Changes to mxnet.metric

benchmark/python/sparse/sparse_end2end.py

@@ -225,7 +225,7 @@ def row_sparse_pull(kv, key, data, slices, weight_array, priority):
                            learning_rate=0.1, rescale_grad=1.0/batch_size/num_worker)
     mod.init_optimizer(optimizer=sgd, kvstore=kv)
     # use accuracy as the metric
-    metric = mx.metric.create('acc')
+    metric = mx.gluon.metric.create('acc')
 
     index = mod._exec_group.param_names.index('w')
     # weight_array bound to executors of the contexts

ci/docker/runtime_functions.sh

@@ -723,7 +723,6 @@ build_ubuntu_gpu_mkldnn() {
     CC=gcc-7 CXX=g++-7 cmake \
         -DCMAKE_BUILD_TYPE="RelWithDebInfo" \
         -DUSE_MKL_IF_AVAILABLE=OFF \
-        -DUSE_TVM_OP=ON \
         -DUSE_CUDA=ON \
         -DMXNET_CUDA_ARCH="$CI_CMAKE_CUDA_ARCH" \
         -DUSE_CPP_PACKAGE=ON \
@@ -737,7 +736,6 @@ build_ubuntu_gpu_mkldnn_nocudnn() {
     CC=gcc-7 CXX=g++-7 cmake \
         -DCMAKE_BUILD_TYPE="RelWithDebInfo" \
         -DUSE_MKL_IF_AVAILABLE=OFF \
-        -DUSE_TVM_OP=ON \
         -DUSE_CUDA=ON \
         -DMXNET_CUDA_ARCH="$CI_CMAKE_CUDA_ARCH" \
         -DUSE_CUDNN=OFF \
@@ -752,7 +750,6 @@ build_ubuntu_gpu_cuda101_cudnn7() {
     CC=gcc-7 CXX=g++-7 cmake \
         -DCMAKE_BUILD_TYPE="RelWithDebInfo" \
         -DUSE_MKL_IF_AVAILABLE=OFF \
-        -DUSE_TVM_OP=ON \
         -DUSE_CUDA=ON \
         -DMXNET_CUDA_ARCH="$CI_CMAKE_CUDA_ARCH" \
         -DUSE_CUDNN=ON \
@@ -775,7 +772,6 @@ build_ubuntu_gpu_cuda101_cudnn7_make() {
         USE_CUDA=1                                \
         USE_CUDA_PATH=/usr/local/cuda             \
         USE_CUDNN=1                               \
-        USE_TVM_OP=1                              \
         USE_CPP_PACKAGE=1                         \
         USE_DIST_KVSTORE=1                        \
         CUDA_ARCH="$CI_CUDA_COMPUTE_CAPABILITIES" \
@@ -795,7 +791,6 @@ build_ubuntu_gpu_cuda101_cudnn7_mkldnn_cpp_test() {
         USE_CUDA=1                                \
         USE_CUDA_PATH=/usr/local/cuda             \
         USE_CUDNN=1                               \
-        USE_TVM_OP=0                              \
         USE_CPP_PACKAGE=1                         \
         USE_DIST_KVSTORE=1                        \
         CUDA_ARCH="$CI_CUDA_COMPUTE_CAPABILITIES" \
@@ -805,23 +800,6 @@ build_ubuntu_gpu_cuda101_cudnn7_mkldnn_cpp_test() {
     make cython PYTHON=python3
 }
 
-build_ubuntu_gpu_cuda101_cudnn7_no_tvm_op() {
-    set -ex
-    cd /work/build
-    CC=gcc-7 CXX=g++-7 cmake \
-        -DCMAKE_BUILD_TYPE="RelWithDebInfo" \
-        -DUSE_MKL_IF_AVAILABLE=OFF \
-        -DUSE_TVM_OP=OFF \
-        -DUSE_CUDA=ON \
-        -DMXNET_CUDA_ARCH="$CI_CMAKE_CUDA_ARCH" \
-        -DUSE_CUDNN=ON \
-        -DUSE_MKLDNN=OFF \
-        -DBUILD_CYTHON_MODULES=ON \
-        -DUSE_DIST_KVSTORE=ON \
-        -G Ninja /work/mxnet
-    ninja
-}
-
 build_ubuntu_amalgamation() {
     set -ex
     # Amalgamation can not be run with -j nproc
@@ -852,7 +830,6 @@ build_ubuntu_gpu_cmake() {
         -DUSE_SIGNAL_HANDLER=ON                 \
         -DUSE_CUDA=ON                           \
         -DUSE_CUDNN=ON                          \
-        -DUSE_TVM_OP=ON                         \
         -DUSE_MKL_IF_AVAILABLE=OFF              \
         -DUSE_MKLML_MKL=OFF                     \
         -DUSE_MKLDNN=OFF                        \
@@ -873,7 +850,6 @@ build_ubuntu_gpu_cmake_no_rtc() {
         -DUSE_SIGNAL_HANDLER=ON                 \
         -DUSE_CUDA=ON                           \
         -DUSE_CUDNN=ON                          \
-        -DUSE_TVM_OP=ON                         \
         -DUSE_MKL_IF_AVAILABLE=OFF              \
         -DUSE_MKLML_MKL=OFF                     \
         -DUSE_MKLDNN=ON                         \
@@ -888,27 +864,6 @@ build_ubuntu_gpu_cmake_no_rtc() {
     ninja
 }
 
-build_ubuntu_gpu_cmake_no_tvm_op() {
-    set -ex
-    cd /work/build
-    CC=gcc-7 CXX=g++-7 cmake \
-        -DUSE_SIGNAL_HANDLER=ON                 \
-        -DUSE_CUDA=ON                           \
-        -DUSE_CUDNN=ON                          \
-        -DUSE_TVM_OP=OFF                        \
-        -DUSE_MKL_IF_AVAILABLE=OFF              \
-        -DUSE_MKLML_MKL=OFF                     \
-        -DUSE_MKLDNN=OFF                        \
-        -DUSE_DIST_KVSTORE=ON                   \
-        -DCMAKE_BUILD_TYPE=Release              \
-        -DMXNET_CUDA_ARCH="$CI_CMAKE_CUDA_ARCH" \
-        -DBUILD_CYTHON_MODULES=1                \
-        -G Ninja                                \
-        /work/mxnet
-
-    ninja
-}
-
 build_ubuntu_cpu_large_tensor() {
     set -ex
     cd /work/build
@@ -931,7 +886,6 @@ build_ubuntu_gpu_large_tensor() {
         -DUSE_SIGNAL_HANDLER=ON                 \
         -DUSE_CUDA=ON                           \
         -DUSE_CUDNN=ON                          \
-        -DUSE_TVM_OP=ON                         \
         -DUSE_MKL_IF_AVAILABLE=OFF              \
         -DUSE_MKLML_MKL=OFF                     \
         -DUSE_MKLDNN=OFF                        \
@@ -989,7 +943,8 @@ cd_unittest_ubuntu() {
 
         # Adding these here as CI doesn't test all CUDA environments
         pytest example/image-classification/test_score.py
-        integrationtest_ubuntu_gpu_dist_kvstore
+        # TODO(szha): fix and reenable the hanging issue. tracked in #18098
+        # integrationtest_ubuntu_gpu_dist_kvstore
     fi
 
     if [[ ${mxnet_variant} = *mkl ]]; then

ci/jenkins/Jenkins_steps.groovy

@@ -289,20 +289,6 @@ def compile_unix_full_gpu_mkldnn_cpp_test() {
     }]
 }
 
-def compile_unix_full_gpu_no_tvm_op() {
-    return ['GPU: CUDA10.1+cuDNN7 TVM_OP OFF': {
-      node(NODE_LINUX_CPU) {
-        ws('workspace/build-gpu-no-tvm-op') {
-          timeout(time: max_time, unit: 'MINUTES') {
-            utils.init_git()
-            utils.docker_run('ubuntu_build_cuda', 'build_ubuntu_gpu_cuda101_cudnn7_no_tvm_op', false)
-            utils.pack_lib('gpu_no_tvm_op', mx_lib_cpp_examples_no_tvm_op)
-          }
-        }
-      }
-    }]
-}
-
 def compile_unix_cmake_gpu() {
     return ['GPU: CMake': {
       node(NODE_LINUX_CPU) {
@@ -317,19 +303,6 @@ def compile_unix_cmake_gpu() {
     }]
 }
 
-def compile_unix_cmake_gpu_no_tvm_op() {
-    return ['GPU: CMake TVM_OP OFF': {
-      node(NODE_LINUX_CPU) {
-        ws('workspace/build-cmake-gpu-no-tvm-op') {
-          timeout(time: max_time, unit: 'MINUTES') {
-            utils.init_git()
-            utils.docker_run('ubuntu_gpu_cu101', 'build_ubuntu_gpu_cmake_no_tvm_op', false)
-          }
-        }
-      }
-    }]
-}
-
 def compile_unix_cmake_gpu_no_rtc() {
     return ['GPU: CMake CUDA RTC OFF': {
         node(NODE_LINUX_CPU) {

ci/jenkins/Jenkinsfile_unix_gpu

@@ -41,8 +41,6 @@ core_logic: {
     custom_steps.compile_unix_cmake_gpu(),
     custom_steps.compile_unix_tensorrt_gpu(),
     custom_steps.compile_unix_int64_gpu(),
-    custom_steps.compile_unix_full_gpu_no_tvm_op(),
-    custom_steps.compile_unix_cmake_gpu_no_tvm_op(),
     custom_steps.compile_unix_cmake_gpu_no_rtc(),
     custom_steps.compile_unix_full_gpu_mkldnn_cpp_test()
   ])

example/adversary/adversary_generation.ipynb

@@ -168,7 +168,7 @@
     "epoch = 3\n",
     "for e in range(epoch):\n",
     "    train_loss = 0.\n",
-    "    acc = mx.metric.Accuracy()\n",
+    "    acc = mx.gluon.metric.Accuracy()\n",
     "    for i, (data, label) in enumerate(train_data):\n",
     "        data = data.as_in_context(ctx)\n",
     "        label = label.as_in_context(ctx)\n",
@@ -223,7 +223,7 @@
     "    l = loss(output, label)\n",
     "l.backward()\n",
     "\n",
-    "acc = mx.metric.Accuracy()\n",
+    "acc = mx.gluon.metric.Accuracy()\n",
     "acc.update(label, output)\n",
     "\n",
     "print(\"Validation batch accuracy {}\".format(acc.get()[1]))"
@@ -256,7 +256,7 @@
     "\n",
     "output = net(data_perturbated)    \n",
     "\n",
-    "acc = mx.metric.Accuracy()\n",
+    "acc = mx.gluon.metric.Accuracy()\n",
     "acc.update(label, output)\n",
     "\n",
     "print(\"Validation batch accuracy after perturbation {}\".format(acc.get()[1]))"

example/autoencoder/variational_autoencoder/VAE_example.ipynb

@@ -610,7 +610,7 @@
    ],
    "source": [
     "# calculate the ELBO which is minus the loss for test set\n",
-    "metric = mx.metric.Loss()\n",
+    "metric = mx.gluon.metric.Loss()\n",
     "model.score(nd_iter_test, metric)"
    ]
   },

example/caffe/caffe_net.py

@@ -140,6 +140,6 @@ def parse_args():
 
     # train
     if use_caffe_loss:
-        train_model.fit(args, net, get_iterator(data_shape, use_caffe_data), mx.metric.Caffe())
+        train_model.fit(args, net, get_iterator(data_shape, use_caffe_data), mx.gluon.metric.Caffe())
     else:
         train_model.fit(args, net, get_iterator(data_shape, use_caffe_data))

example/caffe/train_model.py

@@ -93,7 +93,7 @@ def fit(args, network, data_loader, eval_metrics=None, batch_end_callback=None):
         eval_metrics = ['accuracy']
         # TopKAccuracy only allows top_k > 1
         for top_k in [5, 10, 20]:
-            eval_metrics.append(mx.metric.create('top_k_accuracy', top_k=top_k))
+            eval_metrics.append(mx.gluon.metric.create('top_k_accuracy', top_k=top_k))
 
     if batch_end_callback is not None:
         if not isinstance(batch_end_callback, list):

example/capsnet/capsulenet.py

@@ -122,7 +122,7 @@ def to4d(img):
     return img.reshape(img.shape[0], 1, 28, 28).astype(np.float32)/255
 
 
-class LossMetric(mx.metric.EvalMetric):
+class LossMetric(mx.gluon.metric.EvalMetric):
     """Evaluate the loss function"""
     def __init__(self, batch_size, num_gpus):
         super(LossMetric, self).__init__('LossMetric')

example/ctc/lstm_ocr_train.py

@@ -103,7 +103,7 @@ def main():
         module.fit(train_data=data_train,
                    eval_data=data_val,
                    # use metrics.accuracy or metrics.accuracy_lcs
-                   eval_metric=mx.metric.np(metrics.accuracy, allow_extra_outputs=True),
+                   eval_metric=mx.gluon.metric.np(metrics.accuracy, allow_extra_outputs=True),
                    optimizer='sgd',
                    optimizer_params={'learning_rate': hp.learning_rate,
                                      'momentum': hp.momentum,

example/deep-embedded-clustering/autoencoder.py

@@ -165,7 +165,7 @@ def l2_norm(label, pred):
             return np.mean(np.square(label-pred))/2.0
         solver = Solver(optimizer, momentum=0.9, wd=decay, learning_rate=l_rate,
                         lr_scheduler=lr_scheduler)
-        solver.set_metric(mx.metric.CustomMetric(l2_norm))
+        solver.set_metric(mx.gluon.metric.CustomMetric(l2_norm))
         solver.set_monitor(Monitor(print_every))
         data_iter = mx.io.NDArrayIter({'data': X}, batch_size=batch_size, shuffle=True,
                                       last_batch_handle='roll_over')
@@ -188,7 +188,7 @@ def l2_norm(label, pred):
             return np.mean(np.square(label-pred))/2.0
         solver = Solver(optimizer, momentum=0.9, wd=decay, learning_rate=l_rate,
                         lr_scheduler=lr_scheduler)
-        solver.set_metric(mx.metric.CustomMetric(l2_norm))
+        solver.set_metric(mx.gluon.metric.CustomMetric(l2_norm))
         solver.set_monitor(Monitor(print_every))
         data_iter = mx.io.NDArrayIter({'data': X}, batch_size=batch_size, shuffle=True,
                                       last_batch_handle='roll_over')

example/deep-embedded-clustering/dec.py

@@ -122,7 +122,7 @@ def cluster(self, X, y=None, update_interval=None):
 
         def ce(label, pred):
             return np.sum(label*np.log(label/(pred+0.000001)))/label.shape[0]
-        solver.set_metric(mx.metric.CustomMetric(ce))
+        solver.set_metric(mx.gluon.metric.CustomMetric(ce))
 
         label_buff = np.zeros((X.shape[0], self.num_centers))
         train_iter = mx.io.NDArrayIter({'data': X}, {'label': label_buff}, batch_size=batch_size,

example/distributed_training-horovod/gluon_mnist.py

@@ -104,7 +104,7 @@ def conv_nets():
 # Function to evaluate accuracy for a model
 def evaluate(model, data_iter, context):
     data_iter.reset()
-    metric = mx.metric.Accuracy()
+    metric = mx.gluon.metric.Accuracy()
     for _, batch in enumerate(data_iter):
         data = batch.data[0].as_in_context(context)
         label = batch.label[0].as_in_context(context)
@@ -149,7 +149,7 @@ def evaluate(model, data_iter, context):
 
 # Create loss function and train metric
 loss_fn = gluon.loss.SoftmaxCrossEntropyLoss()
-metric = mx.metric.Accuracy()
+metric = mx.gluon.metric.Accuracy()
 
 # Train model
 for epoch in range(args.epochs):

example/distributed_training-horovod/module_mnist.py

@@ -157,7 +157,7 @@ def conv_net():
           num_epoch=args.epochs)  # train for at most 10 dataset passes
 
 # Step 7: evaluate model accuracy
-acc = mx.metric.Accuracy()
+acc = mx.gluon.metric.Accuracy()
 model.score(val_iter, acc)
 
 if hvd.rank() == 0:

example/distributed_training-horovod/resnet50_imagenet.py

@@ -286,8 +286,8 @@ def evaluate(epoch):
             return
 
         val_data.reset()
-        acc_top1 = mx.metric.Accuracy()
-        acc_top5 = mx.metric.TopKAccuracy(5)
+        acc_top1 = mx.gluon.metric.Accuracy()
+        acc_top5 = mx.gluon.metric.TopKAccuracy(5)
         for _, batch in enumerate(val_data):
             data, label = batch_fn(batch, context)
             output = net(data.astype(args.dtype, copy=False))
@@ -321,7 +321,7 @@ def evaluate(epoch):
 
     # Create loss function and train metric
     loss_fn = gluon.loss.SoftmaxCrossEntropyLoss()
-    metric = mx.metric.Accuracy()
+    metric = mx.gluon.metric.Accuracy()
 
     # Train model
     for epoch in range(args.num_epochs):
@@ -450,8 +450,8 @@ def train_module():
 
     # Evaluate performance if not using synthetic data
     if args.use_rec:
-        acc_top1 = mx.metric.Accuracy()
-        acc_top5 = mx.metric.TopKAccuracy(5)
+        acc_top1 = mx.gluon.metric.Accuracy()
+        acc_top5 = mx.gluon.metric.TopKAccuracy(5)
         res = mod.score(val_data, [acc_top1, acc_top5])
         for name, val in res:
             logging.info('Epoch[%d] Rank[%d] Validation-%s=%f',

example/distributed_training/cifar10_dist.py

@@ -121,7 +121,7 @@ def evaluate_accuracy(data_iterator, network):
     ----------
     tuple of array element
     """
-    acc = mx.metric.Accuracy()
+    acc = mx.gluon.metric.Accuracy()
 
     # Iterate through data and label
     for i, (data, label) in enumerate(data_iterator):

example/distributed_training/cifar10_kvstore_hvd.py

@@ -123,7 +123,7 @@ def evaluate(data_iterator, network, context):
     ----------
     tuple of array element
     """
-    acc = mx.metric.Accuracy()
+    acc = mx.gluon.metric.Accuracy()
 
     # Iterate through data and label
     for i, (data, label) in enumerate(data_iterator):
@@ -208,7 +208,7 @@ def __len__(self):
                         optimizer_params={'learning_rate': args.lr},
                         kvstore=store)
 
-train_metric = mx.metric.Accuracy()
+train_metric = mx.gluon.metric.Accuracy()
 
 # Run as many epochs as required
 for epoch in range(args.epochs):

example/fcn-xs/solver.py

@@ -23,7 +23,7 @@
 from collections import namedtuple
 from mxnet import optimizer as opt
 from mxnet.optimizer import get_updater
-from mxnet import metric
+from mxnet.gluon import metric
 
 # Parameter to pass to batch_end_callback
 BatchEndParam = namedtuple('BatchEndParams', ['epoch', 'nbatch', 'eval_metric'])

example/gluon/audio/urban_sounds/train.py

@@ -28,7 +28,7 @@
 
 def evaluate_accuracy(data_iterator, net):
     """Function to evaluate accuracy of any data iterator passed to it as an argument"""
-    acc = mx.metric.Accuracy()
+    acc = mx.gluon.metric.Accuracy()
     for data, label in data_iterator:
         output = net(data)
         predictions = nd.argmax(output, axis=1)

example/gluon/dc_gan/dcgan.py

@@ -259,7 +259,7 @@ def main():
     real_label = mx.nd.ones((opt.batch_size,), ctx=ctx)
     fake_label = mx.nd.zeros((opt.batch_size,), ctx=ctx)
 
-    metric = mx.metric.Accuracy()
+    metric = mx.gluon.metric.Accuracy()
     print('Training... ')
     stamp = datetime.now().strftime('%Y_%m_%d-%H_%M')