Cpruce opened a new issue #12669: Dropout output shape not the same as input shape URL: https://github.com/apache/incubator-mxnet/issues/12669 Note: Providing complete information in the most concise form is the best way to get help. This issue template serves as the checklist for essential information to most of the technical issues and bug reports. For non-technical issues and feature requests, feel free to present the information in what you believe is the best form. For Q & A and discussion, please start a discussion thread at https://discuss.mxnet.io ## Description The shape of Dropout's output doesn't match that of the input when following a MaxPool2D: https://mxnet.incubator.apache.org/api/python/gluon/nn.html?highlight=dropout#mxnet.gluon.nn.Dropout (see outputs). I believe this is a bug since Dropout documentation even says it should retain shape and the keras model I am converting this from keeps the same shape, but please let me know if I am missing something or if this is fixed. ## Environment info (Required) ``` ----------Python Info---------- Version : 3.6.5 Compiler : GCC 7.2.0 Build : ('default', 'Mar 29 2018 18:21:58') Arch : ('64bit', '') ------------Pip Info----------- Version : 18.0 Directory : /home/cory/anaconda3/lib/python3.6/site-packages/pip ----------MXNet Info----------- Version : 1.2.1 Directory : /home/cory/anaconda3/lib/python3.6/site-packages/mxnet Commit Hash : 106391a1f0ee012b1ea38764d711e76774ce77e1 ----------System Info---------- Platform : Linux-4.15.0-34-generic-x86_64-with-debian-stretch-sid system : Linux node : sprucemoose release : 4.15.0-34-generic version : #37~16.04.1-Ubuntu SMP Tue Aug 28 10:44:06 UTC 2018 ----------Hardware Info---------- machine : x86_64 processor : x86_64 Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Byte Order: Little Endian CPU(s): 8 On-line CPU(s) list: 0-7 Thread(s) per core: 2 Core(s) per socket: 4 Socket(s): 1 NUMA node(s): 1 Vendor ID: GenuineIntel CPU family: 6 Model: 94 Model name: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz Stepping: 3 CPU MHz: 4200.111 CPU max MHz: 4200.0000 CPU min MHz: 800.0000 BogoMIPS: 8016.00 Virtualization: VT-x L1d cache: 32K L1i cache: 32K L2 cache: 256K L3 cache: 8192K NUMA node0 CPU(s): 0-7 Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 monitor ds_cpl vmx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb invpcid_single pti ssbd ibrs ibpb stibp tpr_shadow vnmi flexpriority ept vpid fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm mpx rdseed adx smap clflushopt intel_pt xsaveopt xsavec xgetbv1 xsaves dtherm ida arat pln pts hwp hwp_notify hwp_act_window hwp_epp flush_l1d ----------Network Test---------- Setting timeout: 10 Timing for MXNet: https://github.com/apache/incubator-mxnet, DNS: 0.0101 sec, LOAD: 0.5889 sec. Timing for Gluon Tutorial(en): http://gluon.mxnet.io, DNS: 0.1829 sec, LOAD: 0.3296 sec. Timing for Gluon Tutorial(cn): https://zh.gluon.ai, DNS: 0.1814 sec, LOAD: 0.2691 sec. Timing for FashionMNIST: https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/dataset/fashion-mnist/train-labels-idx1-ubyte.gz, DNS: 0.0288 sec, LOAD: 0.2088 sec. Timing for PYPI: https://pypi.python.org/pypi/pip, DNS: 0.0113 sec, LOAD: 0.5126 sec. Timing for Conda: https://repo.continuum.io/pkgs/free/, DNS: 0.0156 sec, LOAD: 0.1057 sec. ``` Package used (Python/R/Scala/Julia): Python: mxnet-cu91 ## Error Message: My output is conv1 [(32, 16, 101, 101)] conv1 [(32, 16, 101, 101)] conv1 [(32, 16, 101, 101)] pool1 [(32, 16, 50, 50)] pool1 [(32, 16, 25, 25)] <------- here is the problem. The output shape probably should be (32, 16, 50, 50) which should be similar to conv1 Tensor("conv2d_47/BiasAdd:0", shape=(?, 101, 101, 16), dtype=float32) conv1 Tensor("add_19/add:0", shape=(?, 101, 101, 16), dtype=float32) conv1 Tensor("activation_51/Relu:0", shape=(?, 101, 101, 16), dtype=float32) pool1 Tensor("max_pooling2d_5/MaxPool:0", shape=(?, 50, 50, 16), dtype=float32) pool1 Tensor("dropout_9/cond/Merge:0", shape=(?, 50, 50, 16), dtype=float32) Note that the mxnet model is in nchw and the keras model is nhwc ## Minimum reproducible example def build_model(input_layer, start_neurons, DropoutRatio = 0.5): # 101 -> 50 k_size = (3, 3) same_padding = (k_size[0]//2, k_size[1]//2) #input_layer = mx.sym.transpose(input_layer, [0, 3, 1, 2]) conv1 = mx.gluon.nn.Conv2D(start_neurons * 1, kernel_size=k_size, padding=same_padding)(input_layer) print('conv1', conv1.infer_shape(data=(32, 1, 101, 101))[1]) conv1 = residual_block(conv1,start_neurons * 1) print('conv1', conv1.infer_shape(data=(32, 16, 101, 101))[1]) conv1 = residual_block(conv1,start_neurons * 1, True) print('conv1', conv1.infer_shape(data=(32, 16, 101, 101))[1]) pool1 = mx.gluon.nn.MaxPool2D()(conv1) #(2, 2) print('pool1', pool1.infer_shape(data=(32, 16, 101, 101))[1]) pool1 = mx.gluon.nn.Dropout(DropoutRatio/2)(pool1) print('pool1', pool1.infer_shape(data=(32, 16, 50, 50))[1]) which is converted from def build_model(input_layer, start_neurons, DropoutRatio = 0.5): # 101 -> 50 conv1 = Conv2D(start_neurons * 1, (3, 3), activation=None, padding="same")(input_layer) print('conv1', conv1) conv1 = residual_block(conv1,start_neurons * 1) print('conv1', conv1) conv1 = residual_block(conv1,start_neurons * 1, True) print('conv1', conv1) pool1 = MaxPooling2D((2, 2))(conv1) print('pool1', pool1) pool1 = Dropout(DropoutRatio/2)(pool1) print('pool1', pool1) See attached notebook to reproduce ## Steps to reproduce 1. Download this notebook. [mxnet_kernel.ipynb.zip](https://github.com/apache/incubator-mxnet/files/2417812/mxnet_kernel.ipynb.zip) 2. Remove the data sections as you don't need the data to hit the problem. 3. Run the notebook. ## What have you tried to solve it? 1. Looking at parameters for mxnet MaxPool2D v keras MaxPooling2D as also with Dropout. AFAIK, params are the same. 2. Commenting out the preceding MaxPool2D (Dropout still reduces hw sizes). 3. I've started looking into the codebase and see the implementation in operator/nn/dropout-inl.h and operator/nn/dropout.cc. I've checked out 1.2.1 (my mxnet version) and will likely continue debugging there.
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
