14
$\begingroup$

I struggle to interpret the Keras coding difference for one-to-many (e. g. classification of single images) and many-to-many (e. g. classification of image sequences) sequence labeling. I frequently see two different kind of codes:

Type 1 is where no TimeDistributed applied like this:

model=Sequential()

model.add(Convolution2D(nb_filters, kernel_size[0], kernel_size[1], border_mode="valid", input_shape=[1, 56,14]))
model.add(Activation("relu"))
model.add(Convolution2D(nb_filters, kernel_size[0], kernel_size[1]))
model.add(Activation("relu"))
model.add(MaxPooling2D(pool_size=pool_size))

model.add(Reshape((56*14,)))
model.add(Dropout(0.25))
model.add(LSTM(5))
model.add(Dense(50))
model.add(Dense(nb_classes))
model.add(Activation("softmax"))

Type 2 is where TimeDistributed is applied like this:

model = Sequential()

model.add(InputLayer(input_shape=(5, 224, 224, 3)))
model.add(TimeDistributed(Convolution2D(64, (3, 3))))
model.add(TimeDistributed(MaxPooling2D((2,2), strides=(2,2))))
model.add(LSTM(10))
model.add(Dense(3))

My questions are:

  • Is my assumption correct that Type 1 is the one-to-many kind and Type 2 is the many-to-many kind? Or TimeDistributed has no relevance in this aspect?

  • In either case of one-to-many or many-to-many is the last dense layer supposed to be 1 node "long" (emitting only one value in turn) and
    the previous recurrent layer is responsible to determine how many
    1-long value to emit? Or the last dense layer is supposed to consist of N nodes where N=max sequence length? If so, what is the point of
    using RNN here when we could produce a similar input with multiple
    outputs with N parallel "vanilla" estimators?

  • How to define the number of timesteps in RNNs? Is it somehow
    correlated with the output sequence length or is it just a
    hyperparameter to tune?

  • Inn case of my Type 1 example above what is the point of applying
    LSTM when the model emits only one class prediction (of possible
    nb_classes)? What if one omits the LSTM layer?

$\endgroup$
  • $\begingroup$ Could you provide the summary of both models? $\endgroup$ – Fadi Bakoura May 20 '18 at 0:42
2
$\begingroup$

The point of using any recurrent layer is to have the output be a result of not only a single item independent of other items, but rather a sequence of items, such that the output of the layer's operation on one item in the sequence is the result of both that item and any item before it in the sequence. The number of timesteps defines how long such a sequence is. That is, how many items that should be handled in a sequence, and affect each other's resulting output.

A LSTM layer operates in such a way that it accepts input on the form number_of_timesteps, dimensions_of_each_item. If the parameter return_sequences is set to False, which it is by default, the layer "compounds" the inputs of all the timesteps into a single output. If you consider a sequence of, say 10 items, a LSTM layer with return_sequences set to False will from such a sequence produce a single output item, and the attributes of this single item will be a result of all the items (timesteps) in the sequence. This is what you want in the case of a many-to-one design.

A LSTM layer with return_sequences set to True will for each item (timestep) in an input sequence produce an output. This is done in such a way that at any timestep, the output will depend not only on the item that is currently being operated on, but also the previous items in the sequence. This is what you want in the case of a many-to-many design.

As a LSTM layer takes a sequence of items as input, any layer before a LSTM layer in your model will need to produce a sequence as an output. In the case of your Type 1 model, the first few layers do not operate on sequences, but rather a single item at a time. This does hence not produce a sequence of items on which to operate for the LSTM.

Using TimeDistributed makes it possible to have a layer operate on every item in a sequence without the items affecting each other. TimeDistributed layers thus operate on sequences of items, but there is no recursion.

In the case of your type 2 model, the first layers will produce a sequence 5 timesteps long, and the operations done on each of the items in the sequence will be independent of each other, since the layers wrapped in TimeDistributed are non-recurrent. As the LSTM layer uses the default settings, return_sequences = False, the LSTM layer will produce a single output for each such sequence of 5 items.

The final number of output nodes in your model depends completely on the use case. A single node is suitable for something like binary classification or for producing some sort of score.

| improve this answer | |
$\endgroup$
1
$\begingroup$

I think that you might be able to use my previous work. In this code I create sine waves (of random wavelengths and phases) and train an LSTM to a sequence of points from these sine waves and output a sequence of 150 points completing each sine wave.

This is the model:

    features_num=5 
    latent_dim=40

    ##
    encoder_inputs = Input(shape=(None, features_num))
    encoded = LSTM(latent_dim, return_state=False ,return_sequences=True)(encoder_inputs)
    encoded = LSTM(latent_dim, return_state=False ,return_sequences=True)(encoded)
    encoded = LSTM(latent_dim, return_state=False ,return_sequences=True)(encoded)
    encoded = LSTM(latent_dim, return_state=True)(encoded)

    encoder = Model (input=encoder_inputs, output=encoded)
    ##

    encoder_outputs, state_h, state_c = encoder(encoder_inputs)
    encoder_states = [state_h, state_c]

    decoder_inputs=Input(shape=(1, features_num))
    decoder_lstm_1 = LSTM(latent_dim, return_sequences=True, return_state=True)
    decoder_lstm_2 = LSTM(latent_dim, return_sequences=True, return_state=True)
    decoder_lstm_3 = LSTM(latent_dim, return_sequences=True, return_state=True)
    decoder_lstm_4 = LSTM(latent_dim, return_sequences=True, return_state=True)

    decoder_dense = Dense(features_num)

    all_outputs = []
    inputs = decoder_inputs


    states_1=encoder_states
   # Place holder values:
    states_2=states_1; states_3=states_1; states_4=states_1

    for _ in range(1):
        # Run the decoder on the first timestep
        outputs_1, state_h_1, state_c_1 = decoder_lstm_1(inputs, initial_state=states_1)
        outputs_2, state_h_2, state_c_2 = decoder_lstm_2(outputs_1)
        outputs_3, state_h_3, state_c_3 = decoder_lstm_3(outputs_2)
        outputs_4, state_h_4, state_c_4 = decoder_lstm_4(outputs_3)

        # Store the current prediction (we will concatenate all predictions later)
        outputs = decoder_dense(outputs_4)
        all_outputs.append(outputs)
        # Reinject the outputs as inputs for the next loop iteration
        # as well as update the states
        inputs = outputs
        states_1 = [state_h_1, state_c_1]
        states_2 = [state_h_2, state_c_2]
        states_3 = [state_h_3, state_c_3]
        states_4 = [state_h_4, state_c_4]


    for _ in range(149):
        # Run the decoder on each timestep
        outputs_1, state_h_1, state_c_1 = decoder_lstm_1(inputs, initial_state=states_1)
        outputs_2, state_h_2, state_c_2 = decoder_lstm_2(outputs_1, initial_state=states_2)
        outputs_3, state_h_3, state_c_3 = decoder_lstm_3(outputs_2, initial_state=states_3)
        outputs_4, state_h_4, state_c_4 = decoder_lstm_4(outputs_3, initial_state=states_4)

        # Store the current prediction (we will concatenate all predictions later)
        outputs = decoder_dense(outputs_4)
        all_outputs.append(outputs)
        # Reinject the outputs as inputs for the next loop iteration
        # as well as update the states
        inputs = outputs
        states_1 = [state_h_1, state_c_1]
        states_2 = [state_h_2, state_c_2]
        states_3 = [state_h_3, state_c_3]
        states_4 = [state_h_4, state_c_4]


    # 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')


    print(model.summary())

And this is the entire script:

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
from keras import backend as K

from keras.models import load_model

import os


features_num=5 
latent_dim=40

##
encoder_inputs = Input(shape=(None, features_num))
encoded = LSTM(latent_dim, return_state=False ,return_sequences=True)(encoder_inputs)
encoded = LSTM(latent_dim, return_state=False ,return_sequences=True)(encoded)
encoded = LSTM(latent_dim, return_state=False ,return_sequences=True)(encoded)
encoded = LSTM(latent_dim, return_state=True)(encoded)

encoder = Model (input=encoder_inputs, output=encoded)
##

encoder_outputs, state_h, state_c = encoder(encoder_inputs)
encoder_states = [state_h, state_c]

decoder_inputs=Input(shape=(1, features_num))
decoder_lstm_1 = LSTM(latent_dim, return_sequences=True, return_state=True)
decoder_lstm_2 = LSTM(latent_dim, return_sequences=True, return_state=True)
decoder_lstm_3 = LSTM(latent_dim, return_sequences=True, return_state=True)
decoder_lstm_4 = LSTM(latent_dim, return_sequences=True, return_state=True)

decoder_dense = Dense(features_num)

all_outputs = []
inputs = decoder_inputs

# Place holder values:
states_1=encoder_states
states_2=states_1; states_3=states_1; states_4=states_1

for _ in range(1):
    # Run the decoder on one timestep
    outputs_1, state_h_1, state_c_1 = decoder_lstm_1(inputs, initial_state=states_1)
    outputs_2, state_h_2, state_c_2 = decoder_lstm_2(outputs_1)
    outputs_3, state_h_3, state_c_3 = decoder_lstm_3(outputs_2)
    outputs_4, state_h_4, state_c_4 = decoder_lstm_4(outputs_3)

    # Store the current prediction (we will concatenate all predictions later)
    outputs = decoder_dense(outputs_4)
    all_outputs.append(outputs)
    # Reinject the outputs as inputs for the next loop iteration
    # as well as update the states
    inputs = outputs
    states_1 = [state_h_1, state_c_1]
    states_2 = [state_h_2, state_c_2]
    states_3 = [state_h_3, state_c_3]
    states_4 = [state_h_4, state_c_4]


for _ in range(149):
    # Run the decoder on one timestep
    outputs_1, state_h_1, state_c_1 = decoder_lstm_1(inputs, initial_state=states_1)
    outputs_2, state_h_2, state_c_2 = decoder_lstm_2(outputs_1, initial_state=states_2)
    outputs_3, state_h_3, state_c_3 = decoder_lstm_3(outputs_2, initial_state=states_3)
    outputs_4, state_h_4, state_c_4 = decoder_lstm_4(outputs_3, initial_state=states_4)

    # Store the current prediction (we will concatenate all predictions later)
    outputs = decoder_dense(outputs_4)
    all_outputs.append(outputs)
    # Reinject the outputs as inputs for the next loop iteration
    # as well as update the states
    inputs = outputs
    states_1 = [state_h_1, state_c_1]
    states_2 = [state_h_2, state_c_2]
    states_3 = [state_h_3, state_c_3]
    states_4 = [state_h_4, state_c_4]


# 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')


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 = np.random.randint(150, 300)+150       
        data=make_data(1000,sequence_length,features_num,0) # category=0 in train


    #   decoder_target_data is the same as decoder_input_data but offset by one timestep

        encoder_input_data = data[:,:-150,:] # all but last 150 

        decoder_input_data = data[:,-151,:] # the one before the last 150.
        decoder_input_data=decoder_input_data.reshape((decoder_input_data.shape[0],1,decoder_input_data.shape[1]))


        decoder_target_data = (data[:, -150:, :]) # last 150        
        yield [encoder_input_data, decoder_input_data], decoder_target_data
def val_generator():
    while True:

        sequence_length = np.random.randint(150, 300)+150       
        data=make_data(1000,sequence_length,features_num,2) # category=2 in val

        encoder_input_data = data[:,:-150,:] # all but last 150 

        decoder_input_data = data[:,-151,:] # the one before the last 150.
        decoder_input_data=decoder_input_data.reshape((decoder_input_data.shape[0],1,decoder_input_data.shape[1]))

        decoder_target_data = (data[:, -150:, :]) # last 150        
        yield [encoder_input_data, decoder_input_data], decoder_target_data

filepath_for_w= 'flux_p2p_s2s_model.h5' 
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 150; 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],150,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 = np.random.randint(150, 300)+150       
        data=make_data(470,sequence_length,features_num,4) # category=4 in test

        encoder_input_data = data[:,:-150,:] # all but last 150 

        decoder_input_data = data[:,-151,:] # the one before the last 150.
        decoder_input_data=decoder_input_data.reshape((decoder_input_data.shape[0],1,decoder_input_data.shape[1]))

        decoder_target_data = (data[:, -150:, :]) # last 150        
        return [encoder_input_data, decoder_input_data],    decoder_target_data

x,y= test_maker()   



a=predict_many_waves_from_input (x) # is that right..?
x=x[0] # keep the wave (generated data except last 150 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])))
| improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.