File size: 2,901 Bytes
0d7f4b7 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 |
__author__ = 'Ferdiand John Briones, attempt at pix2code2 through pretrained autoencoders'
from keras.layers import Input, Dense, Dropout, RepeatVector, LSTM, concatenate, Flatten
from keras.models import Sequential, Model
# from keras.optimizers import RMSprop
from tensorflow.keras.optimizers import RMSprop
from keras import *
from .Config import *
from .AModel import *
from .autoencoder_image import *
class pix2code2(AModel):
def __init__(self, input_shape, output_size, output_path):
AModel.__init__(self, input_shape, output_size, output_path)
self.name = "pix2code2"
visual_input = Input(shape=input_shape)
#Load the pre-trained autoencoder model
autoencoder_model = autoencoder_image(input_shape, input_shape, output_path)
autoencoder_model.load('autoencoder')
autoencoder_model.model.load_weights('../bin/autoencoder.h5')
#Get only the model up to the encoded part
hidden_layer_model_freeze = Model(inputs=autoencoder_model.model.input, outputs=autoencoder_model.model.get_layer('encoded_layer').output)
hidden_layer_input = hidden_layer_model_freeze(visual_input)
#Additional layers before concatenation
hidden_layer_model = Flatten()(hidden_layer_input)
hidden_layer_model = Dense(1024, activation='relu')(hidden_layer_model)
hidden_layer_model = Dropout(0.3)(hidden_layer_model)
hidden_layer_model = Dense(1024, activation='relu')(hidden_layer_model)
hidden_layer_model = Dropout(0.3)(hidden_layer_model)
hidden_layer_result = RepeatVector(CONTEXT_LENGTH)(hidden_layer_model)
#Make sure the loaded hidden_layer_model_freeze will no longer be updated
for layer in hidden_layer_model_freeze.layers:
layer.trainable = False
#The same language model that of pix2code by Tony Beltramelli
language_model = Sequential()
language_model.add(LSTM(128, return_sequences=True, input_shape=(CONTEXT_LENGTH, output_size)))
language_model.add(LSTM(128, return_sequences=True))
textual_input = Input(shape=(CONTEXT_LENGTH, output_size))
encoded_text = language_model(textual_input)
decoder = concatenate([hidden_layer_result, encoded_text])
decoder = LSTM(512, return_sequences=True)(decoder)
decoder = LSTM(512, return_sequences=False)(decoder)
decoder = Dense(output_size, activation='softmax')(decoder)
self.model = Model(inputs=[visual_input, textual_input], outputs=decoder)
optimizer = RMSprop(lr=0.0001, clipvalue=1.0)
self.model.compile(loss='categorical_crossentropy', optimizer=optimizer)
def fit_generator(self, generator, steps_per_epoch):
self.model.summary()
self.model.fit_generator(generator, steps_per_epoch=steps_per_epoch, epochs=EPOCHS, verbose=1)
self.save()
def predict(self, image, partial_caption):
return self.model.predict([image, partial_caption], verbose=0)[0]
def predict_batch(self, images, partial_captions):
return self.model.predict([images, partial_captions], verbose=1) |