如何使用wights预定义/训练（hdf5）文件来预测一类新的eeg数据？

Question

我有一个名为（bestmodel.hdf5）的预定义文件，它是使用Keras库（python）和theano创建的。使用以下代码的训练模型。

# set parameters 
batch_size = 1280 
nb_epoch = 3000 #6000 
l1_decay=0.00 
l2_decay=0 # .5 
# 0.01 0.06 
sigma=0.005 
in_drop_rate = .2 
drop_rate = .5 

print (tr_X.shape[1]) 
# set network layout 
model = Sequential() 
model.add(Dense(2184, input_shape=(tr_X.shape[1],) 
       , init='he_normal', W_regularizer=l1_l2(l1=l1_decay, l2=l2_decay))) 
model.add(GaussianNoise(sigma)) 
model.add(Activation('relu')) 
model.add(BatchNormalization()) 
model.add(Dropout(in_drop_rate)) 


model.add(Dense(1310, init='he_normal', W_regularizer=l1_l2(l1=l1_decay, l2=l2_decay))) 
model.add(GaussianNoise(sigma)) 
model.add(Activation('relu')) 
model.add(BatchNormalization()) 
model.add(Dropout(drop_rate)) 

model.add(Dense(786, init='he_normal', W_regularizer=l1_l2(l1=l1_decay, l2=l2_decay))) 
model.add(GaussianNoise(sigma)) 
model.add(Activation('relu')) 
model.add(BatchNormalization()) 
model.add(Dropout(drop_rate)) 

model.add(Dense(472, init='he_normal', W_regularizer=l1_l2(l1=l1_decay, l2=l2_decay))) 
model.add(GaussianNoise(sigma)) 
model.add(Activation('relu')) 
model.add(BatchNormalization()) 
model.add(Dropout(drop_rate)) 


model.add(Dense(4, W_regularizer=l1_l2(l1=l1_decay, l2=l2_decay))) 
model.add(Activation('softmax')) 

# Callbacks 
model_checkpoint = ModelCheckpoint('best_model.hdf5', monitor='val_loss', save_best_only=True) 
early = EarlyStopping(monitor='val_loss', patience=600, verbose=0) 

# fit and evaluate the model 
model.compile(loss='categorical_crossentropy', 
       optimizer=Adam(lr=0.001))#SGD(lr=0.0019, momentum=0.9, decay=0.0, nesterov=True)) 
history = model.fit(tr_X, tr_y, batch_size=batch_size, 
        nb_epoch=nb_epoch, verbose=0, callbacks=[early, model_checkpoint], 
        validation_data=(va_X, va_y)) 
model.load_weights('best_model.hdf5') 
tr_pr = model.predict(tr_X, batch_size=batch_size, verbose=0) 

然而，测试一个真实数据（形式实验），我有一个不同的尺寸作为输入（例如：代替2184，我552）

因此，读取HDF5重量文件和用它来预测数据的类别。我写道：

# set parameters 
batch_size = 4 
l1_decay=0.00 
l2_decay=0 # .5 
# 0.01 0.06 
sigma=0.005 
in_drop_rate = .2 
drop_rate = .5 

# set network layout 
model = Sequential() 
model.add(Dense(552, input_shape=(552,) 
       , init='he_normal', W_regularizer=regularizers.l1_l2(l1=l1_decay, l2=l2_decay))) 
model.add(GaussianNoise(sigma)) 
model.add(Activation('relu')) 
model.add(BatchNormalization()) 
model.add(Dropout(in_drop_rate)) 


model.add(Dense(331, init='he_normal', W_regularizer=regularizers.l1_l2(l1=l1_decay, l2=l2_decay))) 
model.add(GaussianNoise(sigma)) 
model.add(Activation('relu')) 
model.add(BatchNormalization()) 
model.add(Dropout(drop_rate)) 

model.add(Dense(189, init='he_normal', W_regularizer=regularizers.l1_l2(l1=l1_decay, l2=l2_decay))) 
model.add(GaussianNoise(sigma)) 
model.add(Activation('relu')) 
model.add(BatchNormalization()) 
model.add(Dropout(drop_rate)) 

model.add(Dense(119, init='he_normal', W_regularizer=regularizers.l1_l2(l1=l1_decay, l2=l2_decay))) 
model.add(GaussianNoise(sigma)) 
model.add(Activation('relu')) 
model.add(BatchNormalization()) 
model.add(Dropout(drop_rate)) 

model.add(Dense(4, W_regularizer=regularizers.l1_l2(l1=l1_decay, l2=l2_decay))) 
model.add(Activation('softmax')) 


model.load_weights('best_model.hdf5') 
te_pr = model.predict(X, batch_size=batch_size, verbose=0) 

当我运行代码，我得到了以下异常：在理解个问题

C:\Users\M\Desktop\Dr Abeer Folder\Emotion Project_code and dataset\End User\Experiment_Calculation.py:106: UserWarning: Update your `Dense` call to the Keras 2 API: `Dense(119, kernel_regularizer=<keras.reg..., kernel_initializer="he_normal")` 

model.add(Dense(119, init='he_normal', W_regularizer=regularizers.l1_l2(l1=l1_decay, l2=l2_decay))) 


C:\Users\M\Desktop\Dr Abeer Folder\Emotion Project_code and dataset\End User\Experiment_Calculation.py:112: UserWarning: Update your `Dense` call to the Keras 2 API: `Dense(4, kernel_regularizer=<keras.reg...)` 

Traceback (most recent call last): 

File "C:\Users\M\Desktop\Dr Abeer Folder\Emotion Project_code and dataset\End User\main2.py", line 88, in BrowseFileHandler 

expcal.calclate_Experiment() 

File "C:\Users\M\Desktop\Dr Abeer Folder\Emotion Project_code and dataset\End User\Experiment_Calculation.py", line 66, in calclate_Experiment 

predictions = DNN(X) 

File "C:\Users\M\Desktop\Dr Abeer Folder\Emotion Project_code and dataset\End User\Experiment_Calculation.py", line 117, in DNN 

te_pr = model.predict(X, batch_size=batch_size, verbose=0) 

File "C:\Users\M\AppData\Roaming\Python\Python27\site-packages\keras\models.py", line 902, in predict 

return self.model.predict(x, batch_size=batch_size, verbose=verbose) 

File "C:\Users\M\AppData\Roaming\Python\Python27\site-packages\keras\engine\training.py", line 1585, in predict 

batch_size=batch_size, verbose=verbose) 

File "C:\Users\M\AppData\Roaming\Python\Python27\site-packages\keras\engine\training.py", line 1212, in _predict_loop 

batch_outs = f(ins_batch) 

File "C:\Users\M\AppData\Roaming\Python\Python27\site-packages\keras\backend\theano_backend.py", line 1158, in __call__ 

return self.function(*inputs) 

File "C:\Users\M\AppData\Roaming\Python\Python27\site-packages\theano\compile\function_module.py", line 898, in __call__ 

storage_map=getattr(self.fn, 'storage_map', None)) 

File "C:\Users\M\AppData\Roaming\Python\Python27\site-packages\theano\gof\link.py", line 325, in raise_with_op 

reraise(exc_type, exc_value, exc_trace) 

File "C:\Users\M\AppData\Roaming\Python\Python27\site-packages\theano\compile\function_module.py", line 884, in __call__ 

self.fn() if output_subset is None else\ 

ValueError: dimension mismatch in args to gemm (4,552)x(2184,2184)->(4,2184) 

Apply node that caused the error: GpuDot22(GpuFromHost.0, dense_1/kernel) 

Toposort index: 28 

Inputs types: [CudaNdarrayType(float32, matrix), CudaNdarrayType(float32, matrix)] 

Inputs shapes: [(4, 552), (2184, 2184)] 

Inputs strides: [(552, 1), (2184, 1)] 

Inputs values: ['not shown', 'not shown'] 

Outputs clients: [[GpuElemwise{Add}[(0, 0)](GpuDot22.0, 
GpuDimShuffle{x,0}.0), GpuElemwise{Composite{(i0 + i1 + (i2 * i3))}}[(0, 3)] 
(GpuDot22.0, GpuDimShuffle{x,0}.0, CudaNdarrayConstant{[[ 0.005]]}, GpuReshape{2}.0)]] 



HINT: Re-running with most Theano optimization disabled could give you a back-trace of when this node was created. This can be done with by setting the Theano flag 'optimizer=fast_compile'. If that does not work, Theano optimizations can be disabled with 'optimizer=None'. 

HINT: Use the Theano flag 'exception_verbosity=high' for a debugprint and storage map footprint of this apply node. 


model.add(Dense(4, W_regularizer=regularizers.l1_l2(l1=l1_decay, l2=l2_decay))) 

任何人都可以，请帮助。我是该地区的新人，特别是Keras和theano的使用。我该如何解决它？有没有办法改变预测的模型？

最好的问候，

Answer 1

这是非常直接的。

你训练了一个模型，第一层是2184x2184矩阵。所以你节省的体重是为2184输入而训练的，并且它们适合你训练的输入类型。

如果我理解正确的话，你想这个矩阵应用于552长度输入...你正在建设一个模型，其中的第一层是552x552矩阵，你想一个2184x2184矩阵加载到它... ...有是没有办法做到这一点......这不会奏效，你的输入应该是完全相同的大小。您无法更改训练有素的模型。

我希望你明白为什么它不起作用:-)如果没有，请询​​问澄清