|
import tensorflow as tf |
|
from tensorflow import keras |
|
from transformer import TransformerEncoder, TransformerDecoderLayer, get_look_ahead_mask, get_padding_mask |
|
|
|
class TFPort(keras.Model): |
|
def __init__(self, crop_size, embed_dim, encoder_layers, decoder_layers, input_seq_length, |
|
target_seq_length, hidden_dim, num_heads, input_vocab_size, dropout_rate=0.1): |
|
super().__init__() |
|
|
|
self.image_encoder = keras.applications.InceptionV3( |
|
include_top=False, |
|
weights='imagenet', |
|
input_shape=crop_size + (3,), |
|
) |
|
self.image_encoder.trainable = False |
|
self.conv = keras.layers.Conv2D(embed_dim, 1) |
|
self.ingredient_encoder = TransformerEncoder(encoder_layers, input_seq_length, embed_dim, hidden_dim, |
|
num_heads, input_vocab_size, dropout_rate=dropout_rate) |
|
self.portion_embedding = PortionEmbedding(target_seq_length, embed_dim) |
|
self.dropout = keras.layers.Dropout(dropout_rate) |
|
self.decoder_layers = [ |
|
TransformerDecoderLayer(embed_dim, hidden_dim, num_heads, dropout_rate=dropout_rate) |
|
for _ in range(decoder_layers) |
|
] |
|
self.linear = keras.layers.Dense(1, activation="relu") |
|
|
|
def call(self, inputs, training=False): |
|
image, ingredients, targets = inputs |
|
padding_mask = get_padding_mask(ingredients) |
|
encoded_img = self.image_encoder(image, training=False) |
|
encoded_img = self.conv(encoded_img, training=training) |
|
encoded_img = tf.reshape(encoded_img, (tf.shape(encoded_img)[0], -1, tf.shape(encoded_img)[3])) |
|
encoded_ingr = self.ingredient_encoder(ingredients, padding_mask, training=training) |
|
encoder_outputs = tf.concat([encoded_img, encoded_ingr], axis=1) |
|
|
|
img_mask = tf.ones((tf.shape(encoded_img)[0], 1, tf.shape(encoded_img)[1]), dtype=tf.int32) |
|
padding_mask = tf.concat([img_mask, padding_mask], axis=2) |
|
look_ahead_mask = get_look_ahead_mask(targets) |
|
|
|
x = self.portion_embedding(targets) |
|
x = self.dropout(x, training=training) |
|
for i in range(len(self.decoder_layers)): |
|
x = self.decoder_layers[i](x, encoder_outputs, look_ahead_mask, training=training, |
|
padding_mask=padding_mask) |
|
x = self.linear(x) |
|
return tf.squeeze(x) |
|
|
|
|
|
class PortionEmbedding(keras.layers.Layer): |
|
def __init__(self, seq_length, embed_dim): |
|
super().__init__() |
|
|
|
self.linear = keras.layers.Dense(embed_dim) |
|
self.position_embeddings = keras.layers.Embedding( |
|
input_dim=seq_length, output_dim=embed_dim |
|
) |
|
|
|
def call(self, inputs): |
|
positions = tf.range(start=0, limit=tf.shape(inputs)[-1], delta=1) |
|
embedded_portions = self.linear(inputs[:, :, tf.newaxis]) |
|
embedded_positions = self.position_embeddings(positions) |
|
return embedded_portions + embedded_positions |
|
|