scorematrix = np.random.randn(C + 1, T)
scorematrix -= np.max(scorematrix, axis=0)
scorematrix = np.exp(scorematrix)
scorematrix /= np.sum(scorematrix, axis=0)
blank = C
B = 1
for _ in range(10):
seq = np.floor(np.random.rand(L) * C).astype(np.int32)
time0 = time.time()
NLL_cython = ctc_cython(scorematrix=scorematrix, queryseq=np.array(seq, dtype=np.int32), blank=blank)
time1 = time.time()
NLL_numba, alphas = ctc_numba(scorematrix, seq, blank=blank)
time2 = time.time()
NLL_theano_nonbatch = f2(scorematrix, seq, blank)
time3 = time.time()
y = seq.reshape([L, B])
yhat = scorematrix.T.reshape([T, C + 1, B])
NLL_theano_batch = f1(y, yhat, np.ones_like(y, dtype=np.float32), np.ones([T, 1], dtype=np.float32), blank)
time4 = time.time()
print(
"NLL_cython = %f, NLL_numba = %f, NLL_theano_nonbatch = %f, NLL_theano_batch = %f"
% (NLL_cython, NLL_numba, NLL_theano_nonbatch[0], NLL_theano_batch)
)
print(
TE += te
TD += td
LOSS += ctcloss*batch
n += batch
aveloss = LOSS / n
print('epoch = %d, batch = %d, ave_loss = %0.4f, ave_CER = %0.4f, progress = %0.1f' % (j, i, aveloss, TE/TD, n/B))
# print('batch loss = %0.4f, CER = %0.4f' % (ctcloss, cer))
# print()
b, t, cp = scorematrixT.shape
scorematrix = scorematrixT.T.reshape(cp, t)
resultseq2 = best_path_decode(scorematrix)
rl = np.sum(resultseq_mask)
queryseq = y_train2[i,:].astype(np.int32)
ctcloss2 = ctc_numba(scorematrix, queryseq)[0]
te2 = editdist(resultseq2, list(queryseq))
td2 = len(queryseq)
TE2 += te2
TD2 += td2
print('batch loss = %0.4f, ctc_numba = %0.4f, batch_CER = %0.4f, batch_CER2 = %0.4f' % (ctcloss, ctcloss2, cer, te2/td2))
print('ave_CER = %0.4f, ave_CER2 = %0.4f' % (TE/TD, TE2/TD2))
print()
resultseq1 = list((resultseq[0:rl].T).astype(np.int32)[0])
if resultseq1 != resultseq2:
print('result seq not same')
print('resultseq1 =', resultseq1)
print('resultseq2 =', resultseq2)
scorematrix = np.random.randn(C+1, T)
scorematrix -= np.max(scorematrix, axis=0)
scorematrix = np.exp(scorematrix)
scorematrix /= np.sum(scorematrix, axis=0)
queryseq = np.array([0, 1, 2, 1, 1])
blank = -1
# with open('myctcdebug2.pkl', 'rb') as fl:
# scorematrix, queryseq, blank = pickle.load(fl)
# fl.close()
# print(scorematrix.shape)
# print(queryseq.shape)
# print(blank)
time0 = time.time()
result0 = ctc_numba(scorematrix, queryseq, blank)
NLL0 = result0[0]
time1 = time.time()
result1 = f(scorematrix, queryseq, blank)
NLL1 = result1[0]
time2 = time.time()
print('NLL0 = %f, NLL1 = %f' % (NLL0, NLL1))
print('time0 = %0.2f, time1 = %0.2f' %(time1-time0, time2-time1))
# f.profile.print_summary()
# print(result1[1][-1])
# print('result1.shape = ', result1[1][-1].shape)
# print(result0[1])
# print('result0.shape = ', result0[1].shape)
