我正在尝试将使用Python2.7和Keras 1.x的旧代码改编为Python3.7.3和Keras 2.2.4和TensorFlow 1.13.1。代码如下：
from keras.layers import Input, add, Dense, Flatten, concatenate
from keras import activations
from keras import models
from keras import backend as K
import numpy as np
import utils
from NGF.preprocessing import tensorise_smiles, tensorise_smiles_mp
from NGF.layers import NeuralGraphHidden, NeuralGraphOutput
from NGF.models import build_graph_conv_model
from NGF.sparse import GraphTensor, EpochIterator
# ==============================================================================
# ================================ Load the data ===============================
# ==============================================================================
print("{:=^100}".format(' Data preprocessing '))
data, labels = utils.load_delaney()
# Tensorise data
X_atoms, X_bonds, X_edges = tensorise_smiles_mp(data)
print('Atoms:', X_atoms.shape)
print('Bonds:', X_bonds.shape)
print('Edges:', X_edges.shape)
# Load sizes from data shape
num_molecules = X_atoms.shape[0]
max_atoms = X_atoms.shape[1]
max_degree = X_bonds.shape[2]
num_atom_features = X_atoms.shape[-1]
num_bond_features = X_bonds.shape[-1]
# ==============================================================================
# =============== Example 1: Building a 3-layer graph convnet =================
# ==============================================================================
print("{:=^100}".format(' Example 1 '))
# Parameters
conv_width = 8
fp_length = 62
# Define the input layers
atoms0 = Input(name='atom_inputs', shape=(max_atoms, num_atom_features))
bonds = Input(name='bond_inputs', shape=(max_atoms, max_degree, num_bond_features))
edges = Input(name='edge_inputs', shape=(max_atoms, max_degree), dtype='int32')
print("DEBUG: edges=", K.print_tensor(edges))
# Define the convoluted atom feature layers
atoms1 = NeuralGraphHidden(conv_width, activation='relu', use_bias=False)([atoms0, bonds, edges])
atoms2 = NeuralGraphHidden(conv_width, activation='relu', use_bias=False)([atoms1, bonds, edges])
# Define the outputs of each (convoluted) atom feature layer to fingerprint
fp_out0 = NeuralGraphOutput(fp_length, activation='softmax')([atoms0, bonds, edges])
fp_out1 = NeuralGraphOutput(fp_length, activation='softmax')([atoms1, bonds, edges])
fp_out2 = NeuralGraphOutput(fp_length, activation='softmax')([atoms2, bonds, edges])
# Flatten the input before the Dense layer by summing the 3 outputs to obtain fingerprint
# final_fp = merge([fp_out0, fp_out1, fp_out2], mode='sum') # Old Keras 1.x syntax
print("DEBUG: fp_out0.get_shape()=", fp_out0.get_shape())
print("DEBUG: fp_out1.get_shape()=", fp_out1.get_shape())
print("DEBUG: fp_out2.get_shape()=", fp_out2.get_shape())
# final_fp = add([fp_out0, fp_out1, fp_out2])
final_fp = concatenate([fp_out0, fp_out1, fp_out2])
print("DEBUG: final_fp.get_shape()=", final_fp.get_shape())
# Build and compile model for regression.
main_pred = Dense(1, activation='linear', name='main_prediction')(final_fp)
print("DEBUG: main_pred.get_shape()=", main_pred.get_shape())
model = models.Model(inputs=[atoms0, bonds, edges], outputs=[main_pred])
model.compile(optimizer='adagrad', loss='mse')
# Show summary
model.summary()
# Train the model
print("DEBUG: labels.shape", labels.shape)
model.fit(x=[X_atoms, X_bonds, X_edges], y=labels, epochs=20, batch_size=32, validation_split=0.2)
从本质上讲，这是一个定制的卷积神经网络，它以3个不同的可变维数组为输入，并返回标量预测。这是我执行它时的输出：
======================================== Data preprocessing ========================================
Tensorising molecules in batches...
1128/1128 [==================================================] - 1s 740us/step
Merging batch tensors... [DONE]
Atoms: (1128, 55, 62)
Bonds: (1128, 55, 5, 6)
Edges: (1128, 55, 5)
============================================ Example 1 =============================================
DEBUG: edges= Tensor("Print:0", shape=(?, 55, 5), dtype=int32)
DEBUG: fp_out0.get_shape()= (?, 62)
DEBUG: fp_out1.get_shape()= (?, 62)
DEBUG: fp_out2.get_shape()= (?, 62)
DEBUG: final_fp.get_shape()= (?, 186)
DEBUG: main_pred.get_shape()= (?, 1)
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
atom_inputs (InputLayer) (None, 55, 62) 0
__________________________________________________________________________________________________
bond_inputs (InputLayer) (None, 55, 5, 6) 0
__________________________________________________________________________________________________
edge_inputs (InputLayer) (None, 55, 5) 0
__________________________________________________________________________________________________
neural_graph_hidden_1 (NeuralGr [(None, 55, 62), (No 2720 atom_inputs[0][0]
bond_inputs[0][0]
edge_inputs[0][0]
__________________________________________________________________________________________________
neural_graph_hidden_2 (NeuralGr [(None, 55, 62), (No 2720 neural_graph_hidden_1[0][0]
bond_inputs[0][0]
edge_inputs[0][0]
__________________________________________________________________________________________________
neural_graph_output_1 (NeuralGr [(None, 55, 62), (No 4278 atom_inputs[0][0]
bond_inputs[0][0]
edge_inputs[0][0]
__________________________________________________________________________________________________
neural_graph_output_2 (NeuralGr [(None, 55, 62), (No 4278 neural_graph_hidden_1[0][0]
bond_inputs[0][0]
edge_inputs[0][0]
__________________________________________________________________________________________________
neural_graph_output_3 (NeuralGr [(None, 55, 62), (No 4278 neural_graph_hidden_2[0][0]
bond_inputs[0][0]
edge_inputs[0][0]
__________________________________________________________________________________________________
concatenate_1 (Concatenate) (None, 55, 186) 0 neural_graph_output_1[0][0]
neural_graph_output_2[0][0]
neural_graph_output_3[0][0]
__________________________________________________________________________________________________
main_prediction (Dense) (None, 55, 1) 187 concatenate_1[0][0]
==================================================================================================
Total params: 18,461
Trainable params: 18,461
Non-trainable params: 0
__________________________________________________________________________________________________
DEBUG: labels.shape (1128,)
Traceback (most recent call last):
File "/home/thomas/Programs/Anaconda3/lib/python3.7/site-packages/IPython/core/interactiveshell.py", line 3296, in run_code
exec(code_obj, self.user_global_ns, self.user_ns)
File "<ipython-input-2-9a41784534dc>", line 1, in <module>
runfile('/home2/thomas/Programs/keras-neural-graph-fingerprint_Py3/examples.py', wdir='/home2/thomas/Programs/keras-neural-graph-fingerprint_Py3')
File "/home2/thomas/Programs/pycharm-2019.1.1/helpers/pydev/_pydev_bundle/pydev_umd.py", line 197, in runfile
pydev_imports.execfile(filename, global_vars, local_vars) # execute the script
File "/home2/thomas/Programs/pycharm-2019.1.1/helpers/pydev/_pydev_imps/_pydev_execfile.py", line 18, in execfile
exec(compile(contents+"\n", file, 'exec'), glob, loc)
File "/home2/thomas/Programs/keras-neural-graph-fingerprint_Py3/examples.py", line 80, in <module>
model.fit(x=[X_atoms, X_bonds, X_edges], y=labels, epochs=20, batch_size=32, validation_split=0.2)
File "/home/thomas/Programs/Anaconda3/lib/python3.7/site-packages/keras/engine/training.py", line 952, in fit
batch_size=batch_size)
File "/home/thomas/Programs/Anaconda3/lib/python3.7/site-packages/keras/engine/training.py", line 789, in _standardize_user_data
exception_prefix='target')
File "/home/thomas/Programs/Anaconda3/lib/python3.7/site-packages/keras/engine/training_utils.py", line 128, in standardize_input_data
'with shape ' + str(data_shape))
ValueError: Error when checking target: expected main_prediction to have 3 dimensions, but got array with shape (1128, 1)
我怀疑此错误与“标签”数组的形状有关，该数组是平坦的。我究竟做错了什么？ 另外，我为什么得到
调试：final_fp.get_shape（）=（？，186）
但是model.summary（）显示
concatenate_1（串联）（无，55、186）0
这个额外维度（55）来自何处？也许由于某种原因，网络期望标签的尺寸为(1128, 55, 1)，而不是(1128, 1)。
如果您需要更多信息，请问我，我将添加更多调试打印功能。
答案 0 :(得分：0)
您的最后一个密集层main_predictions没有给出二维输出，因为您没有将其输入变平。
您需要在卷积层之后使用
Flatten层，以便Dense的输出是二维的。
main_predictions需要3D标签，但是您正在为其提供2D标签。因此，您会收到错误消息。
您可以在代码中添加Flatten层，例如：
flatten = Flatten()( final_fp )
main_pred = Dense(1, activation='linear', name='main_prediction')( flatten )
然后编译模型。
答案 1 :(得分：0)
在您的代码中，您提到的是以前使用final_fp = merge([fp_out0, fp_out1, fp_out2], mode='sum')的注释。因此，根据this，您将mode的参数用作sum而不是concat。您正在通过添加图层来合并图层。但是在新代码中，您正在使用final_fp = concatenate([fp_out0, fp_out1, fp_out2])。因此，您的旧代码和新代码之间是有区别的。在新代码中，您必须使用keras.layers.Add()才能获得相同的功能。
并且由于形状与标签数据和模型的输出节点不兼容而发生了值错误。尝试用keras.layers.Concatenate更改keras.layers.Add。
更新
final_fp = Input(shape=(55,62)) # I had used as input which can be replace with your final_fp
flatten_1 = Flatten()(final_fp)
main_pred = Dense(1, activation='linear', name='main_prediction')(flatten_1)
model = Model(inputs=final_fp,outputs=main_pred)
print(model.summary())
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) (None, 55, 62) 0
_________________________________________________________________
flatten_1 (Flatten) (None, 3410) 0
_________________________________________________________________
main_prediction (Dense) (None, 1) 3411
=================================================================
Total params: 3,411
Trainable params: 3,411
Non-trainable params: 0
_________________________________________________________________