You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
Paddle/python/paddle/fluid/tests/unittests/rnn/test_rnn_nets.py

270 lines
9.6 KiB

# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import paddle
paddle.set_default_dtype("float64")
from paddle.fluid.layers import sequence_mask
import numpy as np
import unittest
from convert import convert_params_for_net
from rnn_numpy import SimpleRNN, LSTM, GRU
class TestSimpleRNN(unittest.TestCase):
def __init__(self, time_major=True, direction="forward", place="cpu"):
super(TestSimpleRNN, self).__init__("runTest")
self.time_major = time_major
self.direction = direction
self.num_directions = 2 if direction == "bidirectional" else 1
self.place = paddle.CPUPlace() if place == "cpu" \
else paddle.CUDAPlace(0)
def setUp(self):
paddle.disable_static(self.place)
rnn1 = SimpleRNN(
16, 32, 2, time_major=self.time_major, direction=self.direction)
rnn2 = paddle.nn.SimpleRNN(
16, 32, 2, time_major=self.time_major, direction=self.direction)
convert_params_for_net(rnn1, rnn2)
self.rnn1 = rnn1
self.rnn2 = rnn2
def test_with_initial_state(self):
rnn1 = self.rnn1
rnn2 = self.rnn2
x = np.random.randn(12, 4, 16)
if not self.time_major:
x = np.transpose(x, [1, 0, 2])
prev_h = np.random.randn(2 * self.num_directions, 4, 32)
y1, h1 = rnn1(x, prev_h)
y2, h2 = rnn2(paddle.to_variable(x), paddle.to_variable(prev_h))
np.testing.assert_allclose(y1, y2.numpy(), atol=1e-8, rtol=1e-5)
np.testing.assert_allclose(h1, h2.numpy(), atol=1e-8, rtol=1e-5)
def test_with_zero_state(self):
rnn1 = self.rnn1
rnn2 = self.rnn2
x = np.random.randn(12, 4, 16)
if not self.time_major:
x = np.transpose(x, [1, 0, 2])
y1, h1 = rnn1(x)
y2, h2 = rnn2(paddle.to_variable(x))
np.testing.assert_allclose(y1, y2.numpy(), atol=1e-8, rtol=1e-5)
np.testing.assert_allclose(h1, h2.numpy(), atol=1e-8, rtol=1e-5)
def test_with_input_lengths(self):
rnn1 = self.rnn1
rnn2 = self.rnn2
x = np.random.randn(12, 4, 16)
if not self.time_major:
x = np.transpose(x, [1, 0, 2])
sequence_length = np.array([12, 10, 9, 8], dtype=np.int64)
y1, h1 = rnn1(x, sequence_length=sequence_length)
seq_len = paddle.to_variable(sequence_length)
mask = sequence_mask(seq_len, dtype=paddle.get_default_dtype())
if self.time_major:
mask = paddle.transpose(mask, [1, 0])
y2, h2 = rnn2(paddle.to_variable(x), sequence_length=seq_len)
y2 = paddle.multiply(y2, mask, axis=0)
np.testing.assert_allclose(y1, y2.numpy(), atol=1e-8, rtol=1e-5)
np.testing.assert_allclose(h1, h2.numpy(), atol=1e-8, rtol=1e-5)
def runTest(self):
self.test_with_initial_state()
self.test_with_zero_state()
self.test_with_input_lengths()
class TestGRU(unittest.TestCase):
def __init__(self, time_major=True, direction="forward", place="cpu"):
super(TestGRU, self).__init__("runTest")
self.time_major = time_major
self.direction = direction
self.num_directions = 2 if direction == "bidirectional" else 1
self.place = paddle.CPUPlace() if place == "cpu" \
else paddle.CUDAPlace(0)
def setUp(self):
paddle.disable_static(self.place)
rnn1 = GRU(16,
32,
2,
time_major=self.time_major,
direction=self.direction)
rnn2 = paddle.nn.GRU(16,
32,
2,
time_major=self.time_major,
direction=self.direction)
convert_params_for_net(rnn1, rnn2)
self.rnn1 = rnn1
self.rnn2 = rnn2
def test_with_initial_state(self):
rnn1 = self.rnn1
rnn2 = self.rnn2
x = np.random.randn(12, 4, 16)
if not self.time_major:
x = np.transpose(x, [1, 0, 2])
prev_h = np.random.randn(2 * self.num_directions, 4, 32)
y1, h1 = rnn1(x, prev_h)
y2, h2 = rnn2(paddle.to_variable(x), paddle.to_variable(prev_h))
np.testing.assert_allclose(y1, y2.numpy(), atol=1e-8, rtol=1e-5)
np.testing.assert_allclose(h1, h2.numpy(), atol=1e-8, rtol=1e-5)
def test_with_zero_state(self):
rnn1 = self.rnn1
rnn2 = self.rnn2
x = np.random.randn(12, 4, 16)
if not self.time_major:
x = np.transpose(x, [1, 0, 2])
y1, h1 = rnn1(x)
y2, h2 = rnn2(paddle.to_variable(x))
np.testing.assert_allclose(y1, y2.numpy(), atol=1e-8, rtol=1e-5)
np.testing.assert_allclose(h1, h2.numpy(), atol=1e-8, rtol=1e-5)
def test_with_input_lengths(self):
rnn1 = self.rnn1
rnn2 = self.rnn2
x = np.random.randn(12, 4, 16)
if not self.time_major:
x = np.transpose(x, [1, 0, 2])
sequence_length = np.array([12, 10, 9, 8], dtype=np.int64)
y1, h1 = rnn1(x, sequence_length=sequence_length)
seq_len = paddle.to_variable(sequence_length)
mask = sequence_mask(seq_len, dtype=paddle.get_default_dtype())
if self.time_major:
mask = paddle.transpose(mask, [1, 0])
y2, h2 = rnn2(paddle.to_variable(x), sequence_length=seq_len)
y2 = paddle.multiply(y2, mask, axis=0)
np.testing.assert_allclose(y1, y2.numpy(), atol=1e-8, rtol=1e-5)
np.testing.assert_allclose(h1, h2.numpy(), atol=1e-8, rtol=1e-5)
def runTest(self):
self.test_with_initial_state()
self.test_with_zero_state()
self.test_with_input_lengths()
class TestLSTM(unittest.TestCase):
def __init__(self, time_major=True, direction="forward", place="cpu"):
super(TestLSTM, self).__init__("runTest")
self.time_major = time_major
self.direction = direction
self.num_directions = 2 if direction == "bidirectional" else 1
self.place = paddle.CPUPlace() if place == "cpu" \
else paddle.CUDAPlace(0)
def setUp(self):
paddle.disable_static(self.place)
rnn1 = LSTM(
16, 32, 2, time_major=self.time_major, direction=self.direction)
rnn2 = paddle.nn.LSTM(
16, 32, 2, time_major=self.time_major, direction=self.direction)
convert_params_for_net(rnn1, rnn2)
self.rnn1 = rnn1
self.rnn2 = rnn2
def test_with_initial_state(self):
rnn1 = self.rnn1
rnn2 = self.rnn2
x = np.random.randn(12, 4, 16)
if not self.time_major:
x = np.transpose(x, [1, 0, 2])
prev_h = np.random.randn(2 * self.num_directions, 4, 32)
prev_c = np.random.randn(2 * self.num_directions, 4, 32)
y1, (h1, c1) = rnn1(x, (prev_h, prev_c))
y2, (h2, c2) = rnn2(
paddle.to_variable(x),
(paddle.to_variable(prev_h), paddle.to_variable(prev_c)))
np.testing.assert_allclose(y1, y2.numpy(), atol=1e-8, rtol=1e-5)
np.testing.assert_allclose(h1, h2.numpy(), atol=1e-8, rtol=1e-5)
np.testing.assert_allclose(c1, c2.numpy(), atol=1e-8, rtol=1e-5)
def test_with_zero_state(self):
rnn1 = self.rnn1
rnn2 = self.rnn2
x = np.random.randn(12, 4, 16)
if not self.time_major:
x = np.transpose(x, [1, 0, 2])
y1, (h1, c1) = rnn1(x)
y2, (h2, c2) = rnn2(paddle.to_variable(x))
np.testing.assert_allclose(y1, y2.numpy(), atol=1e-8, rtol=1e-5)
np.testing.assert_allclose(h1, h2.numpy(), atol=1e-8, rtol=1e-5)
np.testing.assert_allclose(c1, c2.numpy(), atol=1e-8, rtol=1e-5)
def test_with_input_lengths(self):
rnn1 = self.rnn1
rnn2 = self.rnn2
x = np.random.randn(12, 4, 16)
if not self.time_major:
x = np.transpose(x, [1, 0, 2])
sequence_length = np.array([12, 10, 9, 8], dtype=np.int64)
y1, (h1, c1) = rnn1(x, sequence_length=sequence_length)
seq_len = paddle.to_variable(sequence_length)
mask = sequence_mask(seq_len, dtype=paddle.get_default_dtype())
if self.time_major:
mask = paddle.transpose(mask, [1, 0])
y2, (h2, c2) = rnn2(paddle.to_variable(x), sequence_length=seq_len)
y2 = paddle.multiply(y2, mask, axis=0)
np.testing.assert_allclose(y1, y2.numpy(), atol=1e-8, rtol=1e-5)
np.testing.assert_allclose(h1, h2.numpy(), atol=1e-8, rtol=1e-5)
np.testing.assert_allclose(c1, c2.numpy(), atol=1e-8, rtol=1e-5)
def runTest(self):
self.test_with_initial_state()
self.test_with_zero_state()
self.test_with_input_lengths()
def load_tests(loader, tests, pattern):
suite = unittest.TestSuite()
devices = ["cpu", "gpu"] if paddle.fluid.is_compiled_with_cuda() \
else ["cpu"]
for direction in ["forward", "backward", "bidirectional"]:
for time_major in [True, False]:
for device in devices:
for test_class in [TestSimpleRNN, TestLSTM, TestGRU]:
suite.addTest(test_class(time_major, direction, device))
return suite