Spaces:
Runtime error
Runtime error
Update Models Function
Browse files
model.py
CHANGED
@@ -10,7 +10,7 @@ EMBEDDING_DIM = 512
|
|
10 |
UNITS = 512
|
11 |
|
12 |
|
13 |
-
#
|
14 |
vocab = pickle.load(open('vocabulary/vocab_coco.file', 'rb'))
|
15 |
|
16 |
tokenizer = tf.keras.layers.TextVectorization(
|
@@ -25,6 +25,234 @@ idx2word = tf.keras.layers.StringLookup(
|
|
25 |
invert = True
|
26 |
)
|
27 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
28 |
def load_image_from_path(img_path):
|
29 |
img = tf.io.read_file(img_path)
|
30 |
img = tf.io.decode_jpeg(img, channels=3)
|
|
|
10 |
UNITS = 512
|
11 |
|
12 |
|
13 |
+
#LOAD VOCAB FOLDER
|
14 |
vocab = pickle.load(open('vocabulary/vocab_coco.file', 'rb'))
|
15 |
|
16 |
tokenizer = tf.keras.layers.TextVectorization(
|
|
|
25 |
invert = True
|
26 |
)
|
27 |
|
28 |
+
# CREATING MODEL BASED ON KERAS
|
29 |
+
def CNN_Encoder():
|
30 |
+
inception_v3 = tf.keras.applications.InceptionV3(
|
31 |
+
include_top=False,
|
32 |
+
weights='imagenet'
|
33 |
+
)
|
34 |
+
|
35 |
+
output = inception_v3.output
|
36 |
+
output = tf.keras.layers.Reshape(
|
37 |
+
(-1, output.shape[-1]))(output)
|
38 |
+
|
39 |
+
cnn_model = tf.keras.models.Model(inception_v3.input, output)
|
40 |
+
return cnn_model
|
41 |
+
|
42 |
+
|
43 |
+
class TransformerEncoderLayer(tf.keras.layers.Layer):
|
44 |
+
|
45 |
+
def __init__(self, embed_dim, num_heads):
|
46 |
+
super().__init__()
|
47 |
+
self.layer_norm_1 = tf.keras.layers.LayerNormalization()
|
48 |
+
self.layer_norm_2 = tf.keras.layers.LayerNormalization()
|
49 |
+
self.attention = tf.keras.layers.MultiHeadAttention(
|
50 |
+
num_heads=num_heads, key_dim=embed_dim)
|
51 |
+
self.dense = tf.keras.layers.Dense(embed_dim, activation="relu")
|
52 |
+
|
53 |
+
|
54 |
+
def call(self, x, training):
|
55 |
+
x = self.layer_norm_1(x)
|
56 |
+
x = self.dense(x)
|
57 |
+
|
58 |
+
attn_output = self.attention(
|
59 |
+
query=x,
|
60 |
+
value=x,
|
61 |
+
key=x,
|
62 |
+
attention_mask=None,
|
63 |
+
training=training
|
64 |
+
)
|
65 |
+
|
66 |
+
x = self.layer_norm_2(x + attn_output)
|
67 |
+
return x
|
68 |
+
|
69 |
+
|
70 |
+
class Embeddings(tf.keras.layers.Layer):
|
71 |
+
|
72 |
+
def __init__(self, vocab_size, embed_dim, max_len):
|
73 |
+
super().__init__()
|
74 |
+
self.token_embeddings = tf.keras.layers.Embedding(
|
75 |
+
vocab_size, embed_dim)
|
76 |
+
self.position_embeddings = tf.keras.layers.Embedding(
|
77 |
+
max_len, embed_dim, input_shape=(None, max_len))
|
78 |
+
|
79 |
+
|
80 |
+
def call(self, input_ids):
|
81 |
+
length = tf.shape(input_ids)[-1]
|
82 |
+
position_ids = tf.range(start=0, limit=length, delta=1)
|
83 |
+
position_ids = tf.expand_dims(position_ids, axis=0)
|
84 |
+
|
85 |
+
token_embeddings = self.token_embeddings(input_ids)
|
86 |
+
position_embeddings = self.position_embeddings(position_ids)
|
87 |
+
|
88 |
+
return token_embeddings + position_embeddings
|
89 |
+
|
90 |
+
|
91 |
+
class TransformerDecoderLayer(tf.keras.layers.Layer):
|
92 |
+
|
93 |
+
def __init__(self, embed_dim, units, num_heads):
|
94 |
+
super().__init__()
|
95 |
+
self.embedding = Embeddings(
|
96 |
+
tokenizer.vocabulary_size(), embed_dim, MAX_LENGTH)
|
97 |
+
|
98 |
+
self.attention_1 = tf.keras.layers.MultiHeadAttention(
|
99 |
+
num_heads=num_heads, key_dim=embed_dim, dropout=0.1
|
100 |
+
)
|
101 |
+
self.attention_2 = tf.keras.layers.MultiHeadAttention(
|
102 |
+
num_heads=num_heads, key_dim=embed_dim, dropout=0.1
|
103 |
+
)
|
104 |
+
|
105 |
+
self.layernorm_1 = tf.keras.layers.LayerNormalization()
|
106 |
+
self.layernorm_2 = tf.keras.layers.LayerNormalization()
|
107 |
+
self.layernorm_3 = tf.keras.layers.LayerNormalization()
|
108 |
+
|
109 |
+
self.ffn_layer_1 = tf.keras.layers.Dense(units, activation="relu")
|
110 |
+
self.ffn_layer_2 = tf.keras.layers.Dense(embed_dim)
|
111 |
+
|
112 |
+
self.out = tf.keras.layers.Dense(tokenizer.vocabulary_size(), activation="softmax")
|
113 |
+
|
114 |
+
self.dropout_1 = tf.keras.layers.Dropout(0.3)
|
115 |
+
self.dropout_2 = tf.keras.layers.Dropout(0.5)
|
116 |
+
|
117 |
+
|
118 |
+
def call(self, input_ids, encoder_output, training, mask=None):
|
119 |
+
embeddings = self.embedding(input_ids)
|
120 |
+
|
121 |
+
combined_mask = None
|
122 |
+
padding_mask = None
|
123 |
+
|
124 |
+
if mask is not None:
|
125 |
+
causal_mask = self.get_causal_attention_mask(embeddings)
|
126 |
+
padding_mask = tf.cast(mask[:, :, tf.newaxis], dtype=tf.int32)
|
127 |
+
combined_mask = tf.cast(mask[:, tf.newaxis, :], dtype=tf.int32)
|
128 |
+
combined_mask = tf.minimum(combined_mask, causal_mask)
|
129 |
+
|
130 |
+
attn_output_1 = self.attention_1(
|
131 |
+
query=embeddings,
|
132 |
+
value=embeddings,
|
133 |
+
key=embeddings,
|
134 |
+
attention_mask=combined_mask,
|
135 |
+
training=training
|
136 |
+
)
|
137 |
+
|
138 |
+
out_1 = self.layernorm_1(embeddings + attn_output_1)
|
139 |
+
|
140 |
+
attn_output_2 = self.attention_2(
|
141 |
+
query=out_1,
|
142 |
+
value=encoder_output,
|
143 |
+
key=encoder_output,
|
144 |
+
attention_mask=padding_mask,
|
145 |
+
training=training
|
146 |
+
)
|
147 |
+
|
148 |
+
out_2 = self.layernorm_2(out_1 + attn_output_2)
|
149 |
+
|
150 |
+
ffn_out = self.ffn_layer_1(out_2)
|
151 |
+
ffn_out = self.dropout_1(ffn_out, training=training)
|
152 |
+
ffn_out = self.ffn_layer_2(ffn_out)
|
153 |
+
|
154 |
+
ffn_out = self.layernorm_3(ffn_out + out_2)
|
155 |
+
ffn_out = self.dropout_2(ffn_out, training=training)
|
156 |
+
preds = self.out(ffn_out)
|
157 |
+
return preds
|
158 |
+
|
159 |
+
|
160 |
+
def get_causal_attention_mask(self, inputs):
|
161 |
+
input_shape = tf.shape(inputs)
|
162 |
+
batch_size, sequence_length = input_shape[0], input_shape[1]
|
163 |
+
i = tf.range(sequence_length)[:, tf.newaxis]
|
164 |
+
j = tf.range(sequence_length)
|
165 |
+
mask = tf.cast(i >= j, dtype="int32")
|
166 |
+
mask = tf.reshape(mask, (1, input_shape[1], input_shape[1]))
|
167 |
+
mult = tf.concat(
|
168 |
+
[tf.expand_dims(batch_size, -1), tf.constant([1, 1], dtype=tf.int32)],
|
169 |
+
axis=0
|
170 |
+
)
|
171 |
+
return tf.tile(mask, mult)
|
172 |
+
|
173 |
+
|
174 |
+
class ImageCaptioningModel(tf.keras.Model):
|
175 |
+
|
176 |
+
def __init__(self, cnn_model, encoder, decoder, image_aug=None):
|
177 |
+
super().__init__()
|
178 |
+
self.cnn_model = cnn_model
|
179 |
+
self.encoder = encoder
|
180 |
+
self.decoder = decoder
|
181 |
+
self.image_aug = image_aug
|
182 |
+
self.loss_tracker = tf.keras.metrics.Mean(name="loss")
|
183 |
+
self.acc_tracker = tf.keras.metrics.Mean(name="accuracy")
|
184 |
+
|
185 |
+
|
186 |
+
def calculate_loss(self, y_true, y_pred, mask):
|
187 |
+
loss = self.loss(y_true, y_pred)
|
188 |
+
mask = tf.cast(mask, dtype=loss.dtype)
|
189 |
+
loss *= mask
|
190 |
+
return tf.reduce_sum(loss) / tf.reduce_sum(mask)
|
191 |
+
|
192 |
+
|
193 |
+
def calculate_accuracy(self, y_true, y_pred, mask):
|
194 |
+
accuracy = tf.equal(y_true, tf.argmax(y_pred, axis=2))
|
195 |
+
accuracy = tf.math.logical_and(mask, accuracy)
|
196 |
+
accuracy = tf.cast(accuracy, dtype=tf.float32)
|
197 |
+
mask = tf.cast(mask, dtype=tf.float32)
|
198 |
+
return tf.reduce_sum(accuracy) / tf.reduce_sum(mask)
|
199 |
+
|
200 |
+
|
201 |
+
def compute_loss_and_acc(self, img_embed, captions, training=True):
|
202 |
+
encoder_output = self.encoder(img_embed, training=True)
|
203 |
+
y_input = captions[:, :-1]
|
204 |
+
y_true = captions[:, 1:]
|
205 |
+
mask = (y_true != 0)
|
206 |
+
y_pred = self.decoder(
|
207 |
+
y_input, encoder_output, training=True, mask=mask
|
208 |
+
)
|
209 |
+
loss = self.calculate_loss(y_true, y_pred, mask)
|
210 |
+
acc = self.calculate_accuracy(y_true, y_pred, mask)
|
211 |
+
return loss, acc
|
212 |
+
|
213 |
+
|
214 |
+
def train_step(self, batch):
|
215 |
+
imgs, captions = batch
|
216 |
+
|
217 |
+
if self.image_aug:
|
218 |
+
imgs = self.image_aug(imgs)
|
219 |
+
|
220 |
+
img_embed = self.cnn_model(imgs)
|
221 |
+
|
222 |
+
with tf.GradientTape() as tape:
|
223 |
+
loss, acc = self.compute_loss_and_acc(
|
224 |
+
img_embed, captions
|
225 |
+
)
|
226 |
+
|
227 |
+
train_vars = (
|
228 |
+
self.encoder.trainable_variables + self.decoder.trainable_variables
|
229 |
+
)
|
230 |
+
grads = tape.gradient(loss, train_vars)
|
231 |
+
self.optimizer.apply_gradients(zip(grads, train_vars))
|
232 |
+
self.loss_tracker.update_state(loss)
|
233 |
+
self.acc_tracker.update_state(acc)
|
234 |
+
|
235 |
+
return {"loss": self.loss_tracker.result(), "acc": self.acc_tracker.result()}
|
236 |
+
|
237 |
+
|
238 |
+
def test_step(self, batch):
|
239 |
+
imgs, captions = batch
|
240 |
+
|
241 |
+
img_embed = self.cnn_model(imgs)
|
242 |
+
|
243 |
+
loss, acc = self.compute_loss_and_acc(
|
244 |
+
img_embed, captions, training=False
|
245 |
+
)
|
246 |
+
|
247 |
+
self.loss_tracker.update_state(loss)
|
248 |
+
self.acc_tracker.update_state(acc)
|
249 |
+
|
250 |
+
return {"loss": self.loss_tracker.result(), "acc": self.acc_tracker.result()}
|
251 |
+
|
252 |
+
@property
|
253 |
+
def metrics(self):
|
254 |
+
return [self.loss_tracker, self.acc_tracker]
|
255 |
+
|
256 |
def load_image_from_path(img_path):
|
257 |
img = tf.io.read_file(img_path)
|
258 |
img = tf.io.decode_jpeg(img, channels=3)
|