AttributeError：“KerasTensor”对象没有属性“_keras_shape”

Question

嗨，我正在制作一个模型来训练数据集，但是在我的 resnet_model 中，我遇到了属性错误，所以请帮我解决这个错误。代码如下：

from Lib.data_loader import DataLoader
from Lib.resnet_model import Resnet3DBuilder
from Lib.HistoryGraph import HistoryGraph
import Lib.image as img
from Lib.utils import mkdirs
#import tensorflow as tf
import os

from math import ceil

from keras.optimizers import SGD

#from tensorflow.keras.optimizers import SGD

from keras.callbacks import ModelCheckpoint

#from tensorflow.keras.callbacks import ModelCheckpoint

target_size = (64,96)
nb_frames = 16  # here this will get number of pictres from datasets folder
skip = 1 # using resnet we skip different layers
nb_classes = 27
batch_size = 64 
input_shape = (nb_frames,) + target_size + (3,)

workers = 8 
use_multiprocessing = False 
max_queue_size = 20 



resnet_model = Resnet3DBuilder.build_resnet_101(input_shape, nb_classes, drop_rate = 0.5)

optimizer = SGD(lr=0.01, momentum=0.9, decay=0.0001, nesterov=False)

resnet_model.compile(optimizer = optimizer, loss= "categorical_crossentropy" , metrics=["accuracy"])

model_file = os.path.join(path_model, 'resnetmodel.hdf5')

model_checkpointer = ModelCheckpoint(model_file, monitor='val_acc',verbose=1, save_best_only=True, mode='max')

history_graph = HistoryGraph(model_path_name = os.path.join(path_model, "graphs"))

nb_sample_train = data.train_df["video_id"].size
nb_sample_val = data.val_df["video_id"].size

所有上面的代码都工作正常。 错误就在这里：

AttributeError                            Traceback (most recent call last)
<ipython-input-11-be30533fbfba> in <module>
----> 1 resnet_model = Resnet3DBuilder.build_resnet_101(input_shape, nb_classes, drop_rate = 0.5)
      2 
      3 optimizer = SGD(lr=0.01, momentum=0.9, decay=0.0001, nesterov=False)
      4 
      5 resnet_model.compile(optimizer = optimizer, loss= "categorical_crossentropy" , metrics=["accuracy"])

D:\HandGesturesProject\Lib\resnet_model.py in build_resnet_101(input_shape, num_outputs, reg_factor, drop_rate)
    258     def build_resnet_101(input_shape, num_outputs, reg_factor=1e-4, drop_rate=0):
    259         """Build resnet 101."""
--> 260         return Resnet3DBuilder.build(input_shape, num_outputs, bottleneck,
    261                                      [3, 4, 23, 3], reg_factor=reg_factor, drop_rate=drop_rate)

D:\HandGesturesProject\Lib\resnet_model.py in build(input_shape, num_outputs, block_fn, repetitions, reg_factor, drop_rate)
    223         filters = 64
    224         for i, r in enumerate(repetitions):
--> 225             block = _residual_block3d(block_fn, filters=filters,
    226                                       kernel_regularizer=l2(reg_factor),
    227                                       repetitions=r, is_first_layer=(i == 0)

D:\HandGesturesProject\Lib\resnet_model.py in f(input)
    105             if i == 0 and not is_first_layer:
    106                 strides = (2, 2, 2)
--> 107             input = block_function(filters=filters, strides=strides,
    108                                    kernel_regularizer=kernel_regularizer,
    109                                    is_first_block_of_first_layer=(

D:\HandGesturesProject\Lib\resnet_model.py in f(input)
    164                                    )(conv_3_3)
    165 
--> 166         return _shortcut3d(input, residual)
    167 
    168     return f

D:\HandGesturesProject\Lib\resnet_model.py in _shortcut3d(input, residual)
     76 
     77 def _shortcut3d(input, residual):
---> 78     stride_dim1 = input._keras_shape[DIM1_AXIS] \
     79         // residual._keras_shape[DIM1_AXIS]
     80     stride_dim2 = input._keras_shape[DIM2_AXIS] \

AttributeError: 'KerasTensor' 对象没有属性 '_keras_shape'

所以请帮我解决这个错误，即使我正在升级它的库，我也无法理解它。

Answer 1

属性_keras_shape 在张量流> 2.0 中不存在。要么升级你的 lib 使其与 > 2.0 兼容，要么设置一个虚拟环境来支持它。

用来测试的小程序

import tensorflow as tf
from keras.layers import Input

s = Input(shape=[1], dtype=tf.float32, name='1')
print("shape of s is: ",s._keras_shape) # (None, 1)

在tf.__version__ '2.3.1' 这不起作用并抛出AttributeError: 'Tensor' object has no attribute '_keras_shape'

解决方法，如果您还没有为 2.0 做好准备

创建一个虚拟环境：tensor1

virtualenv -p 2.7 tensor1

在张量和 keras 上安装旧版本

pip install tensorflow==1.15
pip install keras==2.2.4 

然后希望这个特定错误会消失，但您可能会遇到新的错误。