|
因为自己在这个坑里折腾了很长时间，才找到原因，所以希望能提示帮到大家。
先写结论
- 网络层必须定义在调用forward前定义完成（比如在__init__中定义），在forward中调用，才能打印出网络结构图和参数。
- 虽然，不能打印出网络的基础结构和参数信息，但不影响使用。
方式1，如下代码，打印不出网络结构和参数，参数数量都是0，但模型代码是可以正常训练和使用的。
import paddle
from paddle.nn import Linear
class S4(paddle.nn.Layer):
def __init__(self):
super().__init__()
def forward(self, inputs):
output_d = Linear(90,4)(inputs)
output_d = Linear(4,1)(output_d)
return output_d
model = S4()
params_info = paddle.summary(model, (2,90))
print(params_info)
运行耗时: 12毫秒
---------------------------------------------------------------------------
Layer (type) Input Shape Output Shape Param
===========================================================================
S [[2, 90]] [2, 1] 0
===========================================================================
Total params: 0
Trainable params: 0
Non-trainable params: 0
---------------------------------------------------------------------------
Input size (MB): 0.00
Forward/backward pass size (MB): 0.00
Params size (MB): 0.00
Estimated Total Size (MB): 0.00
---------------------------------------------------------------------------
{'total_params': 0, 'trainable_params': 0}
方式2，如下写的代码，可以打印出网络结构图和参数。在__init__中定义好各网络层。
import paddle
from paddle.nn import Linear
class S(paddle.nn.Layer):
def __init__(self):
super().__init__()
self.l1= Linear(90,4)
self.l2= Linear(4,1)
def forward(self, inputs):
output_d = self.l1(inputs)
output_d = self.l2(output_d)
return output_d
model = S()
params_info = paddle.summary(model, (2,90))
print(params_info)
---------------------------------------------------------------------------
Layer (type) Input Shape Output Shape Param
===========================================================================
Linear-128 [[2, 90]] [2, 4] 364
Linear-129 [[2, 4]] [2, 1] 5
===========================================================================
Total params: 369
Trainable params: 369
Non-trainable params: 0
---------------------------------------------------------------------------
Input size (MB): 0.00
Forward/backward pass size (MB): 0.00
Params size (MB): 0.00
Estimated Total Size (MB): 0.00
---------------------------------------------------------------------------
{'total_params': 369, 'trainable_params': 369}
深入summary的源码
- 主要通过model(模型的实例)的sublayers()方法，返回网络层的列表。
model.sublayers()
[Linear(in_features=90, out_features=4, dtype=float32),
Linear(in_features=4, out_features=1, dtype=float32)]
vars(model.sublayers()[0])
{'training': True,
'_full_name': 'linear_140',
'_helper': <paddle.fluid.dygraph.layer_object_helper.LayerObjectHelper at 0x7f65a26d0150>,
'_built': True,
'_dtype': 'float32',
'_init_in_dynamic_mode': True,
'_parameters': OrderedDict([('weight', Parameter containing:
Tensor(shape=[90, 4], dtype=float32, place=CPUPlace, stop_gradient=False,
[[-0.20621595, 0.08401936, -0.14570992, 0.11183006],
[-0.01728971, -0.13487765, -0.23390889, 0.24098989],
[-0.13978893, -0.20424819, -0.04855666, -0.16199186],
[-0.19478002, -0.13951692, 0.14115235, -0.14658877],
[ 0.15643370, 0.02453366, 0.19980761, -0.02117334],
[-0.04742610, -0.01464723, -0.15753657, 0.02658525],
[ 0.12675306, 0.21732125, 0.18672404, -0.06408432],
[ 0.16140866, 0.22746661, -0.16478351, -0.00010374],
[-0.22843692, -0.11098161, 0.03810865, -0.24758199],
[-0.11715272, 0.08546367, 0.18643376, 0.00941786],
[ 0.04385796, -0.13718715, -0.13291588, -0.12744799],
[ 0.20507479, 0.08363530, 0.08122411, 0.14772725],
[ 0.12475023, 0.10399523, 0.22499707, 0.04691744],
[ 0.16640592, 0.05567488, 0.25142434, 0.05884814],
[ 0.12723908, -0.21167704, -0.05146568, -0.07546510],
[-0.18740901, -0.13138834, 0.09158534, 0.11389682],
[-0.23955993, -0.04748055, 0.01112020, -0.07551017],
[-0.19948956, 0.07887903, -0.11123639, -0.24801619],
[ 0.17985007, -0.13626945, 0.03615150, -0.09616728],
[ 0.11204135, -0.13762784, -0.10694183, -0.24507296],
[ 0.14161596, -0.18264475, 0.21488816, -0.02595139],
[ 0.01281962, -0.23533244, -0.05832650, -0.18845016],
[ 0.01685080, 0.21017367, 0.04473603, -0.07200867],
[-0.16427593, 0.10090092, -0.08228512, 0.22937572],
[ 0.02535132, 0.16946691, 0.01132673, 0.22891155],
[-0.17698936, 0.14869818, 0.24249542, 0.19769028],
[ 0.01254764, 0.19422016, -0.07282487, 0.09596258],
[ 0.16854528, -0.06393975, 0.20858878, -0.06705859],
[ 0.18959373, -0.08939679, 0.07643029, -0.05605973],
[-0.12243937, -0.14132866, 0.19494379, -0.06451888],
[ 0.19235328, 0.10110179, -0.22526026, -0.21696852],
[-0.07262844, -0.01888850, 0.16245791, 0.22750211],
[-0.22919488, -0.18954048, 0.03602901, 0.21270525],
[-0.13936907, 0.12847206, 0.12586927, -0.16047193],
[-0.23191750, -0.13488717, 0.24885699, 0.08321038],
[-0.13570510, -0.07265873, -0.10631232, 0.09162298],
[-0.12932971, 0.18517035, -0.15347271, -0.02739608],
[ 0.22921789, -0.15096828, 0.23355103, 0.21406570],
[ 0.08929682, 0.15368345, -0.06598282, -0.13409325],
[-0.04817520, -0.10624386, -0.01299433, 0.16373715],
[ 0.06702435, -0.01464282, 0.15239137, -0.06154861],
[-0.08851637, -0.15216920, 0.14518383, 0.03232276],
[-0.00448877, 0.10262579, -0.24696153, 0.06273824],
[ 0.19433418, -0.01608880, -0.06958365, 0.21542296],
[ 0.17663121, 0.24140605, -0.16102412, -0.13844144],
[-0.25249654, -0.11238430, -0.10917042, 0.22930911],
[ 0.18360525, -0.03554909, -0.20884866, 0.10487282],
[-0.24110393, 0.12529787, -0.00518186, 0.09543869],
[ 0.19364417, 0.18657467, -0.05774385, 0.06694761],
[ 0.06235594, 0.18447849, -0.23734707, 0.02031776],
[-0.07527824, 0.06170449, 0.02800348, -0.12043267],
[-0.06529415, 0.02134055, -0.18275535, -0.13008605],
[-0.20935142, 0.05829439, -0.04963452, 0.24142167],
[-0.00459233, -0.17473610, -0.12816842, -0.05939011],
[ 0.25015548, 0.10697207, 0.24913797, 0.23499483],
[-0.09882846, -0.07380529, -0.00138485, -0.05095109],
[ 0.24390355, 0.14776850, -0.23272485, -0.14770921],
[ 0.10985476, 0.04561046, 0.12682661, -0.18175307],
[-0.07764107, -0.23298098, 0.11743194, 0.03062549],
[-0.16420993, 0.06812567, -0.22719657, -0.13849448],
[-0.20762315, 0.15431166, -0.24009264, -0.04427037],
[-0.15257362, -0.11478184, 0.21630144, 0.01226106],
[ 0.20274323, 0.01106474, 0.17747423, 0.19533655],
[-0.09216593, 0.21642286, 0.16395739, 0.18711695],
[-0.16654512, -0.01774178, -0.21099040, -0.15609533],
[ 0.16119608, 0.16110554, 0.22074458, -0.16093451],
[ 0.13975376, -0.13630739, -0.03922784, 0.11074531],
[ 0.17509770, 0.00062934, -0.06153570, 0.14687642],
[ 0.22287810, 0.18240044, 0.14542899, -0.20445868],
[-0.21342079, 0.20837983, -0.05815056, -0.03875078],
[-0.24014582, -0.15421730, -0.12463945, -0.11398430],
[ 0.09004262, -0.02769084, -0.19755839, -0.10233013],
[ 0.19232154, 0.04774907, 0.18493724, 0.25028470],
[-0.12067989, -0.15001965, -0.02605258, 0.22234175],
[-0.00589231, -0.15065691, 0.07578629, -0.23713773],
[ 0.24967501, -0.00311911, -0.11508700, -0.11503294],
[-0.05981220, 0.11477354, -0.03898008, -0.11416386],
[ 0.12355697, 0.00971353, 0.03545964, 0.03065020],
[-0.22309546, -0.07022245, 0.04698494, 0.06996807],
[ 0.00341684, -0.04152401, -0.04922824, -0.19472414],
[-0.07824557, -0.24881583, -0.02940276, -0.09540446],
[-0.07309537, -0.19851863, 0.18788418, -0.14266992],
[ 0.19145533, 0.18813542, -0.13308662, 0.23938963],
[ 0.19193968, 0.15658599, 0.11875460, -0.12196723],
[-0.06095690, 0.14294901, 0.10374212, -0.08170357],
[ 0.15833840, -0.15739962, -0.09807962, 0.09703702],
[-0.15394495, 0.05613059, 0.21851090, 0.02185619],
[ 0.17367855, 0.13805395, 0.12828121, -0.12902705],
[-0.05945474, 0.09741652, 0.21674320, 0.06598496],
[-0.05226740, -0.21138683, 0.21320057, 0.14421508]])),
('bias', Parameter containing:
Tensor(shape=[4], dtype=float32, place=CPUPlace, stop_gradient=False,
[0., 0., 0., 0.]))]),
'_buffers': OrderedDict(),
'_non_persistable_buffer_names_set': set(),
'_sub_layers': OrderedDict(),
'_loaddict_holder': OrderedDict(),
'_forward_pre_hooks': OrderedDict(),
'_forward_post_hooks': OrderedDict(),
'_weight_attr': None,
'_bias_attr': None,
'name': None}
方式1的问题，在于sublayers()返回的列表是空列表。
- sublayers()的源码核心是遍历named_sublayers
- named_sublayers的源码是用类内部属性_sub_layers获取所有的子网络的列表。
例如：
model._sub_layers
OrderedDict([('l1', Linear(in_features=90, out_features=4, dtype=float32)),
('l2', Linear(in_features=4, out_features=1, dtype=float32))])
Examplex:
.. code-block:: python
import paddle
import numpy as np
from collections import OrderedDict
sublayers = OrderedDict([
('conv1d', paddle.nn.Conv1D(3, 2, 3)),
('conv2d', paddle.nn.Conv2D(3, 2, 3)),
('conv3d', paddle.nn.Conv3D(4, 6, (3, 3, 3))),
])
layers_dict = paddle.nn.LayerDict(sublayers=sublayers)
l = layers_dict['conv1d']
for k in layers_dict:
l = layers_dict[k]
len(layers_dict)
del layers_dict['conv2d']
len(layers_dict)
conv1d = layers_dict.pop('conv1d')
len(layers_dict)
layers_dict.clear()
len(layers_dict)
解决方法（只是可行，不一定是优化）
- 生成网络顺序列表。
- 用paddle.nn.LayerList(把1打包成paddle的层列表。
- 在forward中顺序调用。
其中1，2的步骤类似于paddle.nn.Sequential的作用。只不过，可以更加灵活，比如在网络层之间进行特殊的张量操作，而Sequential不行。
例如：
import paddle
from paddle.nn import Linear, BatchNorm, ReLU
class S(paddle.nn.Layer):
def __init__(self):
super().__init__()
self.layer_list = paddle.nn.LayerList(self.make_list())
def linear(self,input_dim, output_dim):
return Linear(input_dim, output_dim)
def linear_g(self, input_dim, output_dim):
l_list = []
l = self.linear(input_dim, output_dim)
l_list.append(l)
l = BatchNorm(l.weight.shape[1])
l_list.append(l)
l_list.append(ReLU())
return l_list
def make_list(self):
l_list = []
l_list.extend(self.linear_g(90,4))
return l_list
def forward(self, x):
for i, l in enumerate(self.layer_list):
x = self.layer_list[i](x)
return x
model = S()
params_info = paddle.summary(model, (1,90))
print(params_info)
print(list(model.named_sublayers()))
print(model.named_children())
---------------------------------------------------------------------------
Layer (type) Input Shape Output Shape Param
===========================================================================
Linear-231 [[1, 90]] [1, 4] 364
BatchNorm-22 [[1, 4]] [1, 4] 16
ReLU-41 [[1, 4]] [1, 4] 0
===========================================================================
Total params: 380
Trainable params: 364
Non-trainable params: 16
---------------------------------------------------------------------------
Input size (MB): 0.00
Forward/backward pass size (MB): 0.00
Params size (MB): 0.00
Estimated Total Size (MB): 0.00
---------------------------------------------------------------------------
{'total_params': 380, 'trainable_params': 364}
[('layer_list', LayerList(
(0): Linear(in_features=90, out_features=4, dtype=float32)
(1): BatchNorm()
(2): ReLU()
)), ('layer_list.0', Linear(in_features=90, out_features=4, dtype=float32)), ('layer_list.1', BatchNorm()), ('layer_list.2', ReLU())]
<generator object Layer.named_children at 0x7f65a427dc50>
定义网络，如果能知道前一层的输出形状？
可以查看各自预定层的源码，例如：paddle.nn.Linear的源码中，
def extra_repr(self):
name_str = ', name={}'.format(self.name) if self.name else ''
return 'in_features={}, out_features={}, dtype={}{}'.format(
self.weight.shape[0], self.weight.shape[1], self._dtype, name_str)
可以看出，本层的输出性质是在self.weight.shape[1]中。
其他的层类似，查看extra_repr函数中的内容。
|