def dealWithOneDoc(rng, doc):
sentenceList = doc
sentenceSize = ttList.length(sentenceList)
sentenceResults, _ = theano.scan(
fn=lambda i, sl: dealWithSentence(rng, sl[i]),
non_sequences=[sentenceList],
sequences=[T.arange(sentenceSize, dtype='int64')])
sentenceResults = sentenceResults.dimshuffle([0, 2])
docW = theano.shared(numpy.asarray(rng.uniform(low=-0.2,
high=0.2,
size=(2, 1, 2)),
dtype=theano.config.floatX),
borrow=True)
doc_out = theano.tensor.signal.conv.conv2d(input=sentenceResults,
filters=docW)
# docPool = T.max(doc_out, axis = [0])
docPool = theano.tensor.signal.downsample.max_pool_2d(doc_out, (1000, 1))
doc_embedding = T.reshape(docPool, (1, 2 * 5)).dimshuffle([1])
# doc_embeddingp = printing.Print('doc_embedding')
# doc_embedding = doc_embeddingp(doc_embedding)
return doc_embedding
def dealWithOneDoc(rng, doc):
sentenceList = doc
sentenceSize = ttList.length(sentenceList)
sentenceResults, _ = theano.scan(fn=lambda i, sl: dealWithSentence(rng, sl[i]),
non_sequences=[sentenceList],
sequences=[T.arange(sentenceSize, dtype='int64')])
sentenceResults = sentenceResults.dimshuffle([0, 2])
docW = theano.shared(
numpy.asarray(
rng.uniform(low=-0.2, high=0.2, size=(2, 1, 2)),
dtype=theano.config.floatX
),
borrow=True
)
doc_out = theano.tensor.signal.conv.conv2d(input=sentenceResults, filters=docW)
# docPool = T.max(doc_out, axis = [0])
docPool = theano.tensor.signal.downsample.max_pool_2d(doc_out, (1000, 1))
doc_embedding = T.reshape(docPool, (1, 2 * 5)).dimshuffle([1])
# doc_embeddingp = printing.Print('doc_embedding')
# doc_embedding = doc_embeddingp(doc_embedding)
return doc_embedding
def main2():
a = [[[[1, 2, 3]], [[4, 5]]]]
tl = ttList.TypedListType(ttList.TypedListType(theano.tensor.fmatrix))()
o = ttList.length(tl)
f = theano.function([tl], o)
testRes = f(a)
print testRes
print "All finished!"
def main2():
a = [
[
[[1, 2, 3]], [[4, 5]]
]
]
tl = ttList.TypedListType(ttList.TypedListType(theano.tensor.fmatrix))()
o = ttList.length(tl)
f = theano.function([tl], o)
testRes = f(a)
print testRes
print "All finished!"
def main():
rng = numpy.random.RandomState(23455)
a = [[[[1, 2, 3, 4], [1, 2, 3, 4], [1, 2, 3, 4]],
[[1, 2, 3, 4], [1, 2, 3, 4]]]]
docList = ttList.TypedListType(ttList.TypedListType(T.fmatrix))("docList")
docSize = ttList.length(docList)
modelResults, _ = theano.scan(fn=lambda i, tl: dealWithOneDoc(rng, tl[i]),
non_sequences=[docList],
sequences=[T.arange(docSize, dtype='int64')])
testFunc = theano.function([docList], modelResults)
rrrr = testFunc(a)
print rrrr[0]
print "All finished!"
def main():
rng = numpy.random.RandomState(23455)
a = [
[
[[1, 2, 3, 4], [1, 2, 3, 4], [1, 2, 3, 4]],
[[1, 2, 3, 4], [1, 2, 3, 4]]
]
]
docList = ttList.TypedListType(ttList.TypedListType(T.fmatrix))("docList")
docSize = ttList.length(docList)
modelResults, _ = theano.scan(fn=lambda i, tl: dealWithOneDoc(rng, tl[i]),
non_sequences=[docList],
sequences=[T.arange(docSize, dtype='int64')])
testFunc = theano.function([docList], modelResults);
rrrr = testFunc(a)
print rrrr[0]
print "All finished!"
emb_matrix_path = 'embedding_matrix_gensim_300D.npy'
Vocabulary_size = word_dim
x_train, word_to_index, index_to_word = load_data(INPUT_DATA_FILE, Vocabulary_size)
x_test = x_train[400000:500000]
x_train = x_train[0:400000]
#iterator counter
t = theano.shared(name = 't', value = np.array(0).astype('int32'))
x = tlist.TypedListType(T.ivector)()
#wl = T.ivector('wl')
l = tlist.length(x)
def get_shapes(index, x):
shape_ex = T.shape(x[index])
return shape_ex[0]
x_shapes, last_output = theano.scan(fn=get_shapes,
outputs_info=None,
non_sequences = [x],
sequences = [T.arange(batch_size, dtype = 'int64')]
)
#f = theano.function([x], T.shape(x[T.argmax(x_shapes)]))
max_x_idx = T.argmax(x_shapes)
