udibr

import sys

if len(sys.argv) < 2:
	print "Usage: %s <file-name>"%sys.argv[0]
	exit(0)
fname = sys.argv[1]

import json

lookup = {'outputs':[], 'execution_count':None}

udibr

import ctypes

# Path to location of libcudart
# Change the path to "cudart<x>_<ver>.dll" to use on Windows, where <x>={32|64} and <ver> is the CUDA version. 
_CUDA = "/usr/local/cuda/lib/libcudart.dylib"
cuda = ctypes.cdll.LoadLibrary(_CUDA)

cuda.cudaMemGetInfo.restype = int
cuda.cudaMemGetInfo.argtypes = [ctypes.c_void_p, ctypes.c_void_p]

udibr

# variation to https://github.com/ryankiros/skip-thoughts/blob/master/decoding/search.py

def keras_rnn_predict(samples, empty=empty, rnn_model=model, maxlen=maxlen):
    """for every sample, calculate probability for every possible label
    you need to supply your RNN model and maxlen - the length of sequences it can handle
    """
    data = sequence.pad_sequences(samples, maxlen=maxlen, value=empty)
    return rnn_model.predict(data, verbose=0)

def beamsearch(predict=keras_rnn_predict,

udibr

import numpy as np
import h5py
import keras.backend as K

def str_shape(x):
    return 'x'.join(map(str, x.shape))

def load_weights(model, filepath, lookup={}, ignore=[], transform=None, verbose=True):
    """Modified version of keras load_weights that loads as much as it can.
    Useful for transfer learning.

udibr

from keras.layers.core import Lambda

class RemoveMask(Lambda):
    def __init__(self):
        super(RemoveMask, self).__init__((lambda x, mask: x))
        self.supports_masking = True

    def compute_mask(self, input, input_mask=None):
        return None

udibr

import numpy as np
from keras.layers import GRU, initializations, K
from collections import OrderedDict

class GRULN(GRU):
    '''Gated Recurrent Unit with Layer Normalization
    Current impelemtation only works with consume_less = 'gpu' which is already
    set.

    # Arguments

udibr

0x4CDE69AE1A924010Ba33F2944A4bb6dfbFB64958