reminisce opened a new issue #16559: Tracking mxnet.numpy operator issues for 1.6.0 release URL: https://github.com/apache/incubator-mxnet/issues/16559 ## Issues from running D2L [numpy2](https://github.com/d2l-ai/d2l-en/tree/numpy2) branch 1. chapter_preliminaries/probability.md ``` TypeError: no implementation found for 'numpy.broadcast_arrays' on types that implement __array_function__: [<class 'numpy.ndarray'>, <class 'mxnet.numpy.ndarray'>] ``` 2. chapter_preliminaries/probability.md ``` MXNetError: [09:16:23] src/operator/numpy/np_true_divide.cc:43: Check failed: lhs_dtype == rhs_dtype (7 vs. 0) : true_divide currently only supports same dtype for dividend and divisor ``` 3. chapter_multilayer-perceptrons/dropout.md ``` mask = np.random.uniform(0, 1, X.shape) > drop_prob return mask * X / (1.0-drop_prob) Fails on the second line there, should be the problem with boolean multiplying with float ``` 4. chapter_deep-learning-computation/parameters.md ``` data[:] = np.random.uniform(-10, 10, data.shape) data *= np.abs(data) >= 5 Fails on the 2nd line, still the multiplication between boolean & float ``` 5. chapter_recurrent-neural-networks/seq2seq.md ``` MXNetError: [10:46:26] src/operator/numpy/np_true_divide.cc:43: Check failed: lhs_dtype == rhs_dtype (0 vs. 6) : true_divide currently only supports same dtype for dividend and divisor<ipython-input-10-23a855fab898> in train_s2s_ch8(model, data_iter, lr, num_epochs, ctx) 24 metric.add(l.sum(), num_tokens) 25 if epoch % 10 == 0: ---> 26 animator.add(epoch, (metric[0]/metric[1],)) 27 print('loss %.3f, %d tokens/sec on %s ' % ( 28 metric[0]/metric[1], metric[1]/timer.stop(), ctx)) -- ``` 6. All notebooks with train_s2s_ch8 ``` MXNetError: [10:46:26] src/operator/numpy/np_true_divide.cc:43: Check failed: lhs_dtype == rhs_dtype (0 vs. 6) : true_divide currently only supports same dtype for dividend and divisor <ipython-input-10-23a855fab898> in train_s2s_ch8(model, data_iter, lr, num_epochs, ctx) 24 metric.add(l.sum(), num_tokens) 25 if epoch % 10 == 0: ---> 26 animator.add(epoch, (metric[0]/metric[1],)) 27 print('loss %.3f, %d tokens/sec on %s ' % ( 28 metric[0]/metric[1], metric[1]/timer.stop(), ctx)) ``` 7. chapter_optimization/optimization-intro.md ``` ValueError: mxnet.numpy operator `<function column_stack at 0x7ff910066e18>` has not been registered in the _NUMPY_ARRAY_FUNCTION_LIST. Please make sure you are using NumPy >= 1.17.0 and the operator implementation is compatible with NumPy. Then add the operator name to the list. ``` 8. chapter_optimization/convexity.md ``` Same as above ``` 9. chapter_natural-language-processing/sentiment-analysis-rnn.md ``` ---------------------------------------------------------------------------MXNetError Traceback (most recent call last)<ipython-input-7-a7e697bf18e7> in <module>* 2* trainer = gluon.Trainer(net.collect_params(), 'adam', {'learning_rate': lr})* 3* loss = gluon.loss.SoftmaxCrossEntropyLoss()----> 4 d2l.train_ch12(net, train_iter, test_iter, loss, trainer, num_epochs, ctx)~/d2l-numpy/d2l/d2l.py in train_ch12(net, train_iter, test_iter, loss, trainer, num_epochs, ctx_list, split_f)* 1054* timer.start()* 1055* l, acc = train_batch_ch12(-> 1056 net, features, labels, loss, trainer, ctx_list, split_f)* 1057* metric.add(l, acc, labels.shape[0], labels.size)* 1058* timer.stop()~/d2l-numpy/d2l/d2l.py in train_batch_ch12(net, features, labels, loss, trainer, ctx_list, split_f)* 1037* l.backward()* 1038* trainer.step(features.shape[0])-> 1039 train_loss_sum = sum([float(l.sum()) for l in ls])* 1040* train_acc_sum = sum(d2l.accuracy(py, y) for py, y in zip(pys, ys))* 1041* return train_loss_sum, train_acc_sum ~/d2l-numpy/d2l/d2l.py in <listcomp>(.0)* 1037* l.backward()* 1038* trainer.step(features.shape[0])-> 1039 train_loss_sum = sum([float(l.sum()) for l in ls])* 1040* train_acc_sum = sum(d2l.accuracy(py, y) for py, y in zip(pys, ys))* 1041* return train_loss_sum, train_acc_sum ~/mxnet_master/python/mxnet/numpy/multiarray.py in __float__(self)* 791* if num_elements != 1:* 792* raise TypeError('only size-1 arrays can be converted to Python scalars')--> 793 return float(self.item())* 794** 795* def __int__(self):~/mxnet_master/python/mxnet/numpy/multiarray.py in item(self, *args)* 830* """* 831* # TODO(junwu): no need to call asnumpy() on the whole array.--> 832 return self.asnumpy().item(*args)* 833** 834* @property ~/mxnet_master/python/mxnet/ndarray/ndarray.py in asnumpy(self)* 2517* self.handle,* 2518* data.ctypes.data_as(ctypes.c_void_p),-> 2519 ctypes.c_size_t(data.size)))* 2520* return data* 2521*~/mxnet_master/python/mxnet/base.py in check_call(ret)* 252* """* 253* if ret != 0:--> 254 raise MXNetError(py_str(_LIB.MXGetLastError()))* 255** 256*MXNetError: [11:23:59] src/operator/./rnn-inl.h:1505: Check failed: e == cudaSuccess || e == cudaErrorCudartUnloading: CUDA: invalid resource handle Stack trace: [bt] (0) /home/ubuntu/mxnet_master/python/mxnet/../../lib/libmxnet.so(dmlc::LogMessageFatal::~LogMessageFatal()+0x43) [0x7fe81ce57ff3] [bt] (1) /home/ubuntu/mxnet_master/python/mxnet/../../lib/libmxnet.so(mxnet::op::RNNOp<mshadow::gpu, float, float>::Backward(mxnet::OpContext const&, std::vector<mxnet::TBlob, std::allocator<mxnet::TBlob> > const&, std::vector<mxnet::TBlob, std::allocator<mxnet::TBlob> > const&, std::vector<mxnet::TBlob, std::allocator<mxnet::TBlob> > const&, std::vector<mxnet::OpReqType, std::allocator<mxnet::OpReqType> > const&, std::vector<mxnet::TBlob, std::allocator<mxnet::TBlob> > const&)+0x129a) [0x7fe82357b35a] [bt] (2) /home/ubuntu/mxnet_master/python/mxnet/../../lib/libmxnet.so(void mxnet::op::RNNStatefulGradCompute<mshadow::gpu>(mxnet::OpStatePtr const&, mxnet::OpContext const&, std::vector<mxnet::TBlob, std::allocator<mxnet::TBlob> > const&, std::vector<mxnet::OpReqType, std::allocator<mxnet::OpReqType> > const&, std::vector<mxnet::TBlob, std::allocator<mxnet::TBlob> > const&)+0x146e) [0x7fe8235b508e] ``` **Note:** Caused by #16391 (https://github.com/apache/incubator-mxnet/pull/16391), verified fix by reverting the PR. Also pinged DickJC123 for a fix 10. chapter_recommender-systems/autorec.md ``` ---------------------------------------------------------------------------MXNetError Traceback (most recent call last)<ipython-input-4-cc01b0d0d033> in <module>* 2* # Load the MovieLens 100K dataset* 3* df, num_users, num_items = d2l.read_data_ml100k()----> 4 train_data, test_data = d2l.split_data_ml100k(df, num_users, num_items)* 5* _, _, _, train_inter_mat = d2l.load_data_ml100k(train_data, num_users,* 6* num_items) ~/d2l-numpy/d2l/d2l.py in split_data_ml100k(data, num_users, num_items, split_mode, test_ratio)* 1432* else:* 1433* mask = [True if x == 1 else False for x in np.random.uniform(-> 1434 0, 1, (len(data))) < 1 - test_ratio]* 1435* neg_mask = [not x for x in mask]* 1436* train_data, test_data = data[mask], data[neg_mask]~/d2l-numpy/d2l/d2l.py in <listcomp>(.0)* 1431* test_data = pd.DataFrame(test_data)* 1432* else:-> 1433 mask = [True if x == 1 else False for x in np.random.uniform(* 1434* 0, 1, (len(data))) < 1 - test_ratio]* 1435* neg_mask = [not x for x in mask]~/mxnet_master/python/mxnet/numpy/multiarray.py in __bool__(self)* 781* return False* 782* elif num_elements == 1:--> 783 return bool(self.item())* 784* else:* 785* raise ValueError("The truth value of an ndarray with multiple elements is ambiguous.")~/mxnet_master/python/mxnet/numpy/multiarray.py in item(self, *args)* 830* """* 831* # TODO(junwu): no need to call asnumpy() on the whole array.--> 832 return self.asnumpy().item(*args)* 833** 834* @property ~/mxnet_master/python/mxnet/ndarray/ndarray.py in asnumpy(self)* 2517* self.handle,* 2518* data.ctypes.data_as(ctypes.c_void_p),-> 2519 ctypes.c_size_t(data.size)))* 2520* return data* 2521*~/mxnet_master/python/mxnet/base.py in check_call(ret)* 252* """* 253* if ret != 0:--> 254 raise MXNetError(py_str(_LIB.MXGetLastError()))* 255** 256*MXNetError: [11:27:58] src/operator/numpy/../tensor/elemwise_binary_scalar_op.h:264: Unknown type enum 7 ``` **Note:** Verified fix by changing that type switch at src/operator/tensor/elemwise_binary_scalar_op.h:264 to the "WITH_BOOL" version
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
