How to correct a model?

TensorFlow General Discussion
, , ,
1

Hi everyone, I make a program and a model to let the PC or a raspberry pi do something later.

I made an XML with sentences, entities and actions. I can use the model, but the model sometimes outputs another entity & action when the input is the same as a sentence. Is there a way to correct the model?

I build the XML up like this:

  • command:
    input: Who is stronger
    entity: search
    action: searchwiki

The code for classifying:

import numpy as np
from keras.models import load_model

labels = open('nlu/entities.txt', 'r', encoding='utf-8').read().split('\n')
model = load_model('nlu/model.h5')

label2idx = {}
idx2label = {}

for k, label in enumerate(labels):
    label2idx[label] = k
    idx2label[k] = label


# Classify any given text into a category of our NLU framework
def classify(text):
    # Create an input array
    x = np.zeros((1, 57, 256), dtype='float32')

    if len(text) > 25:
        text = text[:25]
    # Fill the x array with data from input text
    for k, ch in enumerate(bytes(text.encode('utf-8'))):
        x[0, k, int(ch)] = 1.0

    out = model.predict(x)
    idx = out.argmax()

    # print('Text: "{}" is classified as "{}"'.format(text, idx2label[idx]))
    return {"entity": idx2label[idx], "conf": max(out[0])}

The code for making the model:

import numpy as np
import tensorflow as tf
import yaml
from tensorflow.keras.layers import LSTM, Dense
from tensorflow.keras.models import Sequential
from tensorflow.keras.utils import to_categorical
epochs = 1000

print("Num GPUs Available: ", len(tf.config.list_physical_devices('GPU')))

data = yaml.safe_load(open('train.yml').read())

# Reading the data

inputs, outputs = [], []

for command in data['commands']:
    inputs.append(command['input'].lower())
    outputs.append('{}\{}'.format(command['entity'], command['action']))

# Create a dataset
# Choose a level of tokenization: byte-level -> static vocab, handles out-out-vocabulary

# Create input data

max_sent = max([len((x.encode('utf-8'))) for x in inputs])

# Create arrays one-hot encoding (number of examples, seq length,  vocab_size)
# Create arrays sparse encoding (number of examples, seq length)

input_data = np.zeros((len(inputs), max_sent, 256), dtype='float32')

for i, inp in enumerate(inputs):
    for k, ch in enumerate(bytes(inp.encode('utf-8'))):
        input_data[i, k, int(ch)] = 1.0

# output_data = to_categorical(output_data, len(output_data))

# print(input_data.shape)

print(input_data[0].shape)

# print(len(chars))
# print('Max input seq:', max_sent)

labels = set(outputs)

fwrite = open('entities.txt', 'w', encoding='utf-8')
for label in labels:
    fwrite.write(label + '\n')
fwrite.close()

labels = open('entities.txt', 'r', encoding='utf-8').read().split('\n')

label2idx = {}
idx2label = {}

for k, label in enumerate(labels):
    label2idx[label] = k
    idx2label[k] = label

output_data = []

for output in outputs:
    output_data.append(label2idx[output])

output_data = to_categorical(output_data, len(labels))

model = Sequential()
model.add(LSTM(128))
model.add(Dense(len(labels), activation='softmax'))

model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['acc'])

model.fit(input_data, output_data, epochs=epochs)

model.save('model.h5')
3

Hello @anonymous023

Thank you for using TensorFlow
This issue would be with the output processing, Try to make model predict multiple outputs so that it can learn about the entity and actions also the internal relation between the input, entity and action.To make model learn more complex relations add more LSTM layers to make it learning more about the patterns in the dataset.