Skip to content

P
pytensor
提交 d35d01b1 authored 作者: João Victor Tozatti Risso's avatar João Victor Tozatti Risso
浏览文件
操作

Reuse testcases in both CPU and GPU tests of CTC

Signed-off-by: 's avatarJoão Victor Tozatti Risso <joaovictor.risso@gmail.com>
上级 b4f54eae
隐藏空白字符变更
内嵌 并排
正在显示 包含 77 行增加 和 106 行删除
theano/gpuarray/tests/test_ctc.py
浏览文件 @ d35d01b1
......@@ -9,6 +9,7 @@ from theano.tests import unittest_tools as utt
import theano.gpuarray
from theano.gpuarray.ctc import (gpu_ctc, GpuConnectionistTemporalClassification)
from theano.tensor.nnet.ctc import (ctc, ctc_available, ConnectionistTemporalClassification)
from theano.tensor.nnet.tests.test_ctc import (setup_torch_case, setup_ctc_case, setup_grad_case)
from .config import (mode_with_gpu, mode_without_gpu)
......@@ -118,53 +119,12 @@ class TestCTC(unittest.TestCase):
# Test obtained from Torch tutorial at:
# https://github.com/baidu-research/warp-ctc/blob/master/torch_binding/TUTORIAL.md
def test_torch_case(self):
# Layout, from slowest to fastest changing dimension, is (time, batchSize, inputLayerSize)
activations = np.asarray([[[0, 0, 0, 0, 0], [1, 2, 3, 4, 5], [-5, -4, -3, -2, -1]],
[[0, 0, 0, 0, 0], [6, 7, 8, 9, 10], [-10, -9, -8, -7, -6]],
[[0, 0, 0, 0, 0], [11, 12, 13, 14, 15], [-15, -14, -13, -12, -11]]],
dtype=np.float32)
# Duration of each sequence
activation_times = np.asarray([1, 3, 3], dtype=np.int32)
# Labels for each sequence
labels = np.asarray([[1, -1],
[3, 3],
[2, 3]], dtype=np.int32)
expected_costs = np.asarray([1.609437943, 7.355742931, 4.938849926],
dtype=np.float32)
grads = [[[0.2, -0.8, 0.2, 0.2, 0.2],
[0.01165623125, 0.03168492019, 0.08612854034, -0.7658783197, 0.636408627],
[-0.02115798369, 0.03168492019, -0.8810571432, 0.2341216654, 0.636408627]],
[[0, 0, 0, 0, 0],
[-0.9883437753, 0.03168492019, 0.08612854034, 0.2341216654, 0.636408627],
[-0.02115798369, 0.03168492019, -0.1891518533, -0.4577836394, 0.636408627]],
[[0, 0, 0, 0, 0],
[0.01165623125, 0.03168492019, 0.08612854034, -0.7658783197, 0.636408627],
[-0.02115798369, 0.03168492019, 0.08612854034, -0.7330639958, 0.636408627]]]
expected_gradients = np.asarray(grads, dtype=np.float32)
self.check_ctc(activations, labels, activation_times, expected_costs, expected_gradients)
activations, labels, activation_times, expected_costs, expected_grads = setup_torch_case()
self.check_ctc(activations, labels, activation_times, expected_costs, expected_grads)
def test_ctc(self):
activations = np.asarray([[[0.1, 0.6, 0.1, 0.1, 0.1], [0.1, 0.1, 0.6, 0.1, 0.1]],
[[0.6, 0.1, 0.1, 0.1, 0.1], [0.1, 0.1, 0.5, 0.2, 0.1]]],
dtype=np.float32)
activation_times = np.asarray([2, 2], dtype=np.int32)
labels = np.asarray([[1, 2], [1, 2]], dtype=np.int32)
expected_costs = np.asarray([2.962858438, 3.053659201], dtype=np.float32)
grads = [[[0.177031219, -0.7081246376, 0.177031219, 0.177031219, 0.177031219],
[0.177031219, -0.8229685426, 0.291875124, 0.177031219, 0.177031219]],
[[0.291875124, 0.177031219, -0.8229685426, 0.177031219, 0.177031219],
[0.1786672771, 0.1786672771, -0.7334594727, 0.1974578798, 0.1786672771]]]
expected_gradients = np.asarray(grads, dtype=np.float32)
self.check_ctc(activations, labels, activation_times, expected_costs, expected_gradients)
activations, labels, input_length, expected_costs, expected_grads = setup_ctc_case()
self.check_ctc(activations, labels, input_length, expected_costs, expected_grads)
def test_verify_grad(self):
def ctc_op_functor(labels, in_lengths):
......@@ -175,13 +135,7 @@ class TestCTC(unittest.TestCase):
return gpu_ctc(acts, t_labels, t_activation_times)
return wrapper
activations = np.asarray([[[0.1, 0.6, 0.1, 0.1, 0.1], [0.1, 0.1, 0.6, 0.1, 0.1]],
[[0.6, 0.1, 0.1, 0.1, 0.1], [0.1, 0.1, 0.5, 0.2, 0.1]]],
dtype=np.float32)
activation_times = np.asarray([2, 2], dtype=np.int32)
labels = np.asarray([[1, 2], [1, 2]], dtype=np.int32)
activations, activation_times, labels = setup_grad_case()
ctc_op = ctc_op_functor(labels, activation_times)
......
theano/tensor/nnet/tests/test_ctc.py
浏览文件 @ d35d01b1
......@@ -9,6 +9,72 @@ from theano.tests import unittest_tools as utt
from theano.tensor.nnet.ctc import (ctc_available, ctc, ConnectionistTemporalClassification)
def setup_torch_case():
# Test obtained from Torch tutorial at:
# https://github.com/baidu-research/warp-ctc/blob/master/torch_binding/TUTORIAL.md
# Layout, from slowest to fastest changing dimension, is (time, batchSize, inputLayerSize)
activations = np.asarray([[[0, 0, 0, 0, 0], [1, 2, 3, 4, 5], [-5, -4, -3, -2, -1]],
[[0, 0, 0, 0, 0], [6, 7, 8, 9, 10], [-10, -9, -8, -7, -6]],
[[0, 0, 0, 0, 0], [11, 12, 13, 14, 15], [-15, -14, -13, -12, -11]]],
dtype=np.float32)
# Duration of each sequence
activation_times = np.asarray([1, 3, 3], dtype=np.int32)
# Labels for each sequence
labels = np.asarray([[1, -1],
[3, 3],
[2, 3]], dtype=np.int32)
expected_costs = np.asarray([1.609437943, 7.355742931, 4.938849926],
dtype=np.float32)
grads = [[[0.2, -0.8, 0.2, 0.2, 0.2],
[0.01165623125, 0.03168492019, 0.08612854034, -0.7658783197, 0.636408627],
[-0.02115798369, 0.03168492019, -0.8810571432, 0.2341216654, 0.636408627]],
[[0, 0, 0, 0, 0],
[-0.9883437753, 0.03168492019, 0.08612854034, 0.2341216654, 0.636408627],
[-0.02115798369, 0.03168492019, -0.1891518533, -0.4577836394, 0.636408627]],
[[0, 0, 0, 0, 0],
[0.01165623125, 0.03168492019, 0.08612854034, -0.7658783197, 0.636408627],
[-0.02115798369, 0.03168492019, 0.08612854034, -0.7330639958, 0.636408627]]]
expected_gradients = np.asarray(grads, dtype=np.float32)
return [activations, labels, activation_times, expected_costs, expected_gradients]
def setup_ctc_case():
activations = np.asarray([[[0.1, 0.6, 0.1, 0.1, 0.1], [0.1, 0.1, 0.6, 0.1, 0.1]],
[[0.6, 0.1, 0.1, 0.1, 0.1], [0.1, 0.1, 0.5, 0.2, 0.1]]],
dtype=np.float32)
activation_times = np.asarray([2, 2], dtype=np.int32)
labels = np.asarray([[1, 2], [1, 2]], dtype=np.int32)
expected_costs = np.asarray([2.962858438, 3.053659201], dtype=np.float32)
grads = [[[0.177031219, -0.7081246376, 0.177031219, 0.177031219, 0.177031219],
[0.177031219, -0.8229685426, 0.291875124, 0.177031219, 0.177031219]],
[[0.291875124, 0.177031219, -0.8229685426, 0.177031219, 0.177031219],
[0.1786672771, 0.1786672771, -0.7334594727, 0.1974578798, 0.1786672771]]]
expected_gradients = np.asarray(grads, dtype=np.float32)
return [activations, labels, activation_times, expected_costs, expected_gradients]
def setup_grad_case():
activations = np.asarray([[[0.1, 0.6, 0.1, 0.1, 0.1], [0.1, 0.1, 0.6, 0.1, 0.1]],
[[0.6, 0.1, 0.1, 0.1, 0.1], [0.1, 0.1, 0.5, 0.2, 0.1]]],
dtype=np.float32)
activation_times = np.asarray([2, 2], dtype=np.int32)
labels = np.asarray([[1, 2], [1, 2]], dtype=np.int32)
return [activations, labels, activation_times]
class TestCTC(unittest.TestCase):
"""
Test Baidu CTC wrapper implementation.
......@@ -50,56 +116,13 @@ class TestCTC(unittest.TestCase):
if isinstance(node.op, ConnectionistTemporalClassification):
assert (node.op.compute_grad is False)
# Test obtained from Torch tutorial at:
# https://github.com/baidu-research/warp-ctc/blob/master/torch_binding/TUTORIAL.md
def test_torch_case(self):
# Layout, from slowest to fastest changing dimension, is (time, batchSize, inputLayerSize)
activations = np.asarray([[[0, 0, 0, 0, 0], [1, 2, 3, 4, 5], [-5, -4, -3, -2, -1]],
[[0, 0, 0, 0, 0], [6, 7, 8, 9, 10], [-10, -9, -8, -7, -6]],
[[0, 0, 0, 0, 0], [11, 12, 13, 14, 15], [-15, -14, -13, -12, -11]]],
dtype=np.float32)
# Duration of each sequence
activation_times = np.asarray([1, 3, 3], dtype=np.int32)
# Labels for each sequence
labels = np.asarray([[1, -1],
[3, 3],
[2, 3]], dtype=np.int32)
expected_costs = np.asarray([1.609437943, 7.355742931, 4.938849926],
dtype=np.float32)
grads = [[[0.2, -0.8, 0.2, 0.2, 0.2],
[0.01165623125, 0.03168492019, 0.08612854034, -0.7658783197, 0.636408627],
[-0.02115798369, 0.03168492019, -0.8810571432, 0.2341216654, 0.636408627]],
[[0, 0, 0, 0, 0],
[-0.9883437753, 0.03168492019, 0.08612854034, 0.2341216654, 0.636408627],
[-0.02115798369, 0.03168492019, -0.1891518533, -0.4577836394, 0.636408627]],
[[0, 0, 0, 0, 0],
[0.01165623125, 0.03168492019, 0.08612854034, -0.7658783197, 0.636408627],
[-0.02115798369, 0.03168492019, 0.08612854034, -0.7330639958, 0.636408627]]]
expected_gradients = np.asarray(grads, dtype=np.float32)
self.run_ctc(activations, labels, activation_times, expected_costs, expected_gradients)
activations, labels, input_length, expected_costs, expected_grads = setup_torch_case()
self.run_ctc(activations, labels, input_length, expected_costs, expected_grads)
def test_ctc(self):
activations = np.asarray([[[0.1, 0.6, 0.1, 0.1, 0.1], [0.1, 0.1, 0.6, 0.1, 0.1]],
[[0.6, 0.1, 0.1, 0.1, 0.1], [0.1, 0.1, 0.5, 0.2, 0.1]]],
dtype=np.float32)
activation_times = np.asarray([2, 2], dtype=np.int32)
labels = np.asarray([[1, 2], [1, 2]], dtype=np.int32)
expected_costs = np.asarray([2.962858438, 3.053659201], dtype=np.float32)
grads = [[[0.177031219, -0.7081246376, 0.177031219, 0.177031219, 0.177031219],
[0.177031219, -0.8229685426, 0.291875124, 0.177031219, 0.177031219]],
[[0.291875124, 0.177031219, -0.8229685426, 0.177031219, 0.177031219],
[0.1786672771, 0.1786672771, -0.7334594727, 0.1974578798, 0.1786672771]]]
expected_gradients = np.asarray(grads, dtype=np.float32)
self.run_ctc(activations, labels, activation_times, expected_costs, expected_gradients)
activations, labels, input_length, expected_costs, expected_grads = setup_ctc_case()
self.run_ctc(activations, labels, input_length, expected_costs, expected_grads)
def test_verify_grad(self):
def ctc_op_functor(labels, in_lengths):
......@@ -110,13 +133,7 @@ class TestCTC(unittest.TestCase):
return ctc(acts, t_labels, t_activation_times)
return wrapper
activations = np.asarray([[[0.1, 0.6, 0.1, 0.1, 0.1], [0.1, 0.1, 0.6, 0.1, 0.1]],
[[0.6, 0.1, 0.1, 0.1, 0.1], [0.1, 0.1, 0.5, 0.2, 0.1]]],
dtype=np.float32)
activation_times = np.asarray([2, 2], dtype=np.int32)
labels = np.asarray([[1, 2], [1, 2]], dtype=np.int32)
activations, activation_times, labels = setup_grad_case()
ctc_op = ctc_op_functor(labels, activation_times)
......
Markdown 格式
0% 或
您添加了 0 人 到此讨论。请谨慎行事。
请先完成此评论的编辑！
请 注册 或者 后发表评论