LSTM seq2seq 模型之英语到法语翻译

1. 数据集下载

Machine-Translation-eng-fra | Kaggle

2. 预处理的完整的代码

import os.path

import numpy as np
import tensorflow as tf
import keras
from keras.callbacks import EarlyStopping, ModelCheckpoint
from keras.preprocessing.text import Tokenizer
from keras.utils import pad_sequences
from sklearn.model_selection import train_test_split
import pandas as pd
import numpy as np
import json

from nlp.util.matplotlib_helper import *


batch_size = 36  # Batch size for training.
epochs = 50  # Number of epochs to train for.
latent_dim = 256  # Latent dimensionality of the encoding space.
num_samples = 100000  # Number of samples to train on.
# Path to the data txt file on disk.

# base_dir = "dataset"
base_dir = "C:/apps/ml_datasets/fra-eng"

dataset_path = os.path.join(base_dir, "fra.txt")

START_ = "sostok "
END_ = " eostok"

def get_dataset():
    # Vectorize the data.
    input_texts = []
    target_texts = []
    input_characters = set()
    target_characters = set()
    with open(dataset_path, "r", encoding="utf-8") as f:
        lines = f.read().split("\n")
        print("length of lines:"+ str(len(lines)))
    for line in lines[: min(num_samples, len(lines) - 1)]:
        input_text, target_text, _ = line.split("\t")
        # We use "tab" as the "start sequence" character
        # for the targets, and "\n" as "end sequence" character.
        target_text = START_ + target_text + END_
        input_texts.append(input_text)
        target_texts.append(target_text)
        for char in input_text:
            if char not in input_characters:
                input_characters.add(char)
        for char in target_text:
            if char not in target_characters:
                target_characters.add(char)
    input_characters = sorted(list(input_characters))
    target_characters = sorted(list(target_characters))
    num_encoder_tokens = len(input_characters)
    num_decoder_tokens = len(target_characters)
    max_encoder_seq_length = max([len(txt) for txt in input_texts])
    max_decoder_seq_length = max([len(txt) for txt in target_texts])
    print("Number of samples:", len(input_texts))
    print("Number of unique input tokens:", num_encoder_tokens)
    print("Number of unique output tokens:", num_decoder_tokens)
    print("Max sequence length for inputs:", max_encoder_seq_length)
    print("Max sequence length for outputs:", max_decoder_seq_length)

    return input_texts, target_texts


def check_distribution(x_train ,y_train):
    x_count = []
    y_count = []

    for sent in x_train:
        x_count.append(len(sent.split()))
    for sent in y_train:
        y_count.append(len(sent.split()))

    graph_df = pd.DataFrame()
    graph_df['x_count'] = x_count
    graph_df['y_count'] = y_count

    import matplotlib.pyplot as plt

    graph_df.hist(bins = 5)
    plt.show()

    # Check how much % of summary have 0-15 words
    cnt = 0
    for i in x_train:
        if (len(i.split()) <= 6):
            cnt = cnt + 1
    print(cnt / len(x_train))

    # Check how much % of summary have 0-15 words
    cnt = 0
    for i in y_train:
        if (len(i.split()) <= 10):
            cnt = cnt + 1
    print(cnt / len(y_train))



def get_tokenizers(x_train ,y_train):

    input_tokenizer = Tokenizer()
    input_tokenizer.fit_on_texts(x_train)

    target_tokenizer = Tokenizer()
    target_tokenizer.fit_on_texts(y_train)

    input_sequences = input_tokenizer.texts_to_sequences(x_train)
    target_sequences = target_tokenizer.texts_to_sequences(y_train)

    max_text_len = max(len(seq) for seq in input_sequences)
    max_summary_len = max(len(seq) for seq in target_sequences)
    print('max_text_len:', max_text_len)
    print('max_summary_len:', max_summary_len)


    max_text_len = 6
    max_summary_len = 10

    input_sequences = pad_sequences(input_sequences, maxlen=max_text_len, padding='post')
    target_sequences = pad_sequences(target_sequences, maxlen=max_summary_len, padding='post')

    print(max_text_len, input_sequences[0])
    print(max_summary_len, target_sequences[0])


    x_voc = len(input_tokenizer.word_counts)+1
    print("Size of vocabulary in X = {}".format(x_voc))

    y_voc = len(target_tokenizer.word_counts)+1
    print("Size of vocabulary in Y = {}".format(y_voc))

    y_tokenizer_json = target_tokenizer.to_json()

    # Save the tokenizer to a file
    y_tokenizer_file = os.path.join(base_dir, "y_tokenizer.json")
    with open(y_tokenizer_file, 'w', encoding='utf-8') as f:
        f.write(json.dumps(y_tokenizer_json, ensure_ascii=False))


    x_tokenizer_json = input_tokenizer.to_json()

    # Save the tokenizer to a file
    x_tokenizer_file = os.path.join(base_dir, "x_tokenizer.json")
    with open(x_tokenizer_file, 'w', encoding='utf-8') as f:
        f.write(json.dumps(x_tokenizer_json, ensure_ascii=False))


    return input_tokenizer, target_tokenizer, input_sequences, target_sequences


if __name__ == '__main__':

    x_train ,y_train = get_dataset()

    # check_distribution(x_train, y_train)

    input_tokenizer, target_tokenizer, input_sequences, target_sequences = get_tokenizers(x_train ,y_train);

    # Split the dataset into training and validation sets
    x_train, x_val, y_train, y_val = train_test_split( input_sequences, target_sequences, test_size=0.2, random_state=42 )

    np.savez(os.path.join(base_dir,"train"), x_tr=x_train, y_tr=y_train, x_val=x_val, y_val=y_val)

    # for i in range(5):
    #     print("x = ", x_train[i] )
    #     print("y = ", y_train[i])

3. 训练模型的完整代码

import keras
from keras.layers import Input, LSTM, Embedding, Dense, Concatenate, TimeDistributed
from keras.models import Model
from keras.callbacks import EarlyStopping
from keras.preprocessing.text import tokenizer_from_json
import json
import numpy as np
import os

model_path = "model"

latent_dim = 200
embedding_dim = 100
max_text_len = 6
max_summary_len = 10

base_dir = "C:/apps/ml_datasets/fra-eng"


START_ = "sostok"
END_ = "eostok"

def get_dataset():

    train_set = np.load(os.path.join(base_dir, "train.npz"))
    x_tr = train_set['x_tr']
    y_tr = train_set['y_tr']
    x_val = train_set['x_val']
    y_val = train_set['y_val']

    print("X_train:", x_tr.shape)
    print("y_train:", y_tr.shape)
    print("X_test:", x_val.shape)
    print("y_test:", y_val.shape)

    with open(os.path.join(base_dir, "y_tokenizer.json")) as f:
        data = json.load(f)
        y_tokenizer = tokenizer_from_json(data)

    with open(os.path.join(base_dir,"x_tokenizer.json")) as f:
        data = json.load(f)
        x_tokenizer = tokenizer_from_json(data)

    return x_tr, y_tr, x_val, y_val, x_tokenizer, y_tokenizer

def get_model( max_text_len = 100, x_voc = 33288, y_voc = 11572):
    # K.clear_session()

    # Encoder
    encoder_inputs = Input(shape=(max_text_len,), name="encoder_inputs")

    #embedding layer
    enc_emb =  Embedding(x_voc, embedding_dim,trainable=True, name= "encoder_embedding")(encoder_inputs)

    #encoder lstm 1
    encoder_lstm1 = LSTM(latent_dim,return_sequences=True,return_state=True,dropout=0.4,recurrent_dropout=0.4, name = "encoder_lstm1")
    # encoder_outputs, state_h, state_c = encoder_lstm1(enc_emb)
    encoder_output1, state_h1, state_c1 = encoder_lstm1(enc_emb)

    #encoder lstm 2
    encoder_lstm2 = LSTM(latent_dim,return_sequences=True,return_state=True,dropout=0.4,recurrent_dropout=0.4, name = "encoder_lstm2")
    encoder_output2, state_h2, state_c2 = encoder_lstm2(encoder_output1)

    #encoder lstm 3
    encoder_lstm3=LSTM(latent_dim, return_state=True, return_sequences=True,dropout=0.4,recurrent_dropout=0.4, name = "encoder_lstm3")
    encoder_outputs, state_h, state_c= encoder_lstm3(encoder_output2)

    # Set up the decoder, using `encoder_states` as initial state.
    decoder_inputs = Input(shape=(None,), name="decoder_inputs")

    #embedding layer
    dec_emb_layer = Embedding(y_voc, embedding_dim, trainable=True, name="decoder_embedding")
    dec_emb = dec_emb_layer(decoder_inputs)

    decoder_lstm = LSTM(latent_dim, return_sequences=True, return_state=True,dropout=0.4,recurrent_dropout=0.2, name="decoder_lstm")
    decoder_outputs,decoder_fwd_state, decoder_back_state = decoder_lstm(dec_emb,initial_state=[state_h, state_c])

    #dense layer
    decoder_dense =  Dense(y_voc, activation='softmax', name="decoder_outputs")
    decoder_outputs = decoder_dense(decoder_outputs)

    # Define the model
    model = Model([encoder_inputs, decoder_inputs], decoder_outputs)


    model.summary()

    model.compile(optimizer='rmsprop', loss='sparse_categorical_crossentropy', metrics=["accuracy"])
    # model.compile(optimizer=keras.optimizers.Adam(), loss=keras.losses.SparseCategoricalCrossentropy())

    # model = keras.models.load_model('model.h5')

    return model


def train(model, x_tr,y_tr, x_val, y_val, epochs=50):
    # model, encoder_model, decoder_model, x_tokenizer, y_tokenizer =  load_model()

    x_train_input = x_tr
    y_train_input = y_tr[:, :-1]
    y_train_output = y_tr.reshape(y_tr.shape[0], y_tr.shape[1], 1)[:, 1:]

    print('y_train_input[0]', y_train_input[0])
    print('y_train_output[0]', y_train_output[0])

    x_val_input = x_val
    y_val_input = y_val[:, :-1]
    y_val_output = y_val.reshape(y_val.shape[0], y_val.shape[1], 1)[:, 1:]

    early_stop = EarlyStopping(monitor='val_loss', mode='min', verbose=1, patience=2)
    checkpoint = keras.callbacks.ModelCheckpoint(model_path, monitor='val_loss', save_best_only=True, mode='min', verbose=1)

    callbacks = [ early_stop, checkpoint]

    history = model.fit([x_train_input, y_train_input], y_train_output, epochs=epochs,
                        callbacks=callbacks, batch_size=128, validation_data=([x_val_input, y_val_input], y_val_output))

    model.save(model_path)

    return history

def get_encoder_decoder_model1(model):

    for i, layer in enumerate(model.layers):
        print(f"Layer {i}: {layer.name} - {layer.__class__.__name__}")

    decoder_inputs = model.get_layer("decoder_inputs").input
    encoder_inputs = model.get_layer("encoder_inputs").input

    encoder_outputs, state_h_enc, state_c_enc = model.get_layer("encoder_lstm3").output  # Change index if needed
    decoder_lstm = model.get_layer("decoder_lstm")  # Change index if needed
    decoder_dense = model.get_layer("decoder_outputs") # Change index if needed

    encoder_model = Model(inputs=encoder_inputs, outputs=[encoder_outputs, state_h_enc, state_c_enc])

    decoder_state_input_h = Input(shape=(latent_dim,))
    decoder_state_input_c = Input(shape=(latent_dim,))
    dec_emb2 = model.get_layer("decoder_embedding")(decoder_inputs) # Assuming embedding layer is at index 3, change if needed

    decoder_outputs2, state_h2, state_c2 = decoder_lstm( dec_emb2, initial_state=[decoder_state_input_h, decoder_state_input_c] )

    decoder_outputs2 = decoder_dense(decoder_outputs2)

    decoder_model = Model(
        [decoder_inputs] + [decoder_state_input_h, decoder_state_input_c],
        [decoder_outputs2] + [state_h2, state_c2]
    )



    return decoder_model, encoder_model

def load_model():
    model = keras.models.load_model(model_path)
    decoder_model, encoder_model = get_encoder_decoder_model1(model)

    with open(os.path.join(base_dir, "y_tokenizer.json")) as f:
        data = json.load(f)
        y_tokenizer = tokenizer_from_json(data)

    with open(os.path.join(base_dir, "x_tokenizer.json")) as f:
        data = json.load(f)
        x_tokenizer = tokenizer_from_json(data)

    return model , encoder_model , decoder_model, x_tokenizer, y_tokenizer

def decode_sequence(input_seq, encoder_model, decoder_model,  target_word_index, reverse_target_word_index):

    # Encode the input as state vectors.
    _, e_h, e_c = encoder_model.predict(input_seq)

    # Generate empty target sequence of length 1.
    target_seq = np.zeros((1, 1))

    # Populate the first word of target sequence with the start word.
    target_seq[0, 0] = target_word_index[START_]

    stop_condition = False
    decoded_sentence = ''
    while not stop_condition:

        output_tokens, h, c = decoder_model.predict([target_seq] + [e_h, e_c])

        # Sample a token
        sampled_token_index = np.argmax(output_tokens[0, -1, :])
        print(sampled_token_index)
        if sampled_token_index ==0 :
            sampled_token='0'
        else:
            sampled_token = reverse_target_word_index[sampled_token_index]


        if (sampled_token != END_):
            decoded_sentence += ' ' + sampled_token

        # Exit condition: either hit max length or find stop word.
        if (sampled_token == END_ or len(decoded_sentence.split()) >= (max_summary_len - 1)):
            stop_condition = True

        # Update the target sequence (of length 1).
        target_seq = np.zeros((1, 1))
        target_seq[0, 0] = sampled_token_index

        # Update internal states
        e_h, e_c = h, c

    return decoded_sentence

def seq2summary(input_seq, target_word_index, reverse_target_word_index):
    newString=''
    for i in input_seq:
        if((i!=0 and i!=target_word_index[START_]) and i!=target_word_index[END_]):
            newString=newString+reverse_target_word_index[i]+' '
    return newString

def seq2text(input_seq, reverse_source_word_index):
    newString=''
    for i in input_seq:
        if(i!=0):
            newString=newString+reverse_source_word_index[i]+' '
    return newString


def show_matrix(history):
    # from matplotlib import pyplot
    # pyplot.plot(history.history['loss'], label='train')
    # pyplot.plot(history.history['val_loss'], label='test')
    # pyplot.legend()
    # pyplot.show()
    pass

def test(max_text_len = 100):
    x_tr, y_tr, x_val, y_val, x_tokenizer, y_tokenizer =  get_dataset()
    model , encoder_model , decoder_model, x_tokenizer, y_tokenizer = load_model()
    reverse_target_word_index = y_tokenizer.index_word
    reverse_source_word_index = x_tokenizer.index_word
    target_word_index = y_tokenizer.word_index

    for i in range(5, 50):
        print("Review:", seq2text(x_tr[i], reverse_source_word_index))
        print("Original summary:", seq2summary(y_tr[i], target_word_index, reverse_target_word_index))
        print("Predicted summary:",
              decode_sequence(x_tr[i].reshape(1, max_text_len), encoder_model, decoder_model, target_word_index,
                              reverse_target_word_index))
        print("\n")


if __name__ == '__main__':

    x_voc = 8864
    y_voc = 19131
    epochs = 100

    x_tr, y_tr, x_val, y_val, x_tokenizer, y_tokenizer = get_dataset()

    print(x_tr[0])
    print(y_tr[0])

    model = get_model(max_text_len, x_voc, y_voc)
    history = train(model, x_tr, y_tr, x_val, y_val, epochs)
    show_matrix(history)
    test(max_text_len)

4.训练过程中数据

cpu 上训练时间大概是一整天

X_train: (80000, 6)
y_train: (80000, 10)
X_test: (20000, 6)
y_test: (20000, 10)
[ 30 266  29  12 178   0]
[  1  12  33 355  62  74 112   2   0   0]
Model: "model"
__________________________________________________________________________________________________
 Layer (type)                   Output Shape         Param #     Connected to
==================================================================================================
 encoder_inputs (InputLayer)    [(None, 6)]          0           []

 encoder_embedding (Embedding)  (None, 6, 100)       886400      ['encoder_inputs[0][0]']

 encoder_lstm1 (LSTM)           [(None, 6, 200),     240800      ['encoder_embedding[0][0]']
                                 (None, 200),
                                 (None, 200)]

 decoder_inputs (InputLayer)    [(None, None)]       0           []

 encoder_lstm2 (LSTM)           [(None, 6, 200),     320800      ['encoder_lstm1[0][0]']
                                 (None, 200),
                                 (None, 200)]

 decoder_embedding (Embedding)  (None, None, 100)    1913100     ['decoder_inputs[0][0]']

 encoder_lstm3 (LSTM)           [(None, 6, 200),     320800      ['encoder_lstm2[0][0]']
                                 (None, 200),
                                 (None, 200)]

 decoder_lstm (LSTM)            [(None, None, 200),  240800      ['decoder_embedding[0][0]',
                                 (None, 200),                     'encoder_lstm3[0][1]',
                                 (None, 200)]                     'encoder_lstm3[0][2]']

 decoder_outputs (Dense)        (None, None, 19131)  3845331     ['decoder_lstm[0][0]']

==================================================================================================
Total params: 7,768,031
Trainable params: 7,768,031
Non-trainable params: 0
__________________________________________________________________________________________________
y_train_input[0] [  1  12  33 355  62  74 112   2   0]
y_train_output[0] [[ 12]
 [ 33]
 [355]
 [ 62]
 [ 74]
 [112]
 [  2]
 [  0]
 [  0]]
Epoch 1/100
625/625 [==============================] - ETA: 0s - loss: 3.8344 - accuracy: 0.4682
Epoch 1: val_loss improved from inf to 3.35198, saving model to model

625/625 [==============================] - 523s 824ms/step - loss: 3.8344 - accuracy: 0.4682 - val_loss: 3.3520 - val_accuracy: 0.5184
Epoch 2/100
625/625 [==============================] - ETA: 0s - loss: 3.2284 - accuracy: 0.5280
Epoch 2: val_loss improved from 3.35198 to 3.08983, saving model to model

625/625 [==============================] - 622s 996ms/step - loss: 3.2284 - accuracy: 0.5280 - val_loss: 3.0898 - val_accuracy: 0.5403
Epoch 3/100
625/625 [==============================] - ETA: 0s - loss: 3.0036 - accuracy: 0.5493
Epoch 3: val_loss improved from 3.08983 to 2.89494, saving model to model

625/625 [==============================] - 753s 1s/step - loss: 3.0036 - accuracy: 0.5493 - val_loss: 2.8949 - val_accuracy: 0.5637
Epoch 4/100
625/625 [==============================] - ETA: 0s - loss: 2.8115 - accuracy: 0.5733
Epoch 4: val_loss improved from 2.89494 to 2.70395, saving model to model

625/625 [==============================] - 1114s 2s/step - loss: 2.8115 - accuracy: 0.5733 - val_loss: 2.7039 - val_accuracy: 0.5897
Epoch 5/100
625/625 [==============================] - ETA: 0s - loss: 2.6470 - accuracy: 0.5930
Epoch 5: val_loss improved from 2.70395 to 2.57205, saving model to model

625/625 [==============================] - 1075s 2s/step - loss: 2.6470 - accuracy: 0.5930 - val_loss: 2.5721 - val_accuracy: 0.6046
Epoch 6/100
625/625 [==============================] - ETA: 0s - loss: 2.5275 - accuracy: 0.6063
Epoch 6: val_loss improved from 2.57205 to 2.47006, saving model to model

625/625 [==============================] - 1289s 2s/step - loss: 2.5275 - accuracy: 0.6063 - val_loss: 2.4701 - val_accuracy: 0.6164
Epoch 7/100
625/625 [==============================] - ETA: 0s - loss: 2.4297 - accuracy: 0.6172
Epoch 7: val_loss improved from 2.47006 to 2.37810, saving model to model

625/625 [==============================] - 1403s 2s/step - loss: 2.4297 - accuracy: 0.6172 - val_loss: 2.3781 - val_accuracy: 0.6273
Epoch 8/100
625/625 [==============================] - ETA: 0s - loss: 2.3421 - accuracy: 0.6264
Epoch 8: val_loss improved from 2.37810 to 2.30006, saving model to model

625/625 [==============================] - 1420s 2s/step - loss: 2.3421 - accuracy: 0.6264 - val_loss: 2.3001 - val_accuracy: 0.6363
Epoch 9/100
625/625 [==============================] - ETA: 0s - loss: 2.2645 - accuracy: 0.6353
Epoch 9: val_loss improved from 2.30006 to 2.23482, saving model to model

625/625 [==============================] - 1315s 2s/step - loss: 2.2645 - accuracy: 0.6353 - val_loss: 2.2348 - val_accuracy: 0.6437
Epoch 10/100
625/625 [==============================] - ETA: 0s - loss: 2.1948 - accuracy: 0.6420
Epoch 10: val_loss improved from 2.23482 to 2.17029, saving model to model

625/625 [==============================] - 1338s 2s/step - loss: 2.1948 - accuracy: 0.6420 - val_loss: 2.1703 - val_accuracy: 0.6510
Epoch 11/100
625/625 [==============================] - ETA: 0s - loss: 2.1300 - accuracy: 0.6484
Epoch 11: val_loss improved from 2.17029 to 2.11222, saving model to model

625/625 [==============================] - 1345s 2s/step - loss: 2.1300 - accuracy: 0.6484 - val_loss: 2.1122 - val_accuracy: 0.6572
Epoch 12/100
625/625 [==============================] - ETA: 0s - loss: 2.0684 - accuracy: 0.6553
Epoch 12: val_loss improved from 2.11222 to 2.06084, saving model to model

625/625 [==============================] - 1429s 2s/step - loss: 2.0684 - accuracy: 0.6553 - val_loss: 2.0608 - val_accuracy: 0.6621
Epoch 13/100
625/625 [==============================] - ETA: 0s - loss: 2.0104 - accuracy: 0.6612
Epoch 13: val_loss improved from 2.06084 to 2.00748, saving model to model

625/625 [==============================] - 1236s 2s/step - loss: 2.0104 - accuracy: 0.6612 - val_loss: 2.0075 - val_accuracy: 0.6691
Epoch 14/100
625/625 [==============================] - ETA: 0s - loss: 1.9545 - accuracy: 0.6669
Epoch 14: val_loss improved from 2.00748 to 1.96069, saving model to model

625/625 [==============================] - 1337s 2s/step - loss: 1.9545 - accuracy: 0.6669 - val_loss: 1.9607 - val_accuracy: 0.6737
Epoch 15/100
625/625 [==============================] - ETA: 0s - loss: 1.9030 - accuracy: 0.6721
Epoch 15: val_loss improved from 1.96069 to 1.91793, saving model to model

625/625 [==============================] - 1329s 2s/step - loss: 1.9030 - accuracy: 0.6721 - val_loss: 1.9179 - val_accuracy: 0.6784
Epoch 16/100
625/625 [==============================] - ETA: 0s - loss: 1.8557 - accuracy: 0.6769
Epoch 16: val_loss improved from 1.91793 to 1.87738, saving model to model

625/625 [==============================] - 1336s 2s/step - loss: 1.8557 - accuracy: 0.6769 - val_loss: 1.8774 - val_accuracy: 0.6831
Epoch 17/100
625/625 [==============================] - ETA: 0s - loss: 1.8106 - accuracy: 0.6820
Epoch 17: val_loss improved from 1.87738 to 1.83981, saving model to model

625/625 [==============================] - 1482s 2s/step - loss: 1.8106 - accuracy: 0.6820 - val_loss: 1.8398 - val_accuracy: 0.6875
Epoch 18/100
625/625 [==============================] - ETA: 0s - loss: 1.7678 - accuracy: 0.6865
Epoch 18: val_loss improved from 1.83981 to 1.80432, saving model to model

625/625 [==============================] - 1408s 2s/step - loss: 1.7678 - accuracy: 0.6865 - val_loss: 1.8043 - val_accuracy: 0.6913
Epoch 19/100
625/625 [==============================] - ETA: 0s - loss: 1.7265 - accuracy: 0.6909
Epoch 19: val_loss improved from 1.80432 to 1.77152, saving model to model

625/625 [==============================] - 1336s 2s/step - loss: 1.7265 - accuracy: 0.6909 - val_loss: 1.7715 - val_accuracy: 0.6958
Epoch 20/100
625/625 [==============================] - ETA: 0s - loss: 1.6872 - accuracy: 0.6953
Epoch 20: val_loss improved from 1.77152 to 1.73842, saving model to model

625/625 [==============================] - 1386s 2s/step - loss: 1.6872 - accuracy: 0.6953 - val_loss: 1.7384 - val_accuracy: 0.7000
Epoch 21/100
625/625 [==============================] - ETA: 0s - loss: 1.6492 - accuracy: 0.6996
Epoch 21: val_loss improved from 1.73842 to 1.70509, saving model to model

625/625 [==============================] - 1506s 2s/step - loss: 1.6492 - accuracy: 0.6996 - val_loss: 1.7051 - val_accuracy: 0.7050
Epoch 22/100
625/625 [==============================] - ETA: 0s - loss: 1.6125 - accuracy: 0.7039
Epoch 22: val_loss improved from 1.70509 to 1.67770, saving model to model

625/625 [==============================] - 1492s 2s/step - loss: 1.6125 - accuracy: 0.7039 - val_loss: 1.6777 - val_accuracy: 0.7076
Epoch 23/100
625/625 [==============================] - ETA: 0s - loss: 1.5773 - accuracy: 0.7079
Epoch 23: val_loss improved from 1.67770 to 1.64832, saving model to model

625/625 [==============================] - 1471s 2s/step - loss: 1.5773 - accuracy: 0.7079 - val_loss: 1.6483 - val_accuracy: 0.7108
Epoch 24/100
625/625 [==============================] - ETA: 0s - loss: 1.5442 - accuracy: 0.7121
Epoch 24: val_loss improved from 1.64832 to 1.61894, saving model to model

625/625 [==============================] - 1556s 2s/step - loss: 1.5442 - accuracy: 0.7121 - val_loss: 1.6189 - val_accuracy: 0.7145
Epoch 25/100
625/625 [==============================] - ETA: 0s - loss: 1.5112 - accuracy: 0.7160
Epoch 25: val_loss improved from 1.61894 to 1.59627, saving model to model

625/625 [==============================] - 1467s 2s/step - loss: 1.5112 - accuracy: 0.7160 - val_loss: 1.5963 - val_accuracy: 0.7171
Epoch 26/100
625/625 [==============================] - ETA: 0s - loss: 1.4804 - accuracy: 0.7197
Epoch 26: val_loss improved from 1.59627 to 1.56844, saving model to model

625/625 [==============================] - 1594s 3s/step - loss: 1.4804 - accuracy: 0.7197 - val_loss: 1.5684 - val_accuracy: 0.7220
Epoch 27/100
625/625 [==============================] - ETA: 0s - loss: 1.4507 - accuracy: 0.7234
Epoch 27: val_loss improved from 1.56844 to 1.54407, saving model to model

625/625 [==============================] - 1550s 2s/step - loss: 1.4507 - accuracy: 0.7234 - val_loss: 1.5441 - val_accuracy: 0.7252
Epoch 28/100
625/625 [==============================] - ETA: 0s - loss: 1.4213 - accuracy: 0.7268
Epoch 28: val_loss improved from 1.54407 to 1.52123, saving model to model

625/625 [==============================] - 1529s 2s/step - loss: 1.4213 - accuracy: 0.7268 - val_loss: 1.5212 - val_accuracy: 0.7284
Epoch 29/100
625/625 [==============================] - ETA: 0s - loss: 1.3941 - accuracy: 0.7302
Epoch 29: val_loss improved from 1.52123 to 1.50195, saving model to model

625/625 [==============================] - 1482s 2s/step - loss: 1.3941 - accuracy: 0.7302 - val_loss: 1.5020 - val_accuracy: 0.7308
Epoch 30/100
625/625 [==============================] - ETA: 0s - loss: 1.3673 - accuracy: 0.7337
Epoch 30: val_loss improved from 1.50195 to 1.48183, saving model to model

625/625 [==============================] - 1683s 3s/step - loss: 1.3673 - accuracy: 0.7337 - val_loss: 1.4818 - val_accuracy: 0.7339
Epoch 31/100
625/625 [==============================] - ETA: 0s - loss: 1.3423 - accuracy: 0.7369
Epoch 31: val_loss improved from 1.48183 to 1.46150, saving model to model

625/625 [==============================] - 1621s 3s/step - loss: 1.3423 - accuracy: 0.7369 - val_loss: 1.4615 - val_accuracy: 0.7364
Epoch 32/100
625/625 [==============================] - ETA: 0s - loss: 1.3177 - accuracy: 0.7395
Epoch 32: val_loss improved from 1.46150 to 1.44290, saving model to model

625/625 [==============================] - 1594s 3s/step - loss: 1.3177 - accuracy: 0.7395 - val_loss: 1.4429 - val_accuracy: 0.7392
Epoch 33/100
625/625 [==============================] - ETA: 0s - loss: 1.2939 - accuracy: 0.7428
Epoch 33: val_loss improved from 1.44290 to 1.42519, saving model to model

625/625 [==============================] - 1564s 3s/step - loss: 1.2939 - accuracy: 0.7428 - val_loss: 1.4252 - val_accuracy: 0.7408
Epoch 34/100
625/625 [==============================] - ETA: 0s - loss: 1.2713 - accuracy: 0.7454
Epoch 34: val_loss improved from 1.42519 to 1.40973, saving model to model

625/625 [==============================] - 1410s 2s/step - loss: 1.2713 - accuracy: 0.7454 - val_loss: 1.4097 - val_accuracy: 0.7436
Epoch 35/100
625/625 [==============================] - ETA: 0s - loss: 1.2492 - accuracy: 0.7486
Epoch 35: val_loss improved from 1.40973 to 1.39085, saving model to model

625/625 [==============================] - 1407s 2s/step - loss: 1.2492 - accuracy: 0.7486 - val_loss: 1.3909 - val_accuracy: 0.7464
Epoch 36/100
625/625 [==============================] - ETA: 0s - loss: 1.2286 - accuracy: 0.7511
Epoch 36: val_loss improved from 1.39085 to 1.37574, saving model to model

625/625 [==============================] - 1470s 2s/step - loss: 1.2286 - accuracy: 0.7511 - val_loss: 1.3757 - val_accuracy: 0.7485
Epoch 37/100
625/625 [==============================] - ETA: 0s - loss: 1.2086 - accuracy: 0.7537
Epoch 37: val_loss improved from 1.37574 to 1.36135, saving model to model

625/625 [==============================] - 1561s 2s/step - loss: 1.2086 - accuracy: 0.7537 - val_loss: 1.3613 - val_accuracy: 0.7503
Epoch 38/100
625/625 [==============================] - ETA: 0s - loss: 1.1886 - accuracy: 0.7560
Epoch 38: val_loss improved from 1.36135 to 1.34700, saving model to model

625/625 [==============================] - 1723s 3s/step - loss: 1.1886 - accuracy: 0.7560 - val_loss: 1.3470 - val_accuracy: 0.7519
Epoch 39/100
625/625 [==============================] - ETA: 0s - loss: 1.1690 - accuracy: 0.7591
Epoch 39: val_loss improved from 1.34700 to 1.33595, saving model to model

625/625 [==============================] - 1796s 3s/step - loss: 1.1690 - accuracy: 0.7591 - val_loss: 1.3360 - val_accuracy: 0.7536
Epoch 40/100
625/625 [==============================] - ETA: 0s - loss: 1.1516 - accuracy: 0.7611
Epoch 40: val_loss improved from 1.33595 to 1.32075, saving model to model

625/625 [==============================] - 1883s 3s/step - loss: 1.1516 - accuracy: 0.7611 - val_loss: 1.3208 - val_accuracy: 0.7567
Epoch 41/100
625/625 [==============================] - ETA: 0s - loss: 1.1329 - accuracy: 0.7641
Epoch 41: val_loss improved from 1.32075 to 1.30833, saving model to model

625/625 [==============================] - 1924s 3s/step - loss: 1.1329 - accuracy: 0.7641 - val_loss: 1.3083 - val_accuracy: 0.7582
Epoch 42/100
625/625 [==============================] - ETA: 0s - loss: 1.1154 - accuracy: 0.7661
Epoch 42: val_loss improved from 1.30833 to 1.29805, saving model to model

625/625 [==============================] - 1726s 3s/step - loss: 1.1154 - accuracy: 0.7661 - val_loss: 1.2980 - val_accuracy: 0.7590
Epoch 43/100
625/625 [==============================] - ETA: 0s - loss: 1.0988 - accuracy: 0.7688
Epoch 43: val_loss improved from 1.29805 to 1.28425, saving model to model

625/625 [==============================] - 1588s 3s/step - loss: 1.0988 - accuracy: 0.7688 - val_loss: 1.2843 - val_accuracy: 0.7619
Epoch 44/100
625/625 [==============================] - ETA: 0s - loss: 1.0836 - accuracy: 0.7700
Epoch 44: val_loss improved from 1.28425 to 1.27422, saving model to model

625/625 [==============================] - 1559s 2s/step - loss: 1.0836 - accuracy: 0.7700 - val_loss: 1.2742 - val_accuracy: 0.7623
Epoch 45/100
625/625 [==============================] - ETA: 0s - loss: 1.0658 - accuracy: 0.7731
Epoch 45: val_loss improved from 1.27422 to 1.26191, saving model to model

625/625 [==============================] - 1663s 3s/step - loss: 1.0658 - accuracy: 0.7731 - val_loss: 1.2619 - val_accuracy: 0.7648
Epoch 46/100
625/625 [==============================] - ETA: 0s - loss: 1.0521 - accuracy: 0.7748
Epoch 46: val_loss improved from 1.26191 to 1.25429, saving model to model

625/625 [==============================] - 1829s 3s/step - loss: 1.0521 - accuracy: 0.7748 - val_loss: 1.2543 - val_accuracy: 0.7654
Epoch 47/100
625/625 [==============================] - ETA: 0s - loss: 1.0367 - accuracy: 0.7768
Epoch 47: val_loss improved from 1.25429 to 1.24540, saving model to model

625/625 [==============================] - 1825s 3s/step - loss: 1.0367 - accuracy: 0.7768 - val_loss: 1.2454 - val_accuracy: 0.7665
Epoch 48/100
625/625 [==============================] - ETA: 0s - loss: 1.0228 - accuracy: 0.7788
Epoch 48: val_loss improved from 1.24540 to 1.23448, saving model to model

625/625 [==============================] - 1875s 3s/step - loss: 1.0228 - accuracy: 0.7788 - val_loss: 1.2345 - val_accuracy: 0.7683
Epoch 49/100
625/625 [==============================] - ETA: 0s - loss: 1.0085 - accuracy: 0.7808
Epoch 49: val_loss improved from 1.23448 to 1.22543, saving model to model

625/625 [==============================] - 1919s 3s/step - loss: 1.0085 - accuracy: 0.7808 - val_loss: 1.2254 - val_accuracy: 0.7696
Epoch 50/100
625/625 [==============================] - ETA: 0s - loss: 0.9947 - accuracy: 0.7825
Epoch 50: val_loss improved from 1.22543 to 1.21777, saving model to model

625/625 [==============================] - 1754s 3s/step - loss: 0.9947 - accuracy: 0.7825 - val_loss: 1.2178 - val_accuracy: 0.7700
Epoch 51/100
625/625 [==============================] - ETA: 0s - loss: 0.9813 - accuracy: 0.7848
Epoch 51: val_loss improved from 1.21777 to 1.20699, saving model to model

625/625 [==============================] - 1753s 3s/step - loss: 0.9813 - accuracy: 0.7848 - val_loss: 1.2070 - val_accuracy: 0.7726
Epoch 52/100
625/625 [==============================] - ETA: 0s - loss: 0.9683 - accuracy: 0.7867
Epoch 52: val_loss improved from 1.20699 to 1.19979, saving model to model

625/625 [==============================] - 1705s 3s/step - loss: 0.9683 - accuracy: 0.7867 - val_loss: 1.1998 - val_accuracy: 0.7730
Epoch 53/100
625/625 [==============================] - ETA: 0s - loss: 0.9562 - accuracy: 0.7881
Epoch 53: val_loss improved from 1.19979 to 1.19423, saving model to model

625/625 [==============================] - 1669s 3s/step - loss: 0.9562 - accuracy: 0.7881 - val_loss: 1.1942 - val_accuracy: 0.7740
Epoch 54/100
625/625 [==============================] - ETA: 0s - loss: 0.9433 - accuracy: 0.7902
Epoch 54: val_loss improved from 1.19423 to 1.18470, saving model to model

625/625 [==============================] - 1731s 3s/step - loss: 0.9433 - accuracy: 0.7902 - val_loss: 1.1847 - val_accuracy: 0.7759
Epoch 55/100
625/625 [==============================] - ETA: 0s - loss: 0.9318 - accuracy: 0.7916
Epoch 55: val_loss improved from 1.18470 to 1.17739, saving model to model

625/625 [==============================] - 1711s 3s/step - loss: 0.9318 - accuracy: 0.7916 - val_loss: 1.1774 - val_accuracy: 0.7764
Epoch 56/100
625/625 [==============================] - ETA: 0s - loss: 0.9212 - accuracy: 0.7933
Epoch 56: val_loss improved from 1.17739 to 1.17282, saving model to model

625/625 [==============================] - 1694s 3s/step - loss: 0.9212 - accuracy: 0.7933 - val_loss: 1.1728 - val_accuracy: 0.7773
Epoch 57/100
625/625 [==============================] - ETA: 0s - loss: 0.9108 - accuracy: 0.7944
Epoch 57: val_loss improved from 1.17282 to 1.16557, saving model to model

625/625 [==============================] - 1747s 3s/step - loss: 0.9108 - accuracy: 0.7944 - val_loss: 1.1656 - val_accuracy: 0.7786
Epoch 58/100
625/625 [==============================] - ETA: 0s - loss: 0.8990 - accuracy: 0.7966
Epoch 58: val_loss improved from 1.16557 to 1.16054, saving model to model

625/625 [==============================] - 1629s 3s/step - loss: 0.8990 - accuracy: 0.7966 - val_loss: 1.1605 - val_accuracy: 0.7795
Epoch 59/100
625/625 [==============================] - ETA: 0s - loss: 0.8879 - accuracy: 0.7978
Epoch 59: val_loss improved from 1.16054 to 1.15364, saving model to model

625/625 [==============================] - 1554s 2s/step - loss: 0.8879 - accuracy: 0.7978 - val_loss: 1.1536 - val_accuracy: 0.7797
Epoch 60/100
625/625 [==============================] - ETA: 0s - loss: 0.8777 - accuracy: 0.7996
Epoch 60: val_loss improved from 1.15364 to 1.14653, saving model to model

625/625 [==============================] - 1463s 2s/step - loss: 0.8777 - accuracy: 0.7996 - val_loss: 1.1465 - val_accuracy: 0.7813
Epoch 61/100
625/625 [==============================] - ETA: 0s - loss: 0.8680 - accuracy: 0.8008
Epoch 61: val_loss improved from 1.14653 to 1.14013, saving model to model

625/625 [==============================] - 1493s 2s/step - loss: 0.8680 - accuracy: 0.8008 - val_loss: 1.1401 - val_accuracy: 0.7822
Epoch 62/100
625/625 [==============================] - ETA: 0s - loss: 0.8581 - accuracy: 0.8025
Epoch 62: val_loss improved from 1.14013 to 1.13445, saving model to model

625/625 [==============================] - 1566s 3s/step - loss: 0.8581 - accuracy: 0.8025 - val_loss: 1.1345 - val_accuracy: 0.7835
Epoch 63/100
625/625 [==============================] - ETA: 0s - loss: 0.8482 - accuracy: 0.8041
Epoch 63: val_loss improved from 1.13445 to 1.13095, saving model to model

625/625 [==============================] - 1547s 2s/step - loss: 0.8482 - accuracy: 0.8041 - val_loss: 1.1310 - val_accuracy: 0.7836
Epoch 64/100
625/625 [==============================] - ETA: 0s - loss: 0.8392 - accuracy: 0.8055
Epoch 64: val_loss improved from 1.13095 to 1.12212, saving model to model

625/625 [==============================] - 1585s 3s/step - loss: 0.8392 - accuracy: 0.8055 - val_loss: 1.1221 - val_accuracy: 0.7854
Epoch 65/100
625/625 [==============================] - ETA: 0s - loss: 0.8291 - accuracy: 0.8071
Epoch 65: val_loss improved from 1.12212 to 1.11893, saving model to model

625/625 [==============================] - 1555s 2s/step - loss: 0.8291 - accuracy: 0.8071 - val_loss: 1.1189 - val_accuracy: 0.7851
Epoch 66/100
625/625 [==============================] - ETA: 0s - loss: 0.8204 - accuracy: 0.8083
Epoch 66: val_loss improved from 1.11893 to 1.11534, saving model to model

625/625 [==============================] - 1553s 2s/step - loss: 0.8204 - accuracy: 0.8083 - val_loss: 1.1153 - val_accuracy: 0.7855
Epoch 67/100
625/625 [==============================] - ETA: 0s - loss: 0.8119 - accuracy: 0.8095
Epoch 67: val_loss improved from 1.11534 to 1.11159, saving model to model

625/625 [==============================] - 1587s 3s/step - loss: 0.8119 - accuracy: 0.8095 - val_loss: 1.1116 - val_accuracy: 0.7863
Epoch 68/100
625/625 [==============================] - ETA: 0s - loss: 0.8035 - accuracy: 0.8109
Epoch 68: val_loss improved from 1.11159 to 1.10478, saving model to model

625/625 [==============================] - 1679s 3s/step - loss: 0.8035 - accuracy: 0.8109 - val_loss: 1.1048 - val_accuracy: 0.7874
Epoch 69/100
625/625 [==============================] - ETA: 0s - loss: 0.7956 - accuracy: 0.8119
Epoch 69: val_loss improved from 1.10478 to 1.10363, saving model to model

625/625 [==============================] - 1656s 3s/step - loss: 0.7956 - accuracy: 0.8119 - val_loss: 1.1036 - val_accuracy: 0.7878
Epoch 70/100
625/625 [==============================] - ETA: 0s - loss: 0.7869 - accuracy: 0.8134
Epoch 70: val_loss improved from 1.10363 to 1.09615, saving model to model

625/625 [==============================] - 1660s 3s/step - loss: 0.7869 - accuracy: 0.8134 - val_loss: 1.0962 - val_accuracy: 0.7889
Epoch 71/100
625/625 [==============================] - ETA: 0s - loss: 0.7790 - accuracy: 0.8148
Epoch 71: val_loss improved from 1.09615 to 1.09203, saving model to model

625/625 [==============================] - 1667s 3s/step - loss: 0.7790 - accuracy: 0.8148 - val_loss: 1.0920 - val_accuracy: 0.7894
Epoch 72/100
625/625 [==============================] - ETA: 0s - loss: 0.7707 - accuracy: 0.8164
Epoch 72: val_loss improved from 1.09203 to 1.08892, saving model to model

625/625 [==============================] - 1664s 3s/step - loss: 0.7707 - accuracy: 0.8164 - val_loss: 1.0889 - val_accuracy: 0.7903
Epoch 73/100
625/625 [==============================] - ETA: 0s - loss: 0.7640 - accuracy: 0.8170
Epoch 73: val_loss improved from 1.08892 to 1.08577, saving model to model

625/625 [==============================] - 1701s 3s/step - loss: 0.7640 - accuracy: 0.8170 - val_loss: 1.0858 - val_accuracy: 0.7909
Epoch 74/100
625/625 [==============================] - ETA: 0s - loss: 0.7567 - accuracy: 0.8182
Epoch 74: val_loss improved from 1.08577 to 1.08194, saving model to model

625/625 [==============================] - 1632s 3s/step - loss: 0.7567 - accuracy: 0.8182 - val_loss: 1.0819 - val_accuracy: 0.7915
Epoch 75/100
625/625 [==============================] - ETA: 0s - loss: 0.7482 - accuracy: 0.8195
Epoch 75: val_loss improved from 1.08194 to 1.07837, saving model to model

625/625 [==============================] - 1812s 3s/step - loss: 0.7482 - accuracy: 0.8195 - val_loss: 1.0784 - val_accuracy: 0.7919
Epoch 76/100
625/625 [==============================] - ETA: 0s - loss: 0.7413 - accuracy: 0.8208
Epoch 76: val_loss improved from 1.07837 to 1.07431, saving model to model

625/625 [==============================] - 1680s 3s/step - loss: 0.7413 - accuracy: 0.8208 - val_loss: 1.0743 - val_accuracy: 0.7919
Epoch 77/100
625/625 [==============================] - ETA: 0s - loss: 0.7348 - accuracy: 0.8217
Epoch 77: val_loss improved from 1.07431 to 1.07171, saving model to model

625/625 [==============================] - 1841s 3s/step - loss: 0.7348 - accuracy: 0.8217 - val_loss: 1.0717 - val_accuracy: 0.7922
Epoch 78/100
625/625 [==============================] - ETA: 0s - loss: 0.7276 - accuracy: 0.8231
Epoch 78: val_loss improved from 1.07171 to 1.06717, saving model to model

625/625 [==============================] - 1713s 3s/step - loss: 0.7276 - accuracy: 0.8231 - val_loss: 1.0672 - val_accuracy: 0.7936
Epoch 79/100
625/625 [==============================] - ETA: 0s - loss: 0.7214 - accuracy: 0.8237
Epoch 79: val_loss improved from 1.06717 to 1.06485, saving model to model

625/625 [==============================] - 1678s 3s/step - loss: 0.7214 - accuracy: 0.8237 - val_loss: 1.0648 - val_accuracy: 0.7945
Epoch 80/100
625/625 [==============================] - ETA: 0s - loss: 0.7152 - accuracy: 0.8250
Epoch 80: val_loss improved from 1.06485 to 1.06066, saving model to model

625/625 [==============================] - 1692s 3s/step - loss: 0.7152 - accuracy: 0.8250 - val_loss: 1.0607 - val_accuracy: 0.7942
Epoch 81/100
625/625 [==============================] - ETA: 0s - loss: 0.7088 - accuracy: 0.8258
Epoch 81: val_loss improved from 1.06066 to 1.05814, saving model to model

625/625 [==============================] - 1620s 3s/step - loss: 0.7088 - accuracy: 0.8258 - val_loss: 1.0581 - val_accuracy: 0.7955
Epoch 82/100
625/625 [==============================] - ETA: 0s - loss: 0.7011 - accuracy: 0.8271
Epoch 82: val_loss improved from 1.05814 to 1.05651, saving model to model

625/625 [==============================] - 1587s 3s/step - loss: 0.7011 - accuracy: 0.8271 - val_loss: 1.0565 - val_accuracy: 0.7957
Epoch 83/100
625/625 [==============================] - ETA: 0s - loss: 0.6960 - accuracy: 0.8278
Epoch 83: val_loss improved from 1.05651 to 1.05408, saving model to model

625/625 [==============================] - 1705s 3s/step - loss: 0.6960 - accuracy: 0.8278 - val_loss: 1.0541 - val_accuracy: 0.7956
Epoch 84/100
625/625 [==============================] - ETA: 0s - loss: 0.6901 - accuracy: 0.8287
Epoch 84: val_loss improved from 1.05408 to 1.04978, saving model to model

625/625 [==============================] - 1608s 3s/step - loss: 0.6901 - accuracy: 0.8287 - val_loss: 1.0498 - val_accuracy: 0.7963
Epoch 85/100
625/625 [==============================] - ETA: 0s - loss: 0.6834 - accuracy: 0.8302
Epoch 85: val_loss improved from 1.04978 to 1.04747, saving model to model

625/625 [==============================] - 1574s 3s/step - loss: 0.6834 - accuracy: 0.8302 - val_loss: 1.0475 - val_accuracy: 0.7970
Epoch 86/100
625/625 [==============================] - ETA: 0s - loss: 0.6773 - accuracy: 0.8313
Epoch 86: val_loss improved from 1.04747 to 1.04691, saving model to model

625/625 [==============================] - 1609s 3s/step - loss: 0.6773 - accuracy: 0.8313 - val_loss: 1.0469 - val_accuracy: 0.7963
Epoch 87/100
625/625 [==============================] - ETA: 0s - loss: 0.6718 - accuracy: 0.8319
Epoch 87: val_loss improved from 1.04691 to 1.04322, saving model to model

625/625 [==============================] - 1656s 3s/step - loss: 0.6718 - accuracy: 0.8319 - val_loss: 1.0432 - val_accuracy: 0.7971
Epoch 88/100
625/625 [==============================] - ETA: 0s - loss: 0.6663 - accuracy: 0.8327
Epoch 88: val_loss improved from 1.04322 to 1.04013, saving model to model

625/625 [==============================] - 1749s 3s/step - loss: 0.6663 - accuracy: 0.8327 - val_loss: 1.0401 - val_accuracy: 0.7982
Epoch 89/100
625/625 [==============================] - ETA: 0s - loss: 0.6603 - accuracy: 0.8335
Epoch 89: val_loss improved from 1.04013 to 1.03667, saving model to model

625/625 [==============================] - 1793s 3s/step - loss: 0.6603 - accuracy: 0.8335 - val_loss: 1.0367 - val_accuracy: 0.7981
Epoch 90/100
625/625 [==============================] - ETA: 0s - loss: 0.6550 - accuracy: 0.8348
Epoch 90: val_loss improved from 1.03667 to 1.03568, saving model to model

625/625 [==============================] - 1884s 3s/step - loss: 0.6550 - accuracy: 0.8348 - val_loss: 1.0357 - val_accuracy: 0.7992
Epoch 91/100
104/625 [===>..........................] - ETA: 23:58 - loss: 0.6348 - accuracy: 0.8396

5. 测试

English: that's another issue 
Original Franch: c'est une autre affaire 
Predicted Franch:  c'est une autre affaire


English: we should cut our losses 
Original Franch: nous devrions compenser nos pertes 
Predicted Franch:  nous devrions nous arrêter de nos oreilles


English: jump across 
Original Franch: saute de l'autre côté 
Predicted Franch:  fais de l'autre côté


English: i thought i heard a voice 
Original Franch: j'ai cru entendre une voix 
Predicted Franch:  je pensais entendre une voix


English: let's not go 
Original Franch: n'y allons pas 
Predicted Franch:  ne partons pas


English: does this ring a bell 
Original Franch: ça vous évoque quelque chose 
Predicted Franch:  cela t'évoque quelque chose


English: tom is presumptuous 
Original Franch: tom est présomptueux 
Predicted Franch:  tom est sensible


English: for here or to go 
Original Franch: c'est pour ici ou à emporter  
Predicted Franch:  sur place ou à emporter


English: do what you think is best 
Original Franch: fais ce que tu penses être le mieux 
Predicted Franch:  fais ce que tu penses que c'est mieux


English: i was dazzled 
Original Franch: j'étais éblouie 
Predicted Franch:  j'étais en train de nettoyer


English: i doubt it 
Original Franch: j'en doute 
Predicted Franch:  j'en doute


English: he was sentenced to death 
Original Franch: il fut condamné à mort 
Predicted Franch:  il a été condamné à mort


English: i'm totally confused 
Original Franch: je suis complètement perdue 
Predicted Franch:  je suis complètement confus


English: that wasn't my question 
Original Franch: ce n'était pas ma question 
Predicted Franch:  ce n'était pas ma question


English: what is that smell 
Original Franch: c'est quoi cette odeur 
Predicted Franch:  qu'est ce que ça a t il de l'importance


English: we need an answer 
Original Franch: il nous faut une réponse 
Predicted Franch:  il nous faut une réponse


English: she adores him 
Original Franch: elle le vénère 
Predicted Franch:  elle l'aime


English: they must be fake 
Original Franch: ils doivent être faux 
Predicted Franch:  ils doivent être faux


English: he is mentally handicapped 
Original Franch: il est handicapé mental 
Predicted Franch:  il est américain


English: i'm here by choice 
Original Franch: je suis ici de mon plein gré 
Predicted Franch:  je suis là tout seul


English: tom looks determined 
Original Franch: tom semble déterminé 
Predicted Franch:  tom a l'air déterminé


English: this pen has run dry 
Original Franch: ce stylo n'a plus d'encre 
Predicted Franch:  ce stylo a cessé de lait


English: how did you get hurt 
Original Franch: comment vous êtes vous blessé 
Predicted Franch:  comment t'es tu blessé


English: i'm no dummy 
Original Franch: je ne suis pas abrutie 
Predicted Franch:  je ne suis pas idiot


English: we all cried a lot 
Original Franch: nous avons toutes beaucoup pleuré 
Predicted Franch:  nous avons toutes beaucoup pleuré


English: i've found a better way 
Original Franch: j'ai trouvé un meilleur moyen 
Predicted Franch:  j'ai trouvé un meilleur moyen


English: sorry we're full today 
Original Franch: désolé nous sommes complets aujourd'hui 
Predicted Franch:  désolé nous sommes très occupés aujourd'hui


English: did they say anything 
Original Franch: ont ils dit quelque chose 
Predicted Franch:  ont elles dit quoi que ce soit


English: have you ever been in love 
Original Franch: avez vous jamais été amoureux 
Predicted Franch:  as tu jamais été amoureux


English: here take it 
Original Franch: tiens prends le 
Predicted Franch:  voilà le


English: the wind blew all day 
Original Franch: le vent a soufflé toute la journée 
Predicted Franch:  le vent s'est éclairci toute la journée


English: don't make me choose 
Original Franch: ne m'obligez pas à choisir 
Predicted Franch:  ne m'obligez pas à faire de mes amis


English: we walked to my room 
Original Franch: nous avons marché jusqu'à ma chambre 
Predicted Franch:  nous sommes restés chez moi


English: give it some thought 
Original Franch: réfléchis y un peu 
Predicted Franch:  donne le à le croire


English: i opened the window 
Original Franch: j'ouvris la fenêtre 
Predicted Franch:  j'ai fermé la fenêtre


English: tom is awkward 
Original Franch: tom est maladroit 
Predicted Franch:  tom est blessé


English: i can't wait till summer 
Original Franch: je ne peux pas attendre jusqu'à l'été 
Predicted Franch:  je ne peux pas attendre jusqu'à l'été


English: i'm not lazy 
Original Franch: je ne suis pas paresseux 
Predicted Franch:  je ne suis pas paresseux


English: it's no big deal 
Original Franch: ça n'est pas grave 
Predicted Franch:  ce n'est pas un grand


English: i never was shy 
Original Franch: je n'ai jamais été timide 
Predicted Franch:  je n'ai jamais été timide


English: i need to impress tom 
Original Franch: je dois impressionner tom 
Predicted Franch:  j'ai besoin de tom


English: i want to see your boss 
Original Franch: je veux voir ton patron 
Predicted Franch:  je veux voir ton patron


English: the wind blew all day 
Original Franch: le vent souffla tout le jour 
Predicted Franch:  le vent s'est éclairci toute la journée


English: that umbrella is tom's 
Original Franch: ce parapluie est à tom 
Predicted Franch:  ce parapluie est tom


English: do you mean what you say 
Original Franch: tu dis ça sérieusement  
Predicted Franch:  tu comprends ce que je veux dire