在 keras 中使用 LSTM 进行预测

Question

I'm stuck with a problem with my LSTM. What I want to do is the following:

I have a dataset in the form:

0.04,-9.77,0.71,1,0,0,0
...
...

The first three parameters are data collected by an accelerometer: X acceleration, Y acceleration, Z acceleration

The last four columns are the labels:

[1,0,0,0] [0,1,0,0] [0,0,1,0] [0,0,0,1] [0,0,0,0]

where each one represents a different class.

My network is declared as follows:

 class Config:
        def __init__(self):
            """network parameters"""
            self.batch_size = 16
            self.input_size = 3
            self.seq_max_len = 20
            self.rnn_size = 50
            self.keep_prob = 1
            self.mlp_hidden_size = 100
            self.mlp_projection_activation = tf.nn.tanh
            self.num_classes = 4
            self.learning_rate = 0.001
            self.epochs = 10
    
    
        model = tf.keras.Sequential([
        tf.keras.layers.InputLayer(input_shape=(config.seq_max_len, config.input_size)),
        tf.keras.layers.LSTM(units=config.rnn_size, return_sequences=True, return_state=False),
        tf.keras.layers.Dense(units=config.mlp_hidden_size, activation=config.mlp_projection_activation),
        tf.keras.layers.Dense(units=config.num_classes, activation='softmax'),
    ])
    
    loss_fn = tf.keras.losses.CategoricalCrossentropy(from_logits=True)
    model.compile(optimizer='adam', loss=loss_fn, metrics=['accuracy'])
    
    model.fit(x_train, y_train, batch_size=config.batch_size, epochs=config.epochs)

Now, the problem is that this is not working as i would. When i try to predict, let's say using an array:

arr = np.array([(-0.12,-9.85,0.82),(-1.33,-10,1.61),(-1.57,-10.04,0.9),(0.08,-9.14,0.51),(3.77,-8.36,-0.55),(6.71,-8.43,-1.69),
(9.22,-8.28,-2.63),(10.75,-7.65,-2.98),(9.26,-7.61,-2.35),(6.16,-7.85,-1.77),(2.35,-8.51,-0.78),(-1.10,-8.87,0.71),(-3.61,-9.14,2.31),
                (-5.49,-9.65,3.69),
                (-5.33,-9.49,3.14),
                (-4.24,-9.26,3.30),
                (-2.43,-9.06,2.24),
                (-0.39,-8.87,1.29),
                (3.61,-8.55,-1.22),
                (7.10,-8.28,-1.57)])

Composed by 20 triplette of 3d vectors (accelerations) what i obtain is

predictions = model.predict_classes(arr)
[[0 2 2 0 0 0 0 0 0 0 0 2 2 2 2 2 2 2 0 0]]

which is a vector that represents a prediction for each triplette in the arr vector. What i want, is to have ONLY ONE prediction, after the 20 triplette. This is because my data represent a time series and what i am interested in, is to know if the network is able to classify data after a certain amount of acceleration vectors (20 in this case).

Can help me please?

EDIT

full code:

import tensorflow as tf
import numpy as np
import pandas as pd
import random
import sys
np.set_printoptions(threshold=sys.maxsize)

def get_dataset(filename, config):
    df = pd.read_csv(filename, header=None, skiprows=1)
    x = df[[0, 1, 2]].values
    y = df[[3, 4, 5, 6]].values
    dataset_x, dataset_y = [],[]

    for i in range(x.shape[0]//config.seq_max_len):
        sequence_x, sequence_y = [],[]
        for j in range(config.seq_max_len):
            sequence_x.append(x[i*config.seq_max_len + j])
            sequence_y.append(y[i*config.seq_max_len + j])
        dataset_x.append(sequence_x)
        dataset_y.append(sequence_y)

    return np.array(dataset_x), np.array(dataset_y)


class Config:
    def __init__(self):
        """definizione dei parametri della rete"""
        self.batch_size = 16
        self.input_size = 3
        self.seq_max_len = 20
        self.rnn_size = 50
        self.keep_prob = 1
        self.mlp_hidden_size = 100
        self.mlp_projection_activation = tf.nn.tanh
        self.num_classes = 4
        self.learning_rate = 0.001
        self.epochs = 10

config = Config()

x_train, y_train = get_dataset('data_new.csv', config)

arr = np.array([(-0.12,-9.85,0.82),(-1.33,-10,1.61),(-1.57,-10.04,0.9),(0.08,-9.14,0.51),(3.77,-8.36,-0.55),(6.71,-8.43,-1.69),
(9.22,-8.28,-2.63),(10.75,-7.65,-2.98),(9.26,-7.61,-2.35),(6.16,-7.85,-1.77),(2.35,-8.51,-0.78),(-1.10,-8.87,0.71),(-3.61,-9.14,2.31),
                (-5.49,-9.65,3.69),
                (-5.33,-9.49,3.14),
                (-4.24,-9.26,3.30),
                (-2.43,-9.06,2.24),
                (-0.39,-8.87,1.29),
                (3.61,-8.55,-1.22),
                (7.10,-8.28,-1.57)])
arr2 = np.reshape(arr,(1,20,3))
print(arr2.shape)


model = tf.keras.Sequential([
    tf.keras.layers.InputLayer(input_shape=(config.seq_max_len, config.input_size)),
    tf.keras.layers.LSTM(units=config.rnn_size, return_sequences=True, return_state=False),
    tf.keras.layers.Dense(units=config.mlp_hidden_size, activation=config.mlp_projection_activation),
    tf.keras.layers.Dense(units=config.num_classes, activation='softmax'),
])


loss_fn = tf.keras.losses.CategoricalCrossentropy(from_logits=True)
model.compile(optimizer='adam', loss=loss_fn, metrics=['accuracy'])

model.fit(x_train, y_train, batch_size=config.batch_size, epochs=config.epochs)
predictions = model.predict(arr2)
predictions = np.argmax(predictions, axis=-1)
print("PREDIZIONI---------")
print(predictions.shape)
print(predictions)

Answer 1

There are two possible issues. One is if you set

tf.keras.layers.LSTM(units=.., return_sequences=True, return_state=False)

You would get as follows if you print model. summary() model. summary() in the final layer of your model. And that probably not what you want in your final layer.

dense_5 (Dense)              (None, 20, 4)             404       
=================================================================

So, you should use return_sequence = False to get the final layer output shape as follows:

dense_7 (Dense)              (None, 4)                 404       
=================================================================

Second, you set in your loss function that

 ....CategoricalCrossentropy(from_logits=True)

but you set activation='softma' in the final layer to get the probabilities instead of logits.

....Dense(units=config.num_classes, activation='softmax')

So, based on this set the parameter as follows:

....LSTM(units=.., return_sequences=False, return_state=False)
...
....CategoricalCrossentropy(from_logits=False) # compute probabilities 
...
y_pred = model.predict(arr)
y_pred = np.argmax(y_pred, axis=-1)