def test_convolution_backward_bias(self):
logging.debug("ENTER: test_convolution_backward_bias")
bperr_data = numpy.zeros((100, 64, 104, 226), dtype=numpy.float32)
bperr_data[:] = 0.1
bperr_desc = cudnn.TensorDescriptor()
bperr_desc.set_4d(cudnn.CUDNN_TENSOR_NCHW, cudnn.CUDNN_DATA_FLOAT,
*bperr_data.shape)
bperr_buf = cu.MemAlloc(self.ctx, bperr_data)
gd_data = numpy.zeros(64, dtype=numpy.float32)
gd_desc = cudnn.TensorDescriptor()
gd_desc.set_4d(cudnn.CUDNN_TENSOR_NCHW, cudnn.CUDNN_DATA_FLOAT, 1,
gd_data.size, 1, 1)
gd_buf = cu.MemAlloc(self.ctx, gd_data)
alpha = numpy.ones(1, dtype=numpy.float32)
beta = numpy.zeros(1, dtype=numpy.float32)
self.cudnn.convolution_backward_bias(alpha, bperr_desc, bperr_buf,
beta, gd_desc, gd_buf)
gd_buf.to_host(gd_data)
self.assertEqual(numpy.count_nonzero(gd_data), gd_data.size)
logging.debug("EXIT: test_convolution_backward_bias")
def test_transform_tensor(self):
logging.debug("ENTER: test_transform_tensor")
sh_interleaved = (2, 5, 5, 3)
sh_splitted = (2, 3, 5, 5)
inp_data = numpy.arange(numpy.prod(sh_interleaved),
dtype=numpy.float32).reshape(sh_interleaved)
inp_desc = cudnn.TensorDescriptor()
inp_desc.set_4d(cudnn.CUDNN_TENSOR_NHWC, cudnn.CUDNN_DATA_FLOAT,
*sh_splitted)
inp_buf = cu.MemAlloc(self.ctx, inp_data)
out_data = numpy.zeros(sh_splitted, dtype=numpy.float32)
out_desc = cudnn.TensorDescriptor()
out_desc.set_4d(cudnn.CUDNN_TENSOR_NCHW, cudnn.CUDNN_DATA_FLOAT,
*sh_splitted)
out_buf = cu.MemAlloc(self.ctx, out_data)
alpha = numpy.ones(1, dtype=numpy.float32)
beta = numpy.zeros(1, dtype=numpy.float32)
self.cudnn.transform_tensor(alpha, inp_desc, inp_buf, beta, out_desc,
out_buf)
out_buf.to_host(out_data)
max_diff = numpy.fabs(out_data - inp_data.transpose(0, 3, 1, 2)).max()
self.assertEqual(max_diff, 0.0)
logging.debug("EXIT: test_transform_tensor")
def test_convolution_backward_data(self):
logging.debug("ENTER: test_convolution_backward_data")
conv_desc = cudnn.ConvolutionDescriptor()
conv_desc.set_2d(5, 5, 1, 1)
inp_data = numpy.zeros((100, 8, 96, 96), dtype=numpy.float32)
inp_desc = cudnn.TensorDescriptor()
inp_desc.set_4d(cudnn.CUDNN_TENSOR_NCHW, cudnn.CUDNN_DATA_FLOAT,
*inp_data.shape)
inp_buf = cu.MemAlloc(self.ctx, inp_data)
filter_data = numpy.zeros((64, 8, 11, 11), dtype=numpy.float32)
filter_data[:] = 0.1
filter_desc = cudnn.FilterDescriptor()
filter_desc.set_4d(cudnn.CUDNN_DATA_FLOAT, *filter_data.shape)
filter_buf = cu.MemAlloc(self.ctx, filter_data)
bperr_data = numpy.zeros((100, 64, 96, 96), dtype=numpy.float32)
bperr_data[:] = 0.1
bperr_desc = cudnn.TensorDescriptor()
bperr_desc.set_4d(cudnn.CUDNN_TENSOR_NCHW, cudnn.CUDNN_DATA_FLOAT,
*bperr_data.shape)
bperr_buf = cu.MemAlloc(self.ctx, bperr_data)
alpha = numpy.ones(1, dtype=numpy.float32)
beta = numpy.zeros(1, dtype=numpy.float32)
self.cudnn.convolution_backward_data(alpha, filter_desc, filter_buf,
bperr_desc, bperr_buf, conv_desc,
beta, inp_desc, inp_buf)
inp_buf.to_host(inp_data)
self.assertEqual(numpy.count_nonzero(inp_data), inp_data.size)
if self.cudnn.version >= 4000:
algo = self.cudnn.get_convolution_backward_data_algorithm(
filter_desc, bperr_desc, conv_desc, inp_desc)
logging.debug("Fastest algo is %d", algo)
sz = self.cudnn.get_convolution_backward_data_workspace_size(
filter_desc, bperr_desc, conv_desc, inp_desc, algo)
logging.debug("Workspace size for it is %d", sz)
algo = self.cudnn.get_convolution_backward_data_algorithm(
filter_desc, bperr_desc, conv_desc, inp_desc,
cudnn.CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT,
512 * 1024 * 1024)
logging.debug("With 512 Mb limit: %d", algo)
workspace = cu.MemAlloc(self.ctx, 512 * 1024 * 1024)
inp_buf.memset32_async()
self.cudnn.convolution_backward_data(alpha, filter_desc,
filter_buf, bperr_desc,
bperr_buf, conv_desc, beta,
inp_desc, inp_buf, algo,
workspace, workspace.size)
inp_buf.to_host(inp_data)
self.assertEqual(numpy.count_nonzero(inp_data), inp_data.size)
logging.debug("EXIT: test_convolution_backward_data")
def test_pooling(self):
logging.debug("ENTER: test_pooling")
input_data = numpy.zeros((5, 96, 64, 48), dtype=numpy.float32)
input_data[:] = numpy.random.rand(input_data.size).reshape(
input_data.shape) - 0.5
input_desc = cudnn.TensorDescriptor()
input_desc.set_4d(cudnn.CUDNN_TENSOR_NCHW, cudnn.CUDNN_DATA_FLOAT,
*input_data.shape)
input_buf = cu.MemAlloc(self.ctx, input_data)
pooling_desc = cudnn.PoolingDescriptor()
pooling_desc.set_2d((5, 3), (2, 1), (3, 2))
if self.cudnn.version < 4000:
output_shape = (5, 96, 22, 24)
else:
output_shape = cudnn.CUDNN.get_pooling_2d_forward_output_dim(
pooling_desc, input_desc)
self.assertEqual(len(output_shape), 4)
logging.debug("Output shape is %s", output_shape)
output_data = numpy.zeros(output_shape, dtype=numpy.float32)
output_data[:] = numpy.nan
output_desc = cudnn.TensorDescriptor()
output_desc.set_4d(cudnn.CUDNN_TENSOR_NCHW, cudnn.CUDNN_DATA_FLOAT,
*output_data.shape)
output_buf = cu.MemAlloc(self.ctx, output_data)
np_one = numpy.ones(1, dtype=numpy.float32)
np_zero = numpy.zeros(1, dtype=numpy.float32)
self.cudnn.pooling_forward(pooling_desc, np_one, input_desc, input_buf,
np_zero, output_desc, output_buf)
output_buf.to_host(output_data)
self.assertEqual(numpy.count_nonzero(numpy.isnan(output_data)), 0)
diff_desc = output_desc
diff_buf = cu.MemAlloc(self.ctx, output_data)
grad_desc = input_desc
grad_data = input_data.copy()
grad_data[:] = numpy.nan
grad_buf = cu.MemAlloc(self.ctx, grad_data)
self.cudnn.pooling_backward(pooling_desc, np_one, output_desc,
output_buf, diff_desc, diff_buf,
input_desc, input_buf, np_zero, grad_desc,
grad_buf)
grad_buf.to_host(grad_data)
self.assertEqual(numpy.count_nonzero(numpy.isnan(grad_data)), 0)
logging.debug("EXIT: test_pooling")
def _init_descriptors(self, include_out=False):
conv = cudnn.ConvolutionDescriptor()
conv.set_2d(5, 4, 2, 1)
inp = cudnn.TensorDescriptor()
inp.set_4d(cudnn.CUDNN_TENSOR_NCHW, cudnn.CUDNN_DATA_FLOAT, 100, 8,
208, 224)
filter = cudnn.FilterDescriptor()
filter.set_4d(cudnn.CUDNN_DATA_FLOAT, 64, 8, 11, 7)
if not include_out:
return conv, inp, filter
n, c, h, w = cudnn.CUDNN.get_convolution_2d_forward_output_dim(
conv, inp, filter)
out = cudnn.TensorDescriptor()
out.set_4d(cudnn.CUDNN_TENSOR_NCHW, cudnn.CUDNN_DATA_FLOAT, n, c, h, w)
return conv, inp, filter, out
def test_convolution_forward(self):
logging.debug("ENTER: test_convolution_forward")
conv_desc = cudnn.ConvolutionDescriptor()
conv_desc.set_2d(5, 4, 2, 1)
inp_data = numpy.zeros((100, 8, 104, 112), dtype=numpy.float32)
inp_data[:] = 0.1
inp_desc = cudnn.TensorDescriptor()
inp_desc.set_4d(cudnn.CUDNN_TENSOR_NCHW, cudnn.CUDNN_DATA_FLOAT,
*inp_data.shape)
inp_buf = cu.MemAlloc(self.ctx, inp_data)
filter_data = numpy.zeros((64, 8, 11, 7), dtype=numpy.float32)
filter_data[:] = 0.3
filter_desc = cudnn.FilterDescriptor()
filter_desc.set_4d(cudnn.CUDNN_DATA_FLOAT, *filter_data.shape)
filter_buf = cu.MemAlloc(self.ctx, filter_data)
n, c, h, w = cudnn.CUDNN.get_convolution_2d_forward_output_dim(
conv_desc, inp_desc, filter_desc)
out_data = numpy.zeros((n, c, h, w), dtype=numpy.float32)
out_desc = cudnn.TensorDescriptor()
out_desc.set_4d(cudnn.CUDNN_TENSOR_NCHW, cudnn.CUDNN_DATA_FLOAT,
*out_data.shape)
out_buf = cu.MemAlloc(self.ctx, out_data)
workspace = cu.MemAlloc(self.ctx, 512 * 1024 * 1024)
algo = self.cudnn.get_convolution_forward_algorithm(
inp_desc, filter_desc, conv_desc, out_desc,
cudnn.CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT,
workspace.size)
alpha = numpy.ones(1, dtype=numpy.float32)
beta = numpy.zeros(1, dtype=numpy.float32)
self.cudnn.convolution_forward(alpha, inp_desc, inp_buf, filter_desc,
filter_buf, conv_desc, algo, workspace,
workspace.size, beta, out_desc, out_buf)
out_buf.to_host(out_data)
self.assertEqual(numpy.count_nonzero(out_data), out_data.size)
logging.debug("EXIT: test_convolution_forward")
def test_tensor_descriptor(self):
logging.debug("ENTER: test_tensor_descriptor")
d = cudnn.TensorDescriptor()
self.assertIsNotNone(d.handle)
for dt in (cudnn.CUDNN_DATA_DOUBLE, cudnn.CUDNN_DATA_FLOAT):
for fmt in (cudnn.CUDNN_TENSOR_NCHW, cudnn.CUDNN_TENSOR_NHWC):
d.set_4d(fmt, dt, 100, 50, 217, 215)
self.assertEqual(d.fmt, fmt)
self.assertEqual(d.data_type, dt)
self.assertEqual(d.n, 100)
self.assertEqual(d.c, 50)
self.assertEqual(d.h, 217)
self.assertEqual(d.w, 215)
d.get_nd(4)
if fmt == cudnn.CUDNN_TENSOR_NCHW:
logging.debug("NCHW: %s %s", d.dims, d.strides)
self.assertEqual(d.dims, (100, 50, 217, 215))
self.assertEqual(d.strides,
(50 * 217 * 215, 217 * 215, 215, 1))
elif fmt == cudnn.CUDNN_TENSOR_NHWC:
logging.debug("NHWC: %s %s", d.dims, d.strides)
self.assertEqual(d.dims, (100, 50, 217, 215))
self.assertEqual(d.strides,
(50 * 217 * 215, 1, 215 * 50, 50))
d = cudnn.TensorDescriptor()
self.assertEqual(len(d.dims), 0)
self.assertEqual(len(d.strides), 0)
d.set_nd(cudnn.CUDNN_DATA_FLOAT, (2, 3, 4))
self.assertEqual(d.data_type, cudnn.CUDNN_DATA_FLOAT)
self.assertEqual(d.dims, (2, 3, 4))
self.assertEqual(d.strides, (12, 4, 1))
d._data_type = -1
d._fmt = -1
d._dims = tuple()
d.get_nd(3)
self.assertEqual(d.data_type, cudnn.CUDNN_DATA_FLOAT)
self.assertEqual(d.dims, (2, 3, 4))
self.assertEqual(d.strides, (12, 4, 1))
logging.debug("EXIT: test_tensor_descriptor")
def test_softmax(self):
logging.debug("ENTER: test_softmax")
x_arr = numpy.zeros((7, 37, 1, 1), dtype=numpy.float32)
x_arr[:] = numpy.random.rand(x_arr.size).reshape(x_arr.shape)
y_gold = x_arr.copy().reshape(x_arr.shape[0],
x_arr.size // x_arr.shape[0])
y_gold[:] = (y_gold.transpose() - y_gold.max(axis=1)).transpose()
numpy.exp(y_gold, y_gold)
y_gold[:] = (y_gold.transpose() / y_gold.sum(axis=1)).transpose()
x_desc = cudnn.TensorDescriptor()
x_desc.set_4d(cudnn.CUDNN_TENSOR_NCHW, cudnn.CUDNN_DATA_FLOAT,
*x_arr.shape)
x = cu.MemAlloc(self.ctx, x_arr)
y_arr = numpy.zeros_like(x_arr)
y = cu.MemAlloc(self.ctx, y_arr)
np_one = numpy.ones(1, dtype=numpy.float32)
np_zero = numpy.zeros(1, dtype=numpy.float32)
self.cudnn.softmax_forward(np_one, x_desc, x, np_zero, x_desc, y)
y.to_host(y_arr)
max_diff = numpy.fabs(y_arr.ravel() - y_gold.ravel()).max()
self.assertLess(max_diff, 1.0e-5)
# Backpropagtion test
dy_arr = numpy.zeros_like(y_gold)
dy_arr[:] = numpy.random.rand(dy_arr.size).reshape(dy_arr.shape) + 0.01
dy = cu.MemAlloc(self.ctx, dy_arr)
dx_arr = numpy.zeros_like(dy_arr)
dx = cu.MemAlloc(self.ctx, dx_arr)
self.cudnn.softmax_backward(np_one, x_desc, y, x_desc, dy, np_zero,
x_desc, dx)
dx.to_host(dx_arr)
self.assertEqual(numpy.count_nonzero(dx_arr), dx_arr.size)
# TODO(a.kazantsev): add test for gradient correctness.
logging.debug("EXIT: test_softmax")
def test_rnn(self):
if self.cudnn.version < 5000:
return
logging.debug("ENTER: test_rnn")
drop = cudnn.DropoutDescriptor()
drop_states = cu.MemAlloc(self.ctx, self.cudnn.dropout_states_size)
self.cudnn.set_dropout_descriptor(drop, 0.5, drop_states,
drop_states.size, 1234)
rnn = cudnn.RNNDescriptor()
self.assertEqual(rnn.hidden_size, 0)
self.assertEqual(rnn.num_layers, 0)
self.assertIsNone(rnn.dropout_desc)
self.assertEqual(rnn.input_mode, -1)
self.assertEqual(rnn.direction, -1)
self.assertEqual(rnn.mode, -1)
self.assertEqual(rnn.data_type, -1)
self.assertEqual(rnn.num_linear_layers, 0)
batch_size = 4
x = numpy.zeros(
(batch_size, 32), # minibatch, input size
dtype=numpy.float32)
numpy.random.seed(1234)
x[:] = numpy.random.rand(x.size).reshape(x.shape) - 0.5
x_desc = cudnn.TensorDescriptor()
# Set input as 3-dimensional like in cudnn example:
# minibatch, input_size, 1
x_desc.set_nd(cudnn.CUDNN_DATA_FLOAT, (x.shape[0], x.shape[1], 1))
n_unroll = 16
hidden_size = 64
n_layers = 3
def assert_values():
self.assertEqual(rnn.hidden_size, hidden_size)
self.assertEqual(rnn.num_layers, n_layers)
self.assertIs(rnn.dropout_desc, drop)
self.assertEqual(rnn.input_mode, cudnn.CUDNN_LINEAR_INPUT)
self.assertEqual(rnn.direction, cudnn.CUDNN_UNIDIRECTIONAL)
self.assertEqual(rnn.mode, cudnn.CUDNN_LSTM)
self.assertEqual(rnn.data_type, cudnn.CUDNN_DATA_FLOAT)
self.assertEqual(rnn.num_linear_layers, 8)
# Short syntax
rnn.set(hidden_size, n_layers, drop)
assert_values()
# Check num_linear_layers property
for mode, n in ((cudnn.CUDNN_RNN_RELU, 2), (cudnn.CUDNN_RNN_TANH, 2),
(cudnn.CUDNN_GRU, 6)):
rnn = cudnn.RNNDescriptor()
rnn.set(hidden_size, n_layers, drop, mode=mode)
self.assertEqual(rnn.num_linear_layers, n)
# Full syntax
rnn = cudnn.RNNDescriptor()
rnn.set(hidden_size,
n_layers,
drop,
input_mode=cudnn.CUDNN_LINEAR_INPUT,
direction=cudnn.CUDNN_UNIDIRECTIONAL,
mode=cudnn.CUDNN_LSTM,
data_type=cudnn.CUDNN_DATA_FLOAT)
assert_values()
def get_sz(func):
sz = func(rnn, (x_desc for _i in range(n_unroll)))
self.assertIsInstance(sz, int)
return sz
sz_work = get_sz(self.cudnn.get_rnn_workspace_size)
logging.debug("RNN workspace size for %s with %d unrolls is %d",
x.shape, n_unroll, sz_work)
sz_train = get_sz(self.cudnn.get_rnn_training_reserve_size)
logging.debug("RNN train size for %s with %d unrolls is %d", x.shape,
n_unroll, sz_train)
sz_params = self.cudnn.get_rnn_params_size(rnn, x_desc)
logging.debug("RNN params size for %s is %d", x.shape, sz_params)
x_desc2 = cudnn.TensorDescriptor()
x_desc2.set_nd(cudnn.CUDNN_DATA_DOUBLE, (x.shape[0], x.shape[1], 1))
sz_params2 = self.cudnn.get_rnn_params_size(rnn, x_desc2,
cudnn.CUDNN_DATA_DOUBLE)
self.assertEqual(sz_params2, sz_params * 2)
params_desc = cudnn.FilterDescriptor()
params_desc.set_nd(cudnn.CUDNN_DATA_FLOAT, (sz_params >> 2, 1, 1))
params = cu.MemAlloc(self.ctx, sz_params)
params.memset32_async()
w_desc = cudnn.FilterDescriptor()
w = self.cudnn.get_rnn_lin_layer_matrix_params(rnn, 0, x_desc,
params_desc, params, 0,
w_desc)
logging.debug("Got matrix 0 of dimensions: %s, fmt=%d, sz=%d",
w_desc.dims, w_desc.fmt, w.size)
self.assertEqual(w.size, hidden_size * x.shape[1] * 4)
b_desc = cudnn.FilterDescriptor()
b = self.cudnn.get_rnn_lin_layer_bias_params(rnn, 0, x_desc,
params_desc, params, 0,
b_desc)
logging.debug("Got bias 0 of dimensions: %s, fmt=%d, sz=%d",
b_desc.dims, b_desc.fmt, b.size)
self.assertEqual(b.size, hidden_size * 4)
workspace = cu.MemAlloc(self.ctx, sz_work)
x_buf = cu.MemAlloc(self.ctx, x.nbytes * n_unroll)
for i in range(n_unroll): # will feed the same input
x_buf.to_device(x, x.nbytes * i, x.nbytes)
y_buf = cu.MemAlloc(self.ctx, 4 * hidden_size * batch_size * n_unroll)
hx_buf = cu.MemAlloc(self.ctx, 4 * hidden_size * batch_size * n_layers)
hx_buf.memset32_async()
hy_buf = cu.MemAlloc(self.ctx, 4 * hidden_size * batch_size * n_layers)
cx_buf = cu.MemAlloc(self.ctx, 4 * hidden_size * batch_size * n_layers)
cx_buf.memset32_async()
cy_buf = cu.MemAlloc(self.ctx, 4 * hidden_size * batch_size * n_layers)
y_desc = cudnn.TensorDescriptor()
y_desc.set_nd(cudnn.CUDNN_DATA_FLOAT, (batch_size, hidden_size, 1))
h_desc = cudnn.TensorDescriptor()
h_desc.set_nd(cudnn.CUDNN_DATA_FLOAT,
(n_layers, batch_size, hidden_size))
self.ctx.synchronize()
logging.debug("Starting forward inference")
for i in range(5):
self.cudnn.rnn_forward_inference(
rnn, (x_desc for _i in range(n_unroll)), x_buf, h_desc, hx_buf,
h_desc, cx_buf, params_desc, params,
(y_desc for _i in range(n_unroll)), y_buf, h_desc, hy_buf,
h_desc, cy_buf, workspace, sz_work)
if i == 0:
self.ctx.synchronize()
t0 = time.time()
self.ctx.synchronize()
logging.debug("Forward inference done in %.6f sec",
(time.time() - t0) / 4)
train_space = cu.MemAlloc(self.ctx, sz_train)
self.cudnn.rnn_forward_training(
rnn, (x_desc for _i in range(n_unroll)), x_buf, h_desc, hx_buf,
h_desc, cx_buf, params_desc, params,
(y_desc for _i in range(n_unroll)), y_buf, h_desc, hy_buf, h_desc,
cy_buf, workspace, sz_work, train_space, sz_train)
self.ctx.synchronize()
logging.debug("Forward training done")
dy_buf = cu.MemAlloc(self.ctx, 4 * hidden_size * batch_size * n_unroll)
dy_buf.from_device_async(y_buf)
dhy_buf = cu.MemAlloc(self.ctx,
4 * hidden_size * batch_size * n_layers)
dhy_buf.memset32_async()
dcy_buf = cu.MemAlloc(self.ctx,
4 * hidden_size * batch_size * n_layers)
dcy_buf.memset32_async()
dx_buf = cu.MemAlloc(self.ctx, x_buf.size)
dhx_buf = cu.MemAlloc(self.ctx,
4 * hidden_size * batch_size * n_layers)
dcx_buf = cu.MemAlloc(self.ctx,
4 * hidden_size * batch_size * n_layers)
self.ctx.synchronize()
logging.debug("Starting backpropagation")
for i in range(5):
self.cudnn.rnn_backward_data(
rnn, (y_desc for _i in range(n_unroll)), y_buf,
(y_desc for _i in range(n_unroll)), dy_buf, h_desc, dhy_buf,
h_desc, dcy_buf, params_desc, params, h_desc, hx_buf, h_desc,
cx_buf, (x_desc
for _i in range(n_unroll)), dx_buf, h_desc, dhx_buf,
h_desc, dcx_buf, workspace, sz_work, train_space, sz_train)
if i == 0:
self.ctx.synchronize()
t0 = time.time()
self.ctx.synchronize()
logging.debug("Backpropagation done in %.6f sec",
(time.time() - t0) / 4)
dw = cu.MemAlloc(self.ctx, params.size)
logging.debug("Starting gradient computation")
for i in range(5):
self.cudnn.rnn_backward_weights(
rnn, (x_desc for _i in range(n_unroll)), x_buf, h_desc, hx_buf,
(y_desc for _i in range(n_unroll)), y_buf, workspace, sz_work,
params_desc, dw, train_space, sz_train)
if i == 0:
self.ctx.synchronize()
t0 = time.time()
self.ctx.synchronize()
logging.debug("Gradient computation done in %.6f sec",
(time.time() - t0) / 4)
logging.debug("EXIT: test_rnn")
def test_dropout(self):
if self.cudnn.version < 5000:
return
logging.debug("ENTER: test_dropout")
drop_ss = self.cudnn.dropout_states_size
self.assertIsInstance(drop_ss, int)
logging.debug("Dropout states size is %d", drop_ss)
input_data = numpy.zeros((5, 16, 32, 24), dtype=numpy.float32)
numpy.random.seed(1234)
input_data[:] = numpy.random.rand(input_data.size).reshape(
input_data.shape) - 0.5
input_desc = cudnn.TensorDescriptor()
input_desc.set_4d(cudnn.CUDNN_TENSOR_NCHW, cudnn.CUDNN_DATA_FLOAT,
*input_data.shape)
drop_rss = input_desc.dropout_reserve_space_size
self.assertIsInstance(drop_rss, int)
logging.debug("Dropout reserve space size for %s is %d",
input_data.shape, drop_rss)
drop = cudnn.DropoutDescriptor()
self.assertIsNone(self.cudnn.dropout_desc)
self.assertIsNone(self.cudnn.dropout_states)
self.cudnn.set_dropout_descriptor(drop) # with default parameters
self.assertIs(self.cudnn.dropout_desc, drop)
self.assertIsNone(self.cudnn.dropout_states)
states = cu.MemAlloc(self.ctx, drop_ss)
self.cudnn.set_dropout_descriptor(drop, 0.5, states, drop_ss, 1234)
self.assertIs(self.cudnn.dropout_desc, drop)
self.assertIs(self.cudnn.dropout_states, states)
input_buf = cu.MemAlloc(self.ctx, input_data)
output_buf = cu.MemAlloc(self.ctx, input_buf.size)
reserve = cu.MemAlloc(self.ctx, drop_rss)
self.cudnn.dropout_forward(drop, input_desc, input_buf, input_desc,
output_buf, reserve, reserve.size)
output_data = numpy.ones_like(input_data)
output_buf.to_host(output_data)
n_z = 0
for i, y in numpy.ndenumerate(output_data):
if not y:
n_z += 1
continue
x = input_data[i]
self.assertEqual(y, x * 2)
self.assertGreater(n_z,
(input_data.size >> 1) - (input_data.size >> 5))
err_data = numpy.ones_like(input_data)
err_data[:] = numpy.random.rand(output_data.size).reshape(
output_data.shape) - 0.5
err_buf = cu.MemAlloc(self.ctx, err_data)
self.cudnn.dropout_backward(drop, input_desc, err_buf, input_desc,
input_buf, reserve, reserve.size)
input_buf.to_host(input_data)
n_z = 0
for i, x in numpy.ndenumerate(input_data):
if not output_data[i]:
self.assertEqual(x, 0.0)
n_z += 1
continue
y = err_data[i]
if self.cudnn.version == 5004:
self.assertEqual(x, y) # strangely, it doesn't scale gradient
else:
self.assertEqual(x, y * 2)
self.assertGreater(n_z,
(input_data.size >> 1) - (input_data.size >> 5))
logging.debug("EXIT: test_dropout")
def test_lstm(self):
if self.cudnn.version < 5000:
return
logging.debug("ENTER: test_lstm")
drop = cudnn.DropoutDescriptor()
drop_states = cu.MemAlloc(self.ctx, self.cudnn.dropout_states_size)
self.cudnn.set_dropout_descriptor(drop, 0.0, drop_states,
drop_states.size, 1234)
rnn = cudnn.RNNDescriptor()
self.assertEqual(rnn.hidden_size, 0)
self.assertEqual(rnn.num_layers, 0)
self.assertIsNone(rnn.dropout_desc)
self.assertEqual(rnn.input_mode, -1)
self.assertEqual(rnn.direction, -1)
self.assertEqual(rnn.mode, -1)
self.assertEqual(rnn.data_type, -1)
self.assertEqual(rnn.num_linear_layers, 0)
batch_size = 8
x_arr = numpy.zeros(
(batch_size, 16), # minibatch, input size
dtype=DTYPE)
numpy.random.seed(1234)
x_arr[:] = numpy.random.rand(x_arr.size).reshape(x_arr.shape) - 0.5
x_desc = cudnn.TensorDescriptor()
# Set input as 3-dimensional like in cudnn example.
x_desc.set_nd(CUTYPE, (x_arr.shape[0], x_arr.shape[1], 1))
n_unroll = 5
hidden_size = 16
n_layers = 3
def assert_values():
self.assertEqual(rnn.hidden_size, hidden_size)
self.assertEqual(rnn.num_layers, n_layers)
self.assertIs(rnn.dropout_desc, drop)
self.assertEqual(rnn.input_mode, cudnn.CUDNN_LINEAR_INPUT)
self.assertEqual(rnn.direction, cudnn.CUDNN_UNIDIRECTIONAL)
self.assertEqual(rnn.mode, cudnn.CUDNN_LSTM)
self.assertEqual(rnn.data_type, CUTYPE)
self.assertEqual(rnn.num_linear_layers, 8)
# Full syntax
rnn = cudnn.RNNDescriptor()
rnn.set(hidden_size,
n_layers,
drop,
input_mode=cudnn.CUDNN_LINEAR_INPUT,
direction=cudnn.CUDNN_UNIDIRECTIONAL,
mode=cudnn.CUDNN_LSTM,
data_type=CUTYPE)
assert_values()
x_descs = tuple(x_desc for _i in range(n_unroll))
def get_sz(func):
sz = func(rnn, x_descs)
self.assertIsInstance(sz, int)
return sz
sz_work = get_sz(self.cudnn.get_rnn_workspace_size)
logging.debug("RNN workspace size for %s with %d unrolls is %d",
x_arr.shape, n_unroll, sz_work)
sz_train = get_sz(self.cudnn.get_rnn_training_reserve_size)
logging.debug("RNN train size for %s with %d unrolls is %d",
x_arr.shape, n_unroll, sz_train)
sz_weights = self.cudnn.get_rnn_params_size(rnn, x_desc)
logging.debug("RNN weights size for %s is %d", x_arr.shape, sz_weights)
sz_expected = ITEMSIZE * (
4 * (x_arr.shape[1] + hidden_size + 2) * hidden_size + 4 *
(hidden_size + hidden_size + 2) * hidden_size * (n_layers - 1))
self.assertEqual(sz_weights, sz_expected)
weights_desc = cudnn.FilterDescriptor()
weights_desc.set_nd(CUTYPE, (sz_weights // ITEMSIZE, 1, 1))
weights = cu.MemAlloc(self.ctx, sz_weights)
weights_arr = numpy.random.rand(sz_weights // ITEMSIZE).astype(DTYPE)
weights_arr -= 0.5
weights_arr *= 0.1
weights.to_device(weights_arr)
w_desc = cudnn.FilterDescriptor()
w = self.cudnn.get_rnn_lin_layer_matrix_params(rnn, 0, x_desc,
weights_desc, weights,
0, w_desc)
logging.debug("Got matrix 0 of dimensions: %s, fmt=%d, sz=%d",
w_desc.dims, w_desc.fmt, w.size)
self.assertEqual(w.size, hidden_size * x_arr.shape[1] * ITEMSIZE)
b_desc = cudnn.FilterDescriptor()
b = self.cudnn.get_rnn_lin_layer_bias_params(rnn, 0, x_desc,
weights_desc, weights, 0,
b_desc)
logging.debug("Got bias 0 of dimensions: %s, fmt=%d, sz=%d",
b_desc.dims, b_desc.fmt, b.size)
self.assertEqual(b.size, hidden_size * ITEMSIZE)
work_buf = cu.MemAlloc(self.ctx, sz_work)
work_buf.memset32_async()
x = cu.MemAlloc(self.ctx, x_arr.nbytes * n_unroll)
for i in range(n_unroll): # will feed the same input
x.to_device(x_arr, x_arr.nbytes * i, x_arr.nbytes)
y_arr = numpy.zeros((n_unroll, batch_size, hidden_size), dtype=DTYPE)
y = cu.MemAlloc(self.ctx, y_arr)
hx_arr = numpy.zeros((n_layers, batch_size, hidden_size), dtype=DTYPE)
hx_arr[:] = numpy.random.rand(hx_arr.size).reshape(hx_arr.shape)
hx_arr -= 0.5
hx = cu.MemAlloc(self.ctx, hx_arr)
hy = cu.MemAlloc(self.ctx, hx.size)
hy.memset32_async()
cx_arr = numpy.zeros((n_layers, batch_size, hidden_size), dtype=DTYPE)
cx_arr[:] = numpy.random.rand(cx_arr.size).reshape(cx_arr.shape)
cx_arr -= 0.5
cx = cu.MemAlloc(self.ctx, cx_arr)
cy = cu.MemAlloc(self.ctx, cx.size)
cy.memset32_async()
y_desc = cudnn.TensorDescriptor()
y_desc.set_nd(CUTYPE, (batch_size, hidden_size, 1))
y_descs = tuple(y_desc for _i in range(n_unroll))
h_desc = cudnn.TensorDescriptor()
h_desc.set_nd(CUTYPE, (n_layers, batch_size, hidden_size))
train_buf = cu.MemAlloc(self.ctx, sz_train)
train_buf.memset32_async()
self.cudnn.rnn_forward_training(rnn, x_descs, x, h_desc, hx, h_desc,
cx, weights_desc, weights, y_descs, y,
h_desc, hy, h_desc, cy, work_buf,
sz_work, train_buf, sz_train)
self.ctx.synchronize()
logging.debug("Forward training done")
y.to_host(y_arr)
target = numpy.random.rand(y_arr.size).reshape(y_arr.shape).astype(
y_arr.dtype) - 0.5
dy_arr = y_arr - target
dy = cu.MemAlloc(self.ctx, dy_arr)
dhy = cu.MemAlloc(self.ctx, hx.size)
dhy.memset32_async()
dcy = cu.MemAlloc(self.ctx, cx.size)
dcy.memset32_async()
dx_arr = numpy.zeros_like(x_arr)
dx = cu.MemAlloc(self.ctx, dx_arr)
dhx_arr = numpy.zeros_like(hx_arr)
dhx = cu.MemAlloc(self.ctx, dhx_arr)
dcx_arr = numpy.zeros_like(cx_arr)
dcx = cu.MemAlloc(self.ctx, dcx_arr)
self.cudnn.rnn_backward_data(rnn, y_descs, y, y_descs, dy, h_desc, dhy,
h_desc, dcy, weights_desc, weights,
h_desc, hx, h_desc, cx, x_descs, dx,
h_desc, dhx, h_desc, dcx, work_buf,
sz_work, train_buf, sz_train)
logging.debug("Backpropagation done")
dx.to_host(dx_arr)
dhx.to_host(dhx_arr)
dcx.to_host(dcx_arr)
def forward():
x.to_device_async(x_arr)
hx.to_device_async(hx_arr)
cx.to_device_async(cx_arr)
self.cudnn.rnn_forward_inference(rnn, x_descs, x, h_desc, hx,
h_desc, cx, weights_desc, weights,
y_descs, y, h_desc, hy, h_desc,
cy, work_buf, sz_work)
y.to_host(y_arr)
numdiff = NumDiff()
logging.debug("Checking dx...")
numdiff.check_diff(x_arr, y_arr, target, dx_arr, forward)
logging.debug("Checking dhx...")
numdiff.check_diff(hx_arr, y_arr, target, dhx_arr, forward)
logging.debug("Checking dcx...")
numdiff.check_diff(cx_arr, y_arr, target, dcx_arr, forward)
logging.debug("EXIT: test_lstm")
