Using large number of output classes in RNN model gives error "blob size exceeds INT_MAX"

[Prachi Jain]

In a Language Model, vocabulary size is equal to the number of output of the network.

When using a vocabulary size of ~3,000,000 in my RNN model, I got the following error:

I1020 19:48:51.488100  6972 net.cpp:155] Setting up rnn1_layer
I1020 19:48:51.488250  6972 net.cpp:163] Top shape: 1044 24 200 (5011200)
I1020 19:48:51.488281  6972 layer_factory.hpp:76] Creating layer inner_product
I1020 19:48:51.488297  6972 net.cpp:110] Creating Layer inner_product
I1020 19:48:51.488306  6972 net.cpp:477] inner_product <- rnn1_layer
I1020 19:48:51.488319  6972 net.cpp:433] inner_product -> inner_product
F1020 19:48:55.256937  6972 blob.cpp:29] Check failed: shape[i] <= 2147483647 / count_ (3000002 vs. 85707) blob size exceeds INT_MAX
*** Check failure stack trace: ***
    @     0x7fd9af53bb7d  google::LogMessage::Fail()
    @     0x7fd9af53dc7f  google::LogMessage::SendToLog(

)
    @     0x7fd9af53b76c  google::LogMessage::Flush()
    @     0x7fd9af53e51d  google::LogMessageFatal::~LogMessageFatal()
    @     0x7fd9af879d60  caffe::Blob<>::Reshape()
    @     0x7fd9af919b6d  caffe::InnerProductLayer<>::Reshape()
    @     0x7fd9af9d0083  caffe::Net<>::Init()
    @     0x7fd9af9d1aa4  caffe::Net<>::Net()
    @     0x7fd9af9e99e2  caffe::Solver<>::InitTrainNet()
    @     0x7fd9af9e9fec  caffe::Solver<>::Init()
    @     0x7fd9af9eacc7  caffe::Solver<>::Solver()
    @           0x415dd5  caffe::GetSolver<>()
    @           0x40de29  train()
    @           0x40a7db  main
    @     0x7fd9ae4b676d  (unknown)
    @           0x40a8e9  (unknown)
Aborted (core dumped)

Any help in resolving the error will be great!

[Prachi Jain]

On reducing the vocabulary size (number of output classes) to 82994, training continued smoothly.

But I got the following error during testing:
I1021 09:12:56.200850 21907 layer_factory.hpp:76] Creating layer inner_product
I1021 09:12:56.201725 21907 net.cpp:110] Creating Layer inner_product
I1021 09:12:56.202826 21907 net.cpp:477] inner_product <- rnn1_layer
I1021 09:12:56.204144 21907 net.cpp:433] inner_product -> inner_product
F1021 09:12:56.310900 21907 blob.cpp:29] Check failed: shape[i] <= 2147483647 / count_ (82994 vs. 11427) blob size exceeds INT_MAX

I am unsure about the issue. Any help will be great!