th_dL_dx = get_theano_grads(x, true_labels if sparse else true_labels[:, 0], mask)
else:
get_theano_grads = theano.function([th_x, th_true_labels], T.grad(loss, wrt=th_x))
th_dL_dx = get_theano_grads(x, true_labels if sparse else true_labels[:, 0])
# quagga model
x_gpu = Connector(Matrix.from_npa(x), device_id)
true_labels_gpu = Connector(Matrix.from_npa(true_labels))
mask_gpu = Connector(Matrix.from_npa(mask)) if with_mask else None
softmax_ce_block = SoftmaxCeBlock(x_gpu, true_labels_gpu, mask_gpu)
x_gpu.fprop()
true_labels_gpu.fprop()
if with_mask:
mask_gpu.fprop()
softmax_ce_block.fprop()
softmax_ce_block.bprop()
q_dL_dx = x_gpu.backward_matrix.to_host()
r.append(np.allclose(th_dL_dx, q_dL_dx))
self.assertEqual(sum(r), len(r))
def test_bprop(self):
r = []
for i in xrange(self.N):
matrices = []
nrows = self.rng.random_integers(1, 3000)
ncols = [0]
col_slices = []
device_ids = []
for _ in xrange(self.rng.random_integers(1, 10)):
_ncols = self.rng.random_integers(1, 2000)
ncols.append(ncols[-1] + _ncols)
if self.rng.choice([True, False]):
device_ids.append(0)
col_slices.append((ncols[-2], ncols[-1]))
else:
device_ids.append(None)
matrices.append(self.rng.rand(nrows, _ncols).astype(np.float32))
true_labels = self.rng.randint(ncols[-1], size=(nrows, 1)).astype(np.int32)
if not col_slices:
r.append(True)
continue
output = {}
for processor_type in ['gpu', 'cpu']:
quagga.processor_type = processor_type
qmatrices = [Connector(Matrix.from_npa(m), d_id) for m, d_id in izip(matrices, device_ids)]
qtrue_labels = Connector(Matrix.from_npa(true_labels))
hstack_block = HorizontalStackBlock(*qmatrices)
sce_block = SoftmaxCeBlock(hstack_block.output, qtrue_labels)
for m in qmatrices:
m.fprop()
qtrue_labels.fprop()
hstack_block.fprop()
sce_block.fprop()
sce_block.bprop()
hstack_block.bprop()
output[processor_type] = [m.backward_matrix.to_host()
for m in qmatrices if m.bpropagable]
for dL_dm_gpu, dL_dm_cpu in izip(output['gpu'], output['cpu']):
if not np.allclose(dL_dm_gpu, dL_dm_cpu):
r.append(False)
break
else:
r.append(True)
self.assertEqual(sum(r), self.N)
def test_bprop(self):
"""
compare `bprop` results for cpu and gpu backends
"""
r = []
for i in xrange(self.N):
for sparse in [True, False]:
batch_size, dim = self.rng.random_integers(2000, size=2)
if sparse:
true_labels = np.zeros((batch_size, dim), np.float32)
for k, j in enumerate(self.rng.randint(dim, size=batch_size)):
true_labels[k, j] = 1.0
else:
true_labels = self.rng.randint(dim, size=(batch_size, 1)).astype(np.int32)
x = self.rng.randn(batch_size, dim).astype(np.float32)
mask = (self.rng.rand(batch_size, 1) < 0.8).astype(np.float32)
device_id = 0
for with_mask in [False, True]:
quagga.processor_type = 'gpu'
x_gpu = Connector(Matrix.from_npa(x), device_id)
true_labels_gpu = Connector(Matrix.from_npa(true_labels))
mask_gpu = Connector(Matrix.from_npa(mask)) if with_mask else None
softmax_ce_block = SoftmaxCeBlock(x_gpu, true_labels_gpu, mask_gpu)
x_gpu.fprop()
true_labels_gpu.fprop()
if with_mask:
mask_gpu.fprop()
softmax_ce_block.fprop()
softmax_ce_block.bprop()
dL_dx_gpu = x_gpu.backward_matrix.to_host()
quagga.processor_type = 'cpu'
x_cpu = Connector(Matrix.from_npa(x), device_id)
true_labels_cpu = Connector(Matrix.from_npa(true_labels))
mask_cpu = Connector(Matrix.from_npa(mask)) if with_mask else None
softmax_ce_block = SoftmaxCeBlock(x_cpu, true_labels_cpu, mask_cpu)
x_cpu.fprop()
true_labels_cpu.fprop()
if with_mask:
mask_cpu.fprop()
softmax_ce_block.fprop()
softmax_ce_block.bprop()
dL_dx_cpu = x_cpu.backward_matrix.to_host()
r.append(np.allclose(dL_dx_gpu, dL_dx_cpu))
self.assertEqual(sum(r), len(r))
def test_theano_bprop_vector(self):
r = []
for _ in xrange(self.N):
embd_dim = self.rng.random_integers(10000)
batch_size, output_dim = self.rng.random_integers(2000, size=2)
W = self.get_orthogonal_matrix(embd_dim, output_dim)
row_idxs = self.rng.randint(embd_dim, size=(batch_size, 1)).astype(np.int32)
true_labels = self.rng.randint(output_dim, size=(batch_size, 1)).astype(np.int32)
device_id = 0
quagga.processor_type = 'gpu'
qrow_idxs = Connector(Matrix.from_npa(row_idxs))
qW = Connector(Matrix.from_npa(W), device_id)
qtrue_labels = Connector(Matrix.from_npa(true_labels))
row_slicing_block = RowSlicingBlock(qW, qrow_idxs)
sce_block = SoftmaxCeBlock(row_slicing_block.output, qtrue_labels)
qtrue_labels.fprop()
qW.fprop()
qrow_idxs.fprop()
row_slicing_block.fprop()
sce_block.fprop()
sce_block.bprop()
row_slicing_block.bprop()
qW.add(Context(), qW.backward_matrix)
th_row_idxs = T.ivector()
th_true_labels = T.ivector()
row_slicing_layer = RowSlicingLayer(W)
toutput = row_slicing_layer.get_output_expr(th_row_idxs)
loss = SoftmaxLayer.get_loss(toutput, th_true_labels)
dL_dW = T.grad(loss, row_slicing_layer.W)
fun = theano.function([th_row_idxs, th_true_labels],
updates=[(row_slicing_layer.W, row_slicing_layer.W + dL_dW)])
fun(row_idxs[:, 0], true_labels[:, 0])
r.append(np.allclose(qW.to_host(), row_slicing_layer.W.get_value()))
self.assertEqual(sum(r), len(r))
def test_theano_bprop(self):
r = []
for i in xrange(self.N):
repeats = self.rng.random_integers(42)
axis = self.rng.randint(2)
input_dim, output_dim = self.rng.random_integers(2000, size=2)
x = self.get_normal_matrix(input_dim, output_dim)
input_dim = input_dim if axis else input_dim * repeats
true_labels = self.rng.randint(output_dim, size=(input_dim, 1)).astype(np.int32)
device_id = 0
quagga.processor_type = 'gpu'
qx = Connector(Matrix.from_npa(x), device_id)
qtrue_labels = Connector(Matrix.from_npa(true_labels))
repeat_block = RepeatBlock(qx, repeats, axis)
sce_block = SoftmaxCeBlock(repeat_block.output, qtrue_labels)
qx.fprop()
qtrue_labels.fprop()
repeat_block.fprop()
sce_block.fprop()
sce_block.bprop()
repeat_block.bprop()
q_dL_dx = qx.backward_matrix.to_host()
th_x = T.fmatrix()
th_true_labels = T.ivector()
reps = [1, 1]
reps[axis] = repeats
th_output = T.tile(th_x, reps)
th_output = T.nnet.softmax(th_output)
loss = T.mean(T.nnet.categorical_crossentropy(th_output, th_true_labels))
get_grads = theano.function([th_x, th_true_labels], T.grad(loss, th_x))
th_dL_dx = get_grads(x, true_labels[:, 0])
r.append(np.allclose(q_dL_dx, th_dL_dx))
self.assertEqual(sum(r), len(r))