wehub-resource-sync
387 lines
13 KiB
Python
387 lines
13 KiB
Python
# Copyright (c) 2024 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 unittest
|
|
from collections import OrderedDict
|
|
|
|
import numpy as np
|
|
from op_test import get_device_place
|
|
|
|
import paddle
|
|
from paddle import base
|
|
|
|
|
|
def _make_param(shape, dtype='float32'):
|
|
return paddle.create_parameter(shape=shape, dtype=dtype)
|
|
|
|
|
|
class MyLayer(paddle.nn.Layer):
|
|
def __init__(self, num_stacked_param):
|
|
super().__init__()
|
|
# create ParameterDict with iterable Parameters
|
|
self.params = self.paddle_imperative_ParameterDict(num_stacked_param)
|
|
|
|
def paddle_imperative_ParameterDict(self, num_stacked_param):
|
|
return paddle.nn.ParameterDict(
|
|
[
|
|
(
|
|
't' + str(i),
|
|
paddle.create_parameter(shape=[2, 2], dtype='float32'),
|
|
)
|
|
for i in range(num_stacked_param)
|
|
]
|
|
)
|
|
|
|
def forward(self, x):
|
|
for i, key in enumerate(self.params):
|
|
x = paddle.matmul(x, self.params[key])
|
|
return x
|
|
|
|
|
|
class TestImperativeContainerParameterDict(unittest.TestCase):
|
|
"""Original test: basic forward/backward with list-of-tuples init and update."""
|
|
|
|
def parameter_dict(self):
|
|
self.place = get_device_place()
|
|
data_np = np.random.uniform(-1, 1, [5, 2]).astype('float32')
|
|
with base.dygraph.guard():
|
|
x = paddle.to_tensor(data_np)
|
|
num_stacked_param = 4
|
|
model = MyLayer(num_stacked_param)
|
|
self.assertEqual(len(model.params), num_stacked_param)
|
|
res = model(x)
|
|
self.assertListEqual(res.shape, [5, 2])
|
|
loss = paddle.mean(res)
|
|
loss.backward()
|
|
|
|
model.params['t' + str(num_stacked_param - 1)] = (
|
|
paddle.create_parameter(shape=[2, 3], dtype='float32')
|
|
)
|
|
res = model(x)
|
|
self.assertListEqual(res.shape, [5, 3])
|
|
parameter = OrderedDict(
|
|
[
|
|
(
|
|
't' + str(num_stacked_param),
|
|
paddle.create_parameter(shape=[3, 4], dtype='float32'),
|
|
)
|
|
]
|
|
)
|
|
model.params.update(parameter)
|
|
self.assertEqual(len(model.params), num_stacked_param + 1)
|
|
res = model(x)
|
|
self.assertListEqual(res.shape, [5, 4])
|
|
loss = paddle.mean(res)
|
|
loss.backward()
|
|
|
|
def test_parameter_dict(self):
|
|
self.parameter_dict()
|
|
|
|
|
|
class TestParameterDictInit(unittest.TestCase):
|
|
def test_init_types(self):
|
|
# None, plain dict, OrderedDict, list of tuples
|
|
self.assertEqual(len(paddle.nn.ParameterDict()), 0)
|
|
self.assertEqual(
|
|
len(paddle.nn.ParameterDict({'w': _make_param([2, 3])})), 1
|
|
)
|
|
self.assertEqual(
|
|
len(
|
|
paddle.nn.ParameterDict(
|
|
OrderedDict([('w', _make_param([2, 3]))])
|
|
)
|
|
),
|
|
1,
|
|
)
|
|
self.assertEqual(
|
|
len(paddle.nn.ParameterDict([('w', _make_param([2, 3]))])), 1
|
|
)
|
|
|
|
def test_init_with_parameter_dict(self):
|
|
# ParameterDict as input — exercises the update() fix
|
|
pd1 = paddle.nn.ParameterDict({'w': _make_param([2, 3])})
|
|
pd2 = paddle.nn.ParameterDict(pd1)
|
|
self.assertEqual(len(pd2), 1)
|
|
|
|
def test_init_with_values_alias(self):
|
|
# @param_one_alias: 'values' maps to 'parameters'
|
|
pd = paddle.nn.ParameterDict(values={'w': _make_param([2, 3])})
|
|
self.assertEqual(len(pd), 1)
|
|
|
|
def test_init_preserves_order(self):
|
|
keys_in = ['c', 'a', 'b']
|
|
pd = paddle.nn.ParameterDict(
|
|
OrderedDict([(k, _make_param([1, 2])) for k in keys_in])
|
|
)
|
|
self.assertEqual(list(pd), keys_in)
|
|
|
|
def test_init_errors(self):
|
|
with self.assertRaises((ValueError, TypeError)):
|
|
paddle.nn.ParameterDict([('w', _make_param([2, 3]), 'extra')])
|
|
with self.assertRaises((AssertionError, TypeError)):
|
|
paddle.nn.ParameterDict(42)
|
|
|
|
|
|
class TestParameterDictAccess(unittest.TestCase):
|
|
def setUp(self):
|
|
self.pd = paddle.nn.ParameterDict(
|
|
{'w1': _make_param([2, 3]), 'w2': _make_param([3, 4])}
|
|
)
|
|
|
|
def test_getitem(self):
|
|
self.assertEqual(list(self.pd['w1'].shape), [2, 3])
|
|
self.assertEqual(list(self.pd['w2'].shape), [3, 4])
|
|
|
|
def test_setitem(self):
|
|
self.pd['w1'] = _make_param([2, 5]) # replace
|
|
self.assertEqual(list(self.pd['w1'].shape), [2, 5])
|
|
self.pd['w3'] = _make_param([4, 5]) # add new
|
|
self.assertEqual(len(self.pd), 3)
|
|
|
|
def test_setitem_non_parameter_raises(self):
|
|
with self.assertRaises((AssertionError, TypeError)):
|
|
self.pd['bad'] = paddle.to_tensor([1.0, 2.0])
|
|
|
|
def test_len_iter_contains(self):
|
|
self.assertEqual(len(self.pd), 2)
|
|
self.assertEqual(sorted(self.pd), ['w1', 'w2'])
|
|
self.assertIn('w1', self.pd)
|
|
self.assertNotIn('missing', self.pd)
|
|
|
|
|
|
class TestParameterDictUpdate(unittest.TestCase):
|
|
def setUp(self):
|
|
self.pd = paddle.nn.ParameterDict({'w1': _make_param([2, 3])})
|
|
|
|
def test_update_input_types(self):
|
|
# plain dict, OrderedDict, list of tuples
|
|
self.pd.update({'w2': _make_param([3, 4])})
|
|
self.assertEqual(len(self.pd), 2)
|
|
self.pd.update(OrderedDict([('w3', _make_param([4, 5]))]))
|
|
self.assertEqual(len(self.pd), 3)
|
|
self.pd.update([('w4', _make_param([5, 6]))])
|
|
self.assertEqual(len(self.pd), 4)
|
|
|
|
def test_update_from_parameter_dict(self):
|
|
# ParameterDict as input — exercises the update() fix
|
|
other = paddle.nn.ParameterDict({'w2': _make_param([3, 4])})
|
|
self.pd.update(other)
|
|
self.assertEqual(len(self.pd), 2)
|
|
|
|
def test_update_overwrites(self):
|
|
self.pd.update({'w1': _make_param([2, 5])})
|
|
self.assertEqual(list(self.pd['w1'].shape), [2, 5])
|
|
|
|
def test_update_errors(self):
|
|
with self.assertRaises((ValueError, TypeError)):
|
|
self.pd.update([('w2', _make_param([3, 4]), 'extra')])
|
|
with self.assertRaises((AssertionError, TypeError)):
|
|
self.pd.update(42)
|
|
|
|
|
|
class TestParameterDictRegistration(unittest.TestCase):
|
|
def _make_model(self):
|
|
class M(paddle.nn.Layer):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.pd = paddle.nn.ParameterDict(
|
|
{'w1': _make_param([2, 3]), 'w2': _make_param([3, 4])}
|
|
)
|
|
|
|
def forward(self, x):
|
|
return paddle.matmul(
|
|
paddle.matmul(x, self.pd['w1']), self.pd['w2']
|
|
)
|
|
|
|
return M()
|
|
|
|
def test_registered_in_parameters_named_state_dict(self):
|
|
model = self._make_model()
|
|
self.assertEqual(len(list(model.parameters())), 2)
|
|
named = dict(model.named_parameters())
|
|
self.assertIn('pd.w1', named)
|
|
self.assertIn('pd.w2', named)
|
|
state = model.state_dict()
|
|
self.assertIn('pd.w1', state)
|
|
self.assertIn('pd.w2', state)
|
|
|
|
def test_gradient_flows(self):
|
|
model = self._make_model()
|
|
paddle.matmul(paddle.uniform([2, 2]), model.pd['w1']).sum().backward()
|
|
self.assertIsNotNone(model.pd['w1'].grad)
|
|
|
|
def test_dynamic_setitem_and_update_registered(self):
|
|
class M(paddle.nn.Layer):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.pd = paddle.nn.ParameterDict()
|
|
|
|
def forward(self, x):
|
|
return x
|
|
|
|
model = M()
|
|
model.pd['w'] = _make_param([2, 2])
|
|
model.pd.update({'v': _make_param([2, 2])})
|
|
self.assertEqual(len(list(model.parameters())), 2)
|
|
named = dict(model.named_parameters())
|
|
self.assertIn('pd.w', named)
|
|
self.assertIn('pd.v', named)
|
|
|
|
|
|
class TestParameterDictForwardBackward(unittest.TestCase):
|
|
def _chain_model(self, n):
|
|
class M(paddle.nn.Layer):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.pd = paddle.nn.ParameterDict(
|
|
{f't{i}': _make_param([2, 2]) for i in range(n)}
|
|
)
|
|
|
|
def forward(self, x):
|
|
for key in self.pd:
|
|
x = paddle.matmul(x, self.pd[key])
|
|
return x
|
|
|
|
return M()
|
|
|
|
def test_forward_and_backward(self):
|
|
model = self._chain_model(3)
|
|
x = paddle.uniform([5, 2])
|
|
out = model(x)
|
|
self.assertEqual(list(out.shape), [5, 2])
|
|
paddle.mean(out).backward()
|
|
for key in model.pd:
|
|
self.assertIsNotNone(model.pd[key].grad)
|
|
|
|
def test_replace_param_changes_output_shape(self):
|
|
model = self._chain_model(2)
|
|
x = paddle.uniform([3, 2])
|
|
self.assertEqual(list(model(x).shape), [3, 2])
|
|
model.pd['t1'] = _make_param([2, 5])
|
|
self.assertEqual(list(model(x).shape), [3, 5])
|
|
|
|
def test_float64_params(self):
|
|
pd = paddle.nn.ParameterDict(
|
|
{'w': paddle.create_parameter(shape=[2, 3], dtype='float64')}
|
|
)
|
|
out = paddle.matmul(paddle.uniform([1, 2], dtype='float64'), pd['w'])
|
|
self.assertEqual(list(out.shape), [1, 3])
|
|
self.assertEqual(out.dtype, paddle.float64)
|
|
|
|
|
|
class TestParameterDictPopKeysValues(unittest.TestCase):
|
|
def setUp(self):
|
|
self.pd = paddle.nn.ParameterDict(
|
|
{
|
|
'w1': _make_param([2, 3]),
|
|
'w2': _make_param([3, 4]),
|
|
'w3': _make_param([4, 5]),
|
|
}
|
|
)
|
|
|
|
def test_pop_returns_correct_param(self):
|
|
p = self.pd.pop('w2')
|
|
self.assertEqual(list(p.shape), [3, 4])
|
|
self.assertEqual(len(self.pd), 2)
|
|
self.assertNotIn('w2', list(self.pd.keys()))
|
|
|
|
def test_pop_missing_key_raises(self):
|
|
with self.assertRaises(KeyError):
|
|
self.pd.pop('nonexistent')
|
|
|
|
def test_pop_all_items_leaves_empty(self):
|
|
for k in list(self.pd.keys()):
|
|
self.pd.pop(k)
|
|
self.assertEqual(len(self.pd), 0)
|
|
|
|
def test_keys_returns_all_in_order(self):
|
|
self.assertEqual(list(self.pd.keys()), ['w1', 'w2', 'w3'])
|
|
|
|
def test_keys_after_update(self):
|
|
self.pd.update({'w4': _make_param([5, 6])})
|
|
self.assertIn('w4', list(self.pd.keys()))
|
|
self.assertEqual(len(self.pd), 4)
|
|
|
|
def test_values_shapes(self):
|
|
shapes = [list(v.shape) for v in self.pd.values()]
|
|
self.assertEqual(shapes, [[2, 3], [3, 4], [4, 5]])
|
|
|
|
def test_values_are_parameters(self):
|
|
from paddle.base.framework import Parameter
|
|
|
|
for v in self.pd.values():
|
|
self.assertIsInstance(v, Parameter)
|
|
|
|
def test_values_count_matches_len(self):
|
|
self.assertEqual(len(self.pd.values()), len(self.pd))
|
|
|
|
def test_pop_reduces_values(self):
|
|
self.pd.pop('w1')
|
|
shapes = [list(v.shape) for v in self.pd.values()]
|
|
self.assertEqual(shapes, [[3, 4], [4, 5]])
|
|
|
|
|
|
class TestParameterDictStateDictRoundtrip(unittest.TestCase):
|
|
def _make_model(self):
|
|
class M(paddle.nn.Layer):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.pd = paddle.nn.ParameterDict(
|
|
{
|
|
'w1': _make_param([2, 3]),
|
|
'w2': _make_param([3, 4]),
|
|
}
|
|
)
|
|
|
|
def forward(self, x):
|
|
return paddle.matmul(
|
|
paddle.matmul(x, self.pd['w1']), self.pd['w2']
|
|
)
|
|
|
|
return M()
|
|
|
|
def test_state_dict_roundtrip_values(self):
|
|
model_a = self._make_model()
|
|
state_a = model_a.state_dict()
|
|
w1_a = state_a['pd.w1'].numpy().copy()
|
|
w2_a = state_a['pd.w2'].numpy().copy()
|
|
|
|
model_b = self._make_model()
|
|
model_b.set_state_dict(state_a)
|
|
state_b = model_b.state_dict()
|
|
|
|
np.testing.assert_array_equal(state_b['pd.w1'].numpy(), w1_a)
|
|
np.testing.assert_array_equal(state_b['pd.w2'].numpy(), w2_a)
|
|
|
|
def test_output_matches_after_load(self):
|
|
model_a = self._make_model()
|
|
model_b = self._make_model()
|
|
model_b.set_state_dict(model_a.state_dict())
|
|
|
|
x = paddle.uniform([2, 2])
|
|
np.testing.assert_array_almost_equal(
|
|
model_a(x).numpy(), model_b(x).numpy()
|
|
)
|
|
|
|
def test_state_dict_keys_present(self):
|
|
model = self._make_model()
|
|
state = model.state_dict()
|
|
self.assertIn('pd.w1', state)
|
|
self.assertIn('pd.w2', state)
|
|
self.assertEqual(len(state), 2)
|
|
|
|
|
|
if __name__ == '__main__':
|
|
unittest.main()
|