search

hardmaru / supercell

supercell
190 stars 42 forks source link

Thanks and a question about init the hypercell #4

Open AlbertDu opened 7 years ago

AlbertDu commented 7 years ago

Thanks for you code and the design of Hypercell is amazing! I have a question about the initialization of the cell.

self.cell= supercell.HyperLSTMCell(num_units=args.state_size.....) output, self.last_state = tf.nn.dynamic_rnn(self.cell, self.X, initial_state=??? )

While I use the basic LSTMcell, the tf has a method to init the cell into zero state.

However the zero init perform not very good, do you have any idea of a better way to init the Hypercell?

Thanks a lot again!

hardmaru commented 7 years ago

Hi Albert

I'm in the middle of cleaning this code base up

Let me take a look and get back to you maybe in the next week or two

Thanks

On Sun, Apr 30, 2017 at 8:37 AM AlbertDu notifications@github.com wrote:

Thanks for you code and the design of Hypercell is amazing! I have a question about the initialization of the cell.

self.cell= supercell.HyperLSTMCell(num_units=args.state_size.....) output, self.last_state = tf.nn.dynamic_rnn(self.cell, self.X, initial_state=??? )

While I use the basic LSTMcell, the tf has a method to init the cell into zero state.

However the zero init perform not very good, do you have any idea of a better way to init the Hypercell?

Thanks a lot again!

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/hardmaru/supercell/issues/4, or mute the thread https://github.com/notifications/unsubscribe-auth/AGBoHtyAEz1d-hqpXi8n2kwtC3fiiZmcks5r1KrFgaJpZM4NMmjI .

AlbertDu commented 7 years ago

ok,

Thanks again for you reply.

Best wishes and looking forward to your new version of HyperCell

hardmaru commented 7 years ago

Btw, have you played with using any alternative methods? If you have any suggestions as to which method is faster your current usecases, I'll incorporate them in for the next code release.

On Mon, May 1, 2017 at 6:00 AM AlbertDu notifications@github.com wrote:

ok,

Thanks again for you reply.

Best wishes and looking forward to your new version of HyperCell

— You are receiving this because you commented.

Reply to this email directly, view it on GitHub https://github.com/hardmaru/supercell/issues/4#issuecomment-298329928, or mute the thread https://github.com/notifications/unsubscribe-auth/AGBoHh6bfcLo7fKBDx7qfE0Z57TFeGAMks5r1ddbgaJpZM4NMmjI .

AlbertDu commented 7 years ago

Sorry, after I transplant the cell.zero_state from the BasicLSTMcell, I still could not run the "supercell.py" like this(but if i change the HyperLSTM to LSTM, it works well)

self.cell= supercell.HyperLSTMCell(num_units=args.state_size.....) output, self.last_state = tf.nn.dynamic_rnn(self.cell, self.X, initial_state=??? )

do you have any simple case to import supercell which could run without ERROR?