Merge pull request #4768 from abergeron/gpua_bn

theano/gpuarray/dnn_batchnorm.c

+#section support_code_struct
+
+int dnn_batchnorm_op(PyGpuArrayObject *inp, PyGpuArrayObject *scale,
+                     PyGpuArrayObject *bias, PyGpuArrayObject **outp,
+                     PyGpuArrayObject **x_mean, PyGpuArrayObject **x_invstd,
+                     PyGpuContextObject *c) {
+  if (c_set_tensorNd(inp, bn_input) != 0)
+    return 1;
+  if (c_set_tensorNd(scale, bn_params) != 0)
+    return 1;
+
+  if (theano_prep_output(outp, inp->ga.nd, inp->ga.dimensions, inp->ga.typecode, GA_C_ORDER, c) != 0)
+    return 1;
+  if (theano_prep_output(x_mean, scale->ga.nd, scale->ga.dimensions, scale->ga.typecode, GA_C_ORDER, c) != 0)
+    return 1;
+  if (theano_prep_output(x_invstd, scale->ga.nd, scale->ga.dimensions, scale->ga.typecode, GA_C_ORDER, c) != 0)
+    return 1;
+
+  if (c_set_tensorNd(*outp, bn_output) != 0)
+    return 1;
+
+  {
+    const float falpha = 1.;
+    const float fbeta = 0.;
+    const double dalpha = 1.;
+    const double dbeta = 0.;
+    void *alpha;
+    void *beta;
+    if (inp->ga.typecode == GA_DOUBLE) {
+      alpha = (void *)&dalpha;
+      beta = (void *)&dbeta;
+    } else {
+      alpha = (void *)&falpha;
+      beta = (void *)&fbeta;
+    }
+    cudnnStatus_t err = cudnnBatchNormalizationForwardTraining(
+      APPLY_SPECIFIC(_handle),
+      MODE,
+      alpha,
+      beta,
+      bn_input,
+      PyGpuArray_DEV_DATA(inp),
+      bn_output,
+      PyGpuArray_DEV_DATA(*outp),
+      bn_params,
+      PyGpuArray_DEV_DATA(scale),
+      PyGpuArray_DEV_DATA(bias),
+      0,
+      NULL,  // running mean, deliberately unused
+      NULL,  // running var, deliberately unused
+      EPSILON,
+      PyGpuArray_DEV_DATA(*x_mean),
+      PyGpuArray_DEV_DATA(*x_invstd)
+      );
+    if (err != CUDNN_STATUS_SUCCESS) {
+      PyErr_Format(PyExc_RuntimeError, "Error during batchnorm: %s\n",
+                   cudnnGetErrorString(err));
+      return 1;
+    }
+  }
+  return 0;
+}

theano/gpuarray/dnn_batchnorm_base.c

+#section init_code_struct
+
+{
+  cudnnStatus_t err;
+
+  bn_input = NULL;
+  bn_params = NULL;
+  bn_output = NULL;
+
+  if ((err = cudnnCreateTensorDescriptor(&bn_input)) != CUDNN_STATUS_SUCCESS) {
+    PyErr_Format(PyExc_MemoryError, "could not allocate tensor descriptor "
+                 "(bn_input): %s", cudnnGetErrorString(err));
+    FAIL;
+  }
+  if ((err = cudnnCreateTensorDescriptor(&bn_params)) != CUDNN_STATUS_SUCCESS) {
+    PyErr_Format(PyExc_MemoryError, "could not allocate tensor descriptor "
+                 "(bn_params): %s", cudnnGetErrorString(err));
+    FAIL;
+  }
+  if ((err = cudnnCreateTensorDescriptor(&bn_output)) != CUDNN_STATUS_SUCCESS) {
+    PyErr_Format(PyExc_MemoryError, "could not allocate tensor descriptor "
+                 "(bn_output): %s", cudnnGetErrorString(err));
+    FAIL;
+  }
+}
+
+#section cleanup_code_struct
+
+if (bn_input != NULL)
+  cudnnDestroyTensorDescriptor(bn_input);
+if (bn_params != NULL)
+  cudnnDestroyTensorDescriptor(bn_params);
+if (bn_output != NULL)
+  cudnnDestroyTensorDescriptor(bn_output);
+
+#section support_code_struct
+
+cudnnTensorDescriptor_t bn_input;
+cudnnTensorDescriptor_t bn_params;
+cudnnTensorDescriptor_t bn_output;

theano/gpuarray/dnn_batchnorm_grad.c

+#section init_code_struct
+
+{
+  cudnnStatus_t err;
+
+  bn_doutput = NULL;
+  if ((err = cudnnCreateTensorDescriptor(&bn_doutput)) != CUDNN_STATUS_SUCCESS) {
+    PyErr_Format(PyExc_MemoryError, "could not allocate tensor descriptor "
+                 "(bn_doutput): %s", cudnnGetErrorString(err));
+    FAIL;
+  }
+}
+
+#section cleanup_code_struct
+
+if (bn_doutput != NULL)
+  cudnnDestroyTensorDescriptor(bn_doutput);
+
+#section support_code_struct
+
+cudnnTensorDescriptor_t bn_doutput;
+
+int dnn_batchnorm_grad(PyGpuArrayObject *inp, PyGpuArrayObject *doutp,
+                       PyGpuArrayObject *scale, PyGpuArrayObject *x_mean,
+                       PyGpuArrayObject *x_invstd, PyGpuArrayObject **dinp,
+                       PyGpuArrayObject **dscale, PyGpuArrayObject **dbias,
+                       PyGpuContextObject *c) {
+  if (c_set_tensorNd(inp, bn_input) != 0)
+    return 1;
+  if (c_set_tensorNd(doutp, bn_doutput) != 0)
+    return 1;
+  if (c_set_tensorNd(scale, bn_params) != 0)
+    return 1;
+
+  if (theano_prep_output(dinp, inp->ga.nd, inp->ga.dimensions, inp->ga.typecode, GA_C_ORDER, c) != 0)
+    return 1;
+  if (theano_prep_output(dscale, scale->ga.nd, scale->ga.dimensions, scale->ga.typecode, GA_C_ORDER, c) != 0)
+    return 1;
+  if (theano_prep_output(dbias, scale->ga.nd, scale->ga.dimensions, scale->ga.typecode, GA_C_ORDER, c) != 0)
+    return 1;
+
+  if (c_set_tensorNd(*dinp, bn_output) != 0)
+    return 1;
+
+  {
+    const float falpha = 1.;
+    const float fbeta = 0.;
+    const double dalpha = 1.;
+    const double dbeta = 0.;
+    void *alphaData;
+    void *betaData;
+    void *alphaParam;
+    void *betaParam;
+    if (inp->ga.typecode == GA_DOUBLE) {
+      alphaData = (void *)&dalpha;
+      betaData = (void *)&dbeta;
+      alphaParam = (void *)&dalpha;
+      betaParam = (void *)&dbeta;
+    } else {
+      alphaData = (void *)&falpha;
+      betaData = (void *)&fbeta;
+      alphaParam = (void *)&falpha;
+      betaParam = (void *)&fbeta;
+    }
+    cudnnStatus_t err = cudnnBatchNormalizationBackward(
+      APPLY_SPECIFIC(_handle),
+      MODE,
+      alphaData,
+      betaData,
+      alphaParam,
+      betaParam,
+      bn_input,
+      PyGpuArray_DEV_DATA(inp),
+      bn_doutput,
+      PyGpuArray_DEV_DATA(doutp),
+      bn_output,
+      PyGpuArray_DEV_DATA(*dinp),
+      bn_params,
+      PyGpuArray_DEV_DATA(scale),
+      PyGpuArray_DEV_DATA(*dscale),
+      PyGpuArray_DEV_DATA(*dbias),
+      EPSILON,
+      PyGpuArray_DEV_DATA(x_mean),
+      PyGpuArray_DEV_DATA(x_invstd)
+      );
+    if (err != CUDNN_STATUS_SUCCESS) {
+      PyErr_Format(PyExc_RuntimeError, "Error during batchnorm: %s\n",
+                   cudnnGetErrorString(err));
+      return 1;
+    }
+  }
+  return 0;
+}

theano/gpuarray/dnn_batchnorm_inf.c

+#section support_code_struct
+
+int dnn_batchnorm_op(PyGpuArrayObject *inp, PyGpuArrayObject *scale,
+                     PyGpuArrayObject *bias, PyGpuArrayObject *est_mean,
+                     PyGpuArrayObject *est_var, PyGpuArrayObject **outp,
+                     PyGpuContextObject *c) {
+  if (c_set_tensorNd(inp, bn_input) != 0)
+    return 1;
+  if (c_set_tensorNd(scale, bn_params) != 0)
+    return 1;
+
+  if (theano_prep_output(outp, inp->ga.nd, inp->ga.dimensions, inp->ga.typecode, GA_C_ORDER, c) != 0)
+    return 1;
+
+  if (c_set_tensorNd(*outp, bn_output) != 0)
+    return 1;
+
+  {
+    const float falpha = 1.;
+    const float fbeta = 0.;
+    const double dalpha = 1.;
+    const double dbeta = 0.;
+    void *alpha;
+    void *beta;
+    if (inp->ga.typecode == GA_DOUBLE) {
+      alpha = (void *)&dalpha;
+      beta = (void *)&dbeta;
+    } else {
+      alpha = (void *)&falpha;
+      beta = (void *)&fbeta;
+    }
+    cudnnStatus_t err = cudnnBatchNormalizationForwardInference(
+      APPLY_SPECIFIC(_handle),
+      MODE,
+      alpha,
+      beta,
+      bn_input,
+      PyGpuArray_DEV_DATA(inp),
+      bn_output,
+      PyGpuArray_DEV_DATA(*outp),
+      bn_params,
+      PyGpuArray_DEV_DATA(scale),
+      PyGpuArray_DEV_DATA(bias),
+      PyGpuArray_DEV_DATA(est_mean),
+      PyGpuArray_DEV_DATA(est_var),
+      EPSILON
+      );
+    if (err != CUDNN_STATUS_SUCCESS) {
+      PyErr_Format(PyExc_RuntimeError, "Error during batchnorm: %s\n",
+                   cudnnGetErrorString(err));
+      return 1;
+    }
+  }
+  return 0;
+}

theano/gpuarray/tests/test_dnn.py

@@ -973,3 +973,114 @@ class test_SoftMax(test_nnet.test_SoftMax):
         # Compare the output of the function with the reference function
         inp = numpy.random.normal(0, 1, (5, 6)).astype("float32")
         utt.assert_allclose(f(inp), f_ref(inp))
+
+
+def test_dnn_batchnorm_train():
+    if not dnn.dnn_available(test_ctx_name):
+        raise SkipTest(dnn.dnn_available.msg)
+    if dnn.version(raises=False) < 5000:
+        raise SkipTest("batch normalization requires cudnn v5+")
+    utt.seed_rng()
+
+    for mode in ('per-activation', 'spatial'):
+        for vartype in (T.ftensor4, T.ftensor3, T.fmatrix, T.fvector):
+            x, scale, bias = (vartype(n) for n in ('x', 'scale', 'bias'))
+            ndim = x.ndim
+            eps = 5e-3  # some non-standard value to test if it's used
+
+            # forward pass
+            out, x_mean, x_invstd = dnn.dnn_batch_normalization_train(
+                x, scale, bias, mode, eps)
+            # reference forward pass
+            if mode == 'per-activation':
+                axes = (0,)
+            elif mode == 'spatial':
+                axes = (0,) + tuple(range(2, ndim))
+            x_mean2 = x.mean(axis=axes, keepdims=True)
+            x_invstd2 = T.inv(T.sqrt(x.var(axis=axes, keepdims=True) + eps))
+            scale2 = T.addbroadcast(scale, *axes)
+            bias2 = T.addbroadcast(bias, *axes)
+            out2 = (x - x_mean2) * (scale2 * x_invstd2) + bias2
+            # backward pass
+            dy = vartype('dy')
+            grads = T.grad(None, wrt=[x, scale, bias], known_grads={out: dy})
+            # reference backward pass
+            grads2 = T.grad(None, wrt=[x, scale, bias], known_grads={out2: dy})
+            # compile
+            f = theano.function([x, scale, bias, dy],
+                                [out, x_mean, x_invstd, out2, x_mean2, x_invstd2] +
+                                grads + grads2, mode=mode_with_gpu)
+            # run
+            for data_shape in ((10, 20, 30, 40), (4, 3, 1, 1), (1, 1, 5, 5)):
+                data_shape = data_shape[:ndim]
+                param_shape = tuple(1 if d in axes else s
+                                    for d, s in enumerate(data_shape))
+                X = 4 + 3 * numpy.random.randn(*data_shape).astype('float32')
+                Dy = -1 + 2 * numpy.random.randn(*data_shape).astype('float32')
+                Scale = numpy.random.randn(*param_shape).astype('float32')
+                Bias = numpy.random.randn(*param_shape).astype('float32')
+                outputs = f(X, Scale, Bias, Dy)
+                # compare outputs
+                utt.assert_allclose(outputs[0], outputs[0 + 3])  # out
+                utt.assert_allclose(outputs[1], outputs[1 + 3])  # mean
+                utt.assert_allclose(outputs[2], outputs[2 + 3])  # invstd
+                # compare gradients
+                utt.assert_allclose(outputs[6], outputs[6 + 3])  # dx
+                utt.assert_allclose(outputs[7], outputs[7 + 3], rtol=3e-3)  # dscale
+                utt.assert_allclose(outputs[8], outputs[8 + 3])  # dbias
+
+
+def test_batchnorm_inference():
+    if not dnn.dnn_available(test_ctx_name):
+        raise SkipTest(dnn.dnn_available.msg)
+    if dnn.version(raises=False) < 5000:
+        raise SkipTest("batch normalization requires cudnn v5+")
+    utt.seed_rng()
+
+    for mode in ('per-activation', 'spatial'):
+        for vartype in (T.ftensor4, T.ftensor3, T.fmatrix, T.fvector):
+            x, scale, bias, mean, var = (vartype(n) for n in ('x', 'scale',
+                                                              'bias', 'mean',
+                                                              'var'))
+            ndim = x.ndim
+            eps = 5e-3  # some non-standard value to test if it's used
+
+            # forward pass
+            out = dnn.dnn_batch_normalization_test(x, scale, bias, mean,
+                                                   var, mode, eps)
+            # reference forward pass
+            if mode == 'per-activation':
+                axes = (0,)
+            elif mode == 'spatial':
+                axes = (0,) + tuple(range(2, ndim))
+            scale2, bias2, mean2, var2 = (T.addbroadcast(t, *axes)
+                                          for t in (scale, bias, mean, var))
+            out2 = (x - mean2) * (scale2 / T.sqrt(var2 + eps)) + bias2
+            # backward pass
+            dy = vartype('dy')
+            grads = T.grad(None, wrt=[x, scale, bias, mean, var], known_grads={out: dy})
+            # reference backward pass
+            grads2 = T.grad(None, wrt=[x, scale, bias, mean, var], known_grads={out2: dy})
+            # compile
+            f = theano.function([x, scale, bias, mean, var, dy],
+                                [out, out2] + grads + grads2, mode=mode_with_gpu)
+            # run
+            for data_shape in ((10, 20, 30, 40), (4, 3, 1, 1), (1, 1, 5, 5)):
+                data_shape = data_shape[:ndim]
+                param_shape = tuple(1 if d in axes else s
+                                    for d, s in enumerate(data_shape))
+                X = 4 + 3 * numpy.random.randn(*data_shape).astype('float32')
+                Dy = -1 + 2 * numpy.random.randn(*data_shape).astype('float32')
+                Scale = numpy.random.randn(*param_shape).astype('float32')
+                Bias = numpy.random.randn(*param_shape).astype('float32')
+                Mean = numpy.random.randn(*param_shape).astype('float32')
+                Var = numpy.random.rand(*param_shape).astype('float32')
+                outputs = f(X, Scale, Bias, Mean, Var, Dy)
+                # compare outputs
+                utt.assert_allclose(outputs[0], outputs[1])  # out
+                # compare gradients
+                utt.assert_allclose(outputs[2], outputs[2 + 5])  # dx
+                utt.assert_allclose(outputs[3], outputs[3 + 5])  # dscale
+                utt.assert_allclose(outputs[4], outputs[4 + 5])  # dbias
+                utt.assert_allclose(outputs[5], outputs[5 + 5])  # dmean
+                utt.assert_allclose(outputs[6], outputs[6 + 5], atol=2e-5)  # dvar