Update app.py
Browse files
app.py
CHANGED
@@ -0,0 +1,166 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import string
|
3 |
+
import re
|
4 |
+
import numpy as np
|
5 |
+
import tensorflow as tf
|
6 |
+
from tensorflow import keras
|
7 |
+
from tensorflow.keras import layers
|
8 |
+
from tensorflow.keras.layers import TextVectorization
|
9 |
+
|
10 |
+
strip_chars = pickle.load(open('strip_chars.pkl', 'rb'))
|
11 |
+
|
12 |
+
vocab_size = 15000
|
13 |
+
sequence_length = 20
|
14 |
+
batch_size = 64
|
15 |
+
|
16 |
+
class TransformerEncoder(layers.Layer):
|
17 |
+
def __init__(self, embed_dim, dense_dim, num_heads, **kwargs):
|
18 |
+
super(TransformerEncoder, self).__init__(**kwargs)
|
19 |
+
self.embed_dim = embed_dim
|
20 |
+
self.dense_dim = dense_dim
|
21 |
+
self.num_heads = num_heads
|
22 |
+
self.attention = layers.MultiHeadAttention(
|
23 |
+
num_heads=num_heads, key_dim=embed_dim
|
24 |
+
)
|
25 |
+
self.dense_proj = keras.Sequential(
|
26 |
+
[layers.Dense(dense_dim, activation="relu"), layers.Dense(embed_dim),]
|
27 |
+
)
|
28 |
+
self.layernorm_1 = layers.LayerNormalization()
|
29 |
+
self.layernorm_2 = layers.LayerNormalization()
|
30 |
+
self.supports_masking = True
|
31 |
+
|
32 |
+
def call(self, inputs, mask=None):
|
33 |
+
if mask is not None:
|
34 |
+
padding_mask = tf.cast(mask[:, tf.newaxis, tf.newaxis, :], dtype="int32")
|
35 |
+
attention_output = self.attention(
|
36 |
+
query=inputs, value=inputs, key=inputs, attention_mask=padding_mask
|
37 |
+
)
|
38 |
+
proj_input = self.layernorm_1(inputs + attention_output)
|
39 |
+
proj_output = self.dense_proj(proj_input)
|
40 |
+
return self.layernorm_2(proj_input + proj_output)
|
41 |
+
|
42 |
+
|
43 |
+
class PositionalEmbedding(layers.Layer):
|
44 |
+
def __init__(self, sequence_length, vocab_size, embed_dim, **kwargs):
|
45 |
+
super(PositionalEmbedding, self).__init__(**kwargs)
|
46 |
+
self.token_embeddings = layers.Embedding(
|
47 |
+
input_dim=vocab_size, output_dim=embed_dim
|
48 |
+
)
|
49 |
+
self.position_embeddings = layers.Embedding(
|
50 |
+
input_dim=sequence_length, output_dim=embed_dim
|
51 |
+
)
|
52 |
+
self.sequence_length = sequence_length
|
53 |
+
self.vocab_size = vocab_size
|
54 |
+
self.embed_dim = embed_dim
|
55 |
+
|
56 |
+
def call(self, inputs):
|
57 |
+
length = tf.shape(inputs)[-1]
|
58 |
+
positions = tf.range(start=0, limit=length, delta=1)
|
59 |
+
embedded_tokens = self.token_embeddings(inputs)
|
60 |
+
embedded_positions = self.position_embeddings(positions)
|
61 |
+
return embedded_tokens + embedded_positions
|
62 |
+
|
63 |
+
def compute_mask(self, inputs, mask=None):
|
64 |
+
return tf.math.not_equal(inputs, 0)
|
65 |
+
|
66 |
+
|
67 |
+
class TransformerDecoder(layers.Layer):
|
68 |
+
def __init__(self, embed_dim, latent_dim, num_heads, **kwargs):
|
69 |
+
super(TransformerDecoder, self).__init__(**kwargs)
|
70 |
+
self.embed_dim = embed_dim
|
71 |
+
self.latent_dim = latent_dim
|
72 |
+
self.num_heads = num_heads
|
73 |
+
self.attention_1 = layers.MultiHeadAttention(
|
74 |
+
num_heads=num_heads, key_dim=embed_dim
|
75 |
+
)
|
76 |
+
self.attention_2 = layers.MultiHeadAttention(
|
77 |
+
num_heads=num_heads, key_dim=embed_dim
|
78 |
+
)
|
79 |
+
self.dense_proj = keras.Sequential(
|
80 |
+
[layers.Dense(latent_dim, activation="relu"), layers.Dense(embed_dim),]
|
81 |
+
)
|
82 |
+
self.layernorm_1 = layers.LayerNormalization()
|
83 |
+
self.layernorm_2 = layers.LayerNormalization()
|
84 |
+
self.layernorm_3 = layers.LayerNormalization()
|
85 |
+
self.supports_masking = True
|
86 |
+
|
87 |
+
def call(self, inputs, encoder_outputs, mask=None):
|
88 |
+
causal_mask = self.get_causal_attention_mask(inputs)
|
89 |
+
if mask is not None:
|
90 |
+
padding_mask = tf.cast(mask[:, tf.newaxis, :], dtype="int32")
|
91 |
+
padding_mask = tf.minimum(padding_mask, causal_mask)
|
92 |
+
|
93 |
+
attention_output_1 = self.attention_1(
|
94 |
+
query=inputs, value=inputs, key=inputs, attention_mask=causal_mask
|
95 |
+
)
|
96 |
+
out_1 = self.layernorm_1(inputs + attention_output_1)
|
97 |
+
|
98 |
+
attention_output_2 = self.attention_2(
|
99 |
+
query=out_1,
|
100 |
+
value=encoder_outputs,
|
101 |
+
key=encoder_outputs,
|
102 |
+
attention_mask=padding_mask,
|
103 |
+
)
|
104 |
+
out_2 = self.layernorm_2(out_1 + attention_output_2)
|
105 |
+
|
106 |
+
proj_output = self.dense_proj(out_2)
|
107 |
+
return self.layernorm_3(out_2 + proj_output)
|
108 |
+
|
109 |
+
def get_causal_attention_mask(self, inputs):
|
110 |
+
input_shape = tf.shape(inputs)
|
111 |
+
batch_size, sequence_length = input_shape[0], input_shape[1]
|
112 |
+
i = tf.range(sequence_length)[:, tf.newaxis]
|
113 |
+
j = tf.range(sequence_length)
|
114 |
+
mask = tf.cast(i >= j, dtype="int32")
|
115 |
+
mask = tf.reshape(mask, (1, input_shape[1], input_shape[1]))
|
116 |
+
mult = tf.concat(
|
117 |
+
[tf.expand_dims(batch_size, -1), tf.constant([1, 1], dtype=tf.int32)],
|
118 |
+
axis=0,
|
119 |
+
)
|
120 |
+
return tf.tile(mask, mult)
|
121 |
+
|
122 |
+
custom_objects={'TransformerEncoder': TransformerEncoder, 'TransformerDecoder': TransformerDecoder, 'PositionalEmbedding':PositionalEmbedding}
|
123 |
+
transformer = keras.models.load_model("model.h5", custom_objects=custom_objects)
|
124 |
+
|
125 |
+
def custom_standardization(input_string):
|
126 |
+
lowercase = tf.strings.lower(input_string)
|
127 |
+
return tf.strings.regex_replace(lowercase, "[%s]" % re.escape(strip_chars), "")
|
128 |
+
|
129 |
+
|
130 |
+
eng_vectorization = TextVectorization(
|
131 |
+
max_tokens=vocab_size, output_mode="int", output_sequence_length=sequence_length,
|
132 |
+
)
|
133 |
+
spa_vectorization = TextVectorization(
|
134 |
+
max_tokens=vocab_size,
|
135 |
+
output_mode="int",
|
136 |
+
output_sequence_length=sequence_length + 1,
|
137 |
+
standardize=custom_standardization,
|
138 |
+
)
|
139 |
+
train_eng_texts = [pair[0] for pair in train_pairs]
|
140 |
+
train_spa_texts = [pair[1] for pair in train_pairs]
|
141 |
+
eng_vectorization.adapt(train_eng_texts)
|
142 |
+
spa_vectorization.adapt(train_spa_texts)
|
143 |
+
|
144 |
+
inputs = gr.inputs.Textbox(lines=1, label="Text in Spanish")
|
145 |
+
outputs = [gr.outputs.Textbox(label="Translated text in Quechua")]
|
146 |
+
|
147 |
+
def get_translate(input_sentence):
|
148 |
+
spa_vocab = spa_vectorization.get_vocabulary()
|
149 |
+
spa_index_lookup = dict(zip(range(len(spa_vocab)), spa_vocab))
|
150 |
+
max_decoded_sentence_length = 20
|
151 |
+
tokenized_input_sentence = eng_vectorization([input_sentence])
|
152 |
+
decoded_sentence = "[start]"
|
153 |
+
for i in range(max_decoded_sentence_length):
|
154 |
+
tokenized_target_sentence = spa_vectorization([decoded_sentence])[:, :-1]
|
155 |
+
predictions = transformer([tokenized_input_sentence, tokenized_target_sentence])
|
156 |
+
|
157 |
+
sampled_token_index = np.argmax(predictions[0, i, :])
|
158 |
+
sampled_token = spa_index_lookup[sampled_token_index]
|
159 |
+
decoded_sentence += " " + sampled_token
|
160 |
+
|
161 |
+
if sampled_token == "[end]":
|
162 |
+
break
|
163 |
+
return decoded_sentence.replace("[start]", "").replace("[end]", "")
|
164 |
+
iface=gr.Interface(fn=get_translate,inputs=inputs , outputs=outputs, title='Sakil EnglishToGerman Translator APP')
|
165 |
+
|
166 |
+
iface.launch(debug=True)
|