def build_lookup(repo, output_dico, filename_load):
dwin = 9
with closing(open(os.path.join(repo, output_dico), 'rb')) as f:
dico = pickle.load(f)
n_mot = [len(dico[i]) for i in dico.keys()]
vect_size = [20, 10, 5, 5]
n_hidden = 25
x = T.imatrix('x')
t_nlp = LookUpTrain(dwin, n_mot, vect_size, n_hidden)
t_nlp.initialize()
t_nlp.load(repo, filename_load)
lookup = theano.function(inputs=[x], outputs=t_nlp.embedding(x), allow_input_downcast=True)
return lookup
def test_embedding(): n_mot = [6, 7, 8] vect_size = [23, 5, 7] n_hidden = 14 dwin = 5 window = LookUpTrain(dwin, n_mot, vect_size, n_hidden, n_out=1) window.initialize() x = T.itensor3() x_value = np.zeros((2, 3, dwin)).astype(int) for p in range(2): for i in range(3): for q in range(dwin): x_value[p,i,q] = np.random.randint(n_mot[i]) f = theano.function([x], window.embedding(x), allow_input_downcast=True) print f(x_value).shape