I'm experimenting with LSTMs, specifically, inputting a sequence into an LSTM, transferring the states into another LSTM, and decoding the sequence. I added an autoencoder between the two LSTMs, encoding and then decoding the transferred states via a lower dimensional latent space.
This works fine when I create the model and fit it. However, if I save this model, and then either try to continue training it, or even just use it without additional training, the model does not run and I get the following warning:
Traceback (most recent call last):
File "s2s_AE_2.py", line 140, in <module>
model.fit_generator(train_generator(),callbacks=[checkpointer], steps_per_epoch=30, epochs=2000, verbose=1,validation_data=val_generator(),validation_steps=30)
File "C:\ProgramData\Anaconda3\lib\site-packages\keras\legacy\interfaces.py", line 91, in wrapper
return func(*args, **kwargs)
File "C:\ProgramData\Anaconda3\lib\site-packages\keras\engine\training.py", line 2224, in fit_generator
class_weight=class_weight)
File "C:\ProgramData\Anaconda3\lib\site-packages\keras\engine\training.py", line 1877, in train_on_batch
class_weight=class_weight)
File "C:\ProgramData\Anaconda3\lib\site-packages\keras\engine\training.py", line 1476, in _standardize_user_data
exception_prefix='input')
File "C:\ProgramData\Anaconda3\lib\site-packages\keras\engine\training.py", line 86, in _standardize_input_data
str(len(data)) + ' arrays: ' + str(data)[:200] + '...')
ValueError: Error when checking model input: the list of Numpy arrays that you are passing to your model is not the size the model expected. Expected to see 1 array(s), but instead got the following list of 2 arrays: [array([[[ 0.47338937, 0.75865918, 0.37731877, 0.63840222,
0.14653083],
[ 0.52119932, 0.78308798, 0.45885839, 0.66738276,
0.20393343],
[ 0.5674261 , 0.806364...
My code is as follows:
from keras.models import Model
from keras.layers import Input, LSTM, Dense, TimeDistributed,Lambda, Dropout, Activation ,RepeatVector
from keras.callbacks import ModelCheckpoint
import numpy as np
from keras.layers import Lambda, Concatenate
from keras import backend as K
from keras.models import load_model
import os
seq_length=150
features_num=5
LSTM_latent_dim=40
AE_latent_dim=10
encoder_inputs = Input(shape=(seq_length, features_num))
encoder = LSTM(LSTM_latent_dim, return_state=True)
encoder_outputs, state_h, state_c = encoder(encoder_inputs)
merged_encoder_states = Concatenate(axis=-1)([state_h, state_c])
encoded_states=Dense(AE_latent_dim,activation='relu')(merged_encoder_states)
decoded_states=Dense(LSTM_latent_dim*2, activation='relu')(encoded_states)
decoder_inputs=Input(shape=(1, features_num))
decoder_lstm = LSTM(LSTM_latent_dim, return_sequences=True, return_state=True)
decoder_dense = Dense(features_num)
all_outputs = []
inputs = decoder_inputs
states=[decoded_states[:,:LSTM_latent_dim],decoded_states[:,LSTM_latent_dim:]]
for _ in range(seq_length):
# Run the decoder on one timestep
outputs, state_h, state_c = decoder_lstm(inputs, initial_state=states)
outputs = decoder_dense(outputs)
# Store the current prediction (we will concatenate all predictions later)
all_outputs.append(outputs)
# Reinject the outputs as inputs for the next loop iteration
# as well as update the states
inputs = outputs
states = [state_h, state_c]
# Concatenate all predictions
decoder_outputs = Lambda(lambda x: K.concatenate(x, axis=1))(all_outputs)
model = Model([encoder_inputs, decoder_inputs], decoder_outputs)
#model = load_model('pre_model.h5')
filepath_for_w= 'AE2_p2p_s2s_model.h5'
try:
model = load_model(filepath_for_w) # if model was previouslly run, continue from it
print("loaded model")
except: print("new model")
print(model.summary())
model.compile(loss='mean_squared_error', optimizer='adam')
def create_wavelength(min_wavelength, max_wavelength, fluxes_in_wavelength, category ) :
#category :: 0 - train ; 2 - validate ; 4- test. 1;3;5 - dead space
c=(category+np.random.random())/6
k = fluxes_in_wavelength
#
base= (np.trunc(k*np.random.random()*(max_wavelength-min_wavelength)) +k*min_wavelength) /k
answer=base+c/k
return (answer)
def make_line(length,category):
shift= np.random.random()
wavelength = create_wavelength(30,10,1,category)
a=np.arange(length)
answer=np.sin(a/wavelength+shift)
return answer
def make_data(seq_num,seq_len,dim,category):
data=np.array([]).reshape(0,seq_len,dim)
for i in range (seq_num):
mini_data=np.array([]).reshape(0,seq_len)
for j in range (dim):
line = make_line(seq_len,category)
line=line.reshape(1,seq_len)
mini_data=np.append(mini_data,line,axis=0)
mini_data=np.swapaxes(mini_data,1,0)
mini_data=mini_data.reshape(1,seq_len,dim)
data=np.append(data,mini_data,axis=0)
return (data)
def train_generator():
while True:
sequence_length = seq_length+1
data=make_data(1000,sequence_length,features_num,0) # category=0 in train
#
encoder_input_data =data[:,1:,:] # all
#
decoder_input_data = data[:,0,:] # the first value in the sequence
decoder_input_data=decoder_input_data.reshape((decoder_input_data.shape[0],1,decoder_input_data.shape[1]))
#
#
decoder_target_data = encoder_input_data
yield [encoder_input_data, decoder_input_data], decoder_target_data
def val_generator():
while True:
sequence_length =seq_length+1
data=make_data(1000,sequence_length,features_num,2) # category=2 in val
#
#
# # decoder_target_data is the same as decoder_input_data but offset by one timestep
#
encoder_input_data =data[:,1:,:] # all
#
decoder_input_data = data[:,0,:] # the one before the last one.
decoder_input_data=decoder_input_data.reshape((decoder_input_data.shape[0],1,decoder_input_data.shape[1]))
#
#
decoder_target_data = encoder_input_data
yield [encoder_input_data, decoder_input_data], decoder_target_data
checkpointer=ModelCheckpoint(filepath_for_w, monitor='val_loss', verbose=0, save_best_only=True, mode='auto', period=1)
model.fit_generator(train_generator(),callbacks=[checkpointer], steps_per_epoch=30, epochs=2000, verbose=1,validation_data=val_generator(),validation_steps=30)
model.save(filepath_for_w)
def predict_wave(input_wave,input_for_decoder): # input wave= x[n,:,:], ie points except the last seq_length; each wave has feature_num features. run this function for all such instances (=n)
#print (input_wave.shape)
#print (input_for_decoder.shape)
pred= model.predict([input_wave,input_for_decoder])
#
return pred
def predict_many_waves_from_input(x):
x, x2=x # x == encoder_input_data ; x==2 decoder_input_data
#
instance_num= x.shape[0]
#
#
multi_predict_collection=np.zeros((x.shape[0],seq_length,x.shape[2]))
#
for n in range(instance_num):
input_wave=x[n,:,:].reshape(1,x.shape[1],x.shape[2])
input_for_decoder=x2[n,:,:].reshape(1,x2.shape[1],x2.shape[2])
wave_prediction=predict_wave(input_wave,input_for_decoder)
multi_predict_collection[n,:,:]=wave_prediction
return (multi_predict_collection)
def test_maker():
if True:
sequence_length = seq_length +1
data=make_data(470,sequence_length,features_num,4) # category=4 in test
#
encoder_input_data =data[:,1:,:] # all
#
decoder_input_data = data[:,0,:] # the first value
decoder_input_data=decoder_input_data.reshape((decoder_input_data.shape[0],1,decoder_input_data.shape[1]))
#
#
decoder_target_data = encoder_input_data
return [encoder_input_data, decoder_input_data], decoder_target_data
x,y= test_maker()
a=predict_many_waves_from_input (x)
x=x[0] # keep the wave (generated data except last seq_length time points)
print (x.shape)
print (y.shape)
print (a.shape)
np.save ('a.npy',a)
np.save ('y.npy',y)
np.save ('x.npy',x)
print (np.mean(np.absolute(y[:,:,0]-a[:,:,0])))
print (np.mean(np.absolute(y[:,:,1]-a[:,:,1])))
print (np.mean(np.absolute(y[:,:,2]-a[:,:,2])))
print (np.mean(np.absolute(y[:,:,3]-a[:,:,3])))
print (np.mean(np.absolute(y[:,:,4]-a[:,:,4])))
The culprit might be this line:
states=[decoded_states[:,:LSTM_latent_dim],decoded_states[:,LSTM_latent_dim:]]
After combining the states of the encoding LSTM and passing them through the autoencoder, I split them back into c and h (the cell state and the hidden state, respectively) and feed them into the decoder LSTM.
It seems reasonable to me that this step occurs correctly when the initial model is used, but is somehow incorrectly saved into the model file (or incorrectly loaded from the model file), resulting in a defective loaded model.
Further supporting my assessment, in my opinion, is the fact that when this line is replaced with
states= [state_h, state_c]
, the loaded model is able to run correctly (fitting and predicting), but of course this does away with the state autoencoder so I cannot use it except for zooming in on the bug.
So, I ask your help regarding two questions:
Why does this problem occur?
How do I solve it?
A possible partial solution is to forgo the saving of the model in its entirety, and just save (and load) the model's weights.
Replacing the lines
model = load_model(filepath_for_w)
...
checkpointer=ModelCheckpoint(filepath_for_w, monitor='val_loss', verbose=0, save_best_only=True, mode='auto', period=1)
...
model.save(filepath_for_w)
with
model.load_weights(filepath_for_w)
...
checkpointer=ModelCheckpoint(filepath_for_w, save_weights_only=True, monitor='val_loss', verbose=0, save_best_only=True, mode='auto', period=1)
...
model.save_weights(filepath_for_w)
does the trick. The model can be loaded for further fitting and for prediction.
However this does not allow the saving of the entire model; I still need to keep the architecture in the code in order to populate it with the weights. It also does not explain why does this problem occurs to begin with.