Spaces:
Runtime error
Runtime error
Steven Zhang
commited on
Commit
•
ffd9653
1
Parent(s):
435f2f4
chinese update
Browse files- .gitignore +5 -1
- TestTranslationChinese/translation_model.py +252 -0
- app.py +14 -6
.gitignore
CHANGED
@@ -129,4 +129,8 @@ dmypy.json
|
|
129 |
.pyre/
|
130 |
|
131 |
# my own ckpts from gdown
|
132 |
-
EngToSpanishckpts/
|
|
|
|
|
|
|
|
|
|
129 |
.pyre/
|
130 |
|
131 |
# my own ckpts from gdown
|
132 |
+
EngToSpanishckpts/
|
133 |
+
cmn.txt
|
134 |
+
re-model.h5
|
135 |
+
TestTranslationChinese/cmn.txt
|
136 |
+
TestTranslationChinese/re-model.h5
|
TestTranslationChinese/translation_model.py
ADDED
@@ -0,0 +1,252 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# -*- coding: utf-8 -*-
|
2 |
+
"""Translation Model.ipynb
|
3 |
+
|
4 |
+
Automatically generated by Colaboratory.
|
5 |
+
|
6 |
+
Original file is located at
|
7 |
+
https://colab.research.google.com/drive/1njNMtQLmXo_zVtAKxA91dZJh_bxlLume
|
8 |
+
"""
|
9 |
+
|
10 |
+
import pathlib
|
11 |
+
import random
|
12 |
+
import string
|
13 |
+
import h5py
|
14 |
+
import re
|
15 |
+
import numpy as np
|
16 |
+
import tensorflow as tf
|
17 |
+
from tensorflow import keras
|
18 |
+
from tensorflow.keras import layers
|
19 |
+
from tensorflow.keras.layers.experimental.preprocessing import TextVectorization
|
20 |
+
|
21 |
+
import gdown
|
22 |
+
|
23 |
+
url = "https://drive.google.com/uc?id=1FOC2x5HlgcFTMgnGhPjvLWWlEqVTLQno"
|
24 |
+
gdown.download(url, quiet=False)
|
25 |
+
|
26 |
+
with open('cmn.txt', encoding="utf-8") as f:
|
27 |
+
lines = f.read().split("\n")[:-1]
|
28 |
+
text_pairs = []
|
29 |
+
for line in lines:
|
30 |
+
eng, cmn, o1 = line.split("\t")
|
31 |
+
text_pairs.append((eng, cmn))
|
32 |
+
|
33 |
+
random.shuffle(text_pairs)
|
34 |
+
num_val_samples = int(0.15 * len(text_pairs))
|
35 |
+
num_train_samples = len(text_pairs) - 2 * num_val_samples
|
36 |
+
train_pairs = text_pairs[:num_train_samples]
|
37 |
+
val_pairs = text_pairs[num_train_samples : num_train_samples + num_val_samples]
|
38 |
+
test_pairs = text_pairs[num_train_samples + num_val_samples :]
|
39 |
+
|
40 |
+
strip_chars = string.punctuation + "¿"
|
41 |
+
strip_chars = strip_chars.replace("[", "")
|
42 |
+
strip_chars = strip_chars.replace("]", "")
|
43 |
+
|
44 |
+
vocab_size = 15000
|
45 |
+
sequence_length = 20
|
46 |
+
batch_size = 64
|
47 |
+
|
48 |
+
eng_vectorization = TextVectorization(
|
49 |
+
max_tokens=vocab_size,
|
50 |
+
output_mode="int",
|
51 |
+
output_sequence_length=sequence_length,
|
52 |
+
)
|
53 |
+
cmn_vectorization = TextVectorization(
|
54 |
+
max_tokens=vocab_size,
|
55 |
+
output_mode="int",
|
56 |
+
split='character',
|
57 |
+
output_sequence_length=sequence_length + 1,
|
58 |
+
standardize='strip_punctuation',
|
59 |
+
)
|
60 |
+
train_eng_texts = [pair[0] for pair in train_pairs]
|
61 |
+
train_cmn_texts = [pair[1] for pair in train_pairs]
|
62 |
+
eng_vectorization.adapt(train_eng_texts)
|
63 |
+
cmn_vectorization.adapt(train_cmn_texts)
|
64 |
+
|
65 |
+
def format_dataset(eng, cmn):
|
66 |
+
eng = eng_vectorization(eng)
|
67 |
+
cmn = cmn_vectorization(cmn)
|
68 |
+
return (
|
69 |
+
{
|
70 |
+
"encoder_inputs": eng,
|
71 |
+
"decoder_inputs": cmn[:, :-1],
|
72 |
+
},
|
73 |
+
cmn[:, 1:],
|
74 |
+
)
|
75 |
+
|
76 |
+
|
77 |
+
def make_dataset(pairs):
|
78 |
+
eng_texts, cmn_texts = zip(*pairs)
|
79 |
+
eng_texts = list(eng_texts)
|
80 |
+
cmn_texts = list(cmn_texts)
|
81 |
+
dataset = tf.data.Dataset.from_tensor_slices((eng_texts, cmn_texts))
|
82 |
+
dataset = dataset.batch(batch_size)
|
83 |
+
dataset = dataset.map(format_dataset)
|
84 |
+
return dataset.shuffle(2048).prefetch(16).cache()
|
85 |
+
|
86 |
+
|
87 |
+
train_ds = make_dataset(train_pairs)
|
88 |
+
val_ds = make_dataset(val_pairs)
|
89 |
+
|
90 |
+
class TransformerEncoder(layers.Layer):
|
91 |
+
def __init__(self, embed_dim, dense_dim, num_heads, **kwargs):
|
92 |
+
super(TransformerEncoder, self).__init__(**kwargs)
|
93 |
+
self.embed_dim = embed_dim
|
94 |
+
self.dense_dim = dense_dim
|
95 |
+
self.num_heads = num_heads
|
96 |
+
self.attention = layers.MultiHeadAttention(
|
97 |
+
num_heads=num_heads, key_dim=embed_dim
|
98 |
+
)
|
99 |
+
self.dense_proj = keras.Sequential(
|
100 |
+
[
|
101 |
+
layers.Dense(dense_dim, activation="relu"),
|
102 |
+
layers.Dense(embed_dim),
|
103 |
+
]
|
104 |
+
)
|
105 |
+
self.layernorm_1 = layers.LayerNormalization()
|
106 |
+
self.layernorm_2 = layers.LayerNormalization()
|
107 |
+
self.supports_masking = True
|
108 |
+
|
109 |
+
def call(self, inputs, mask=None):
|
110 |
+
if mask is not None:
|
111 |
+
padding_mask = tf.cast(mask[:, tf.newaxis, tf.newaxis, :], dtype="int32")
|
112 |
+
attention_output = self.attention(
|
113 |
+
query=inputs, value=inputs, key=inputs, attention_mask=padding_mask
|
114 |
+
)
|
115 |
+
proj_input = self.layernorm_1(inputs + attention_output)
|
116 |
+
proj_output = self.dense_proj(proj_input)
|
117 |
+
return self.layernorm_2(proj_input + proj_output)
|
118 |
+
def get_config(self):
|
119 |
+
config = super().get_config()
|
120 |
+
config.update({
|
121 |
+
"embed_dim": self.embed_dim,
|
122 |
+
"dense_dim": self.dense_dim,
|
123 |
+
"num_heads": self.num_heads,
|
124 |
+
})
|
125 |
+
return config
|
126 |
+
|
127 |
+
|
128 |
+
class PositionalEmbedding(layers.Layer):
|
129 |
+
def __init__(self, sequence_length, vocab_size, embed_dim, **kwargs):
|
130 |
+
super(PositionalEmbedding, self).__init__(**kwargs)
|
131 |
+
self.token_embeddings = layers.Embedding(
|
132 |
+
input_dim=vocab_size, output_dim=embed_dim
|
133 |
+
)
|
134 |
+
self.position_embeddings = layers.Embedding(
|
135 |
+
input_dim=sequence_length, output_dim=embed_dim
|
136 |
+
)
|
137 |
+
self.sequence_length = sequence_length
|
138 |
+
self.vocab_size = vocab_size
|
139 |
+
self.embed_dim = embed_dim
|
140 |
+
|
141 |
+
def call(self, inputs):
|
142 |
+
length = tf.shape(inputs)[-1]
|
143 |
+
positions = tf.range(start=0, limit=length, delta=1)
|
144 |
+
embedded_tokens = self.token_embeddings(inputs)
|
145 |
+
embedded_positions = self.position_embeddings(positions)
|
146 |
+
return embedded_tokens + embedded_positions
|
147 |
+
|
148 |
+
def compute_mask(self, inputs, mask=None):
|
149 |
+
return tf.math.not_equal(inputs, 0)
|
150 |
+
def get_config(self):
|
151 |
+
config = super().get_config()
|
152 |
+
config.update({
|
153 |
+
"sequence_length": self.sequence_length,
|
154 |
+
"vocab_size": self.vocab_size,
|
155 |
+
"embed_dim": self.embed_dim,
|
156 |
+
})
|
157 |
+
return config
|
158 |
+
|
159 |
+
|
160 |
+
class TransformerDecoder(layers.Layer):
|
161 |
+
def __init__(self, embed_dim, latent_dim, num_heads, **kwargs):
|
162 |
+
super(TransformerDecoder, self).__init__(**kwargs)
|
163 |
+
self.embed_dim = embed_dim
|
164 |
+
self.latent_dim = latent_dim
|
165 |
+
self.num_heads = num_heads
|
166 |
+
self.attention_1 = layers.MultiHeadAttention(
|
167 |
+
num_heads=num_heads, key_dim=embed_dim
|
168 |
+
)
|
169 |
+
self.attention_2 = layers.MultiHeadAttention(
|
170 |
+
num_heads=num_heads, key_dim=embed_dim
|
171 |
+
)
|
172 |
+
self.dense_proj = keras.Sequential(
|
173 |
+
[
|
174 |
+
layers.Dense(latent_dim, activation="relu"),
|
175 |
+
layers.Dense(embed_dim),
|
176 |
+
]
|
177 |
+
)
|
178 |
+
self.layernorm_1 = layers.LayerNormalization()
|
179 |
+
self.layernorm_2 = layers.LayerNormalization()
|
180 |
+
self.layernorm_3 = layers.LayerNormalization()
|
181 |
+
self.supports_masking = True
|
182 |
+
|
183 |
+
def call(self, inputs, encoder_outputs, mask=None):
|
184 |
+
causal_mask = self.get_causal_attention_mask(inputs)
|
185 |
+
if mask is not None:
|
186 |
+
padding_mask = tf.cast(mask[:, tf.newaxis, :], dtype="int32")
|
187 |
+
padding_mask = tf.minimum(padding_mask, causal_mask)
|
188 |
+
|
189 |
+
attention_output_1 = self.attention_1(
|
190 |
+
query=inputs, value=inputs, key=inputs, attention_mask=causal_mask
|
191 |
+
)
|
192 |
+
out_1 = self.layernorm_1(inputs + attention_output_1)
|
193 |
+
|
194 |
+
attention_output_2 = self.attention_2(
|
195 |
+
query=out_1,
|
196 |
+
value=encoder_outputs,
|
197 |
+
key=encoder_outputs,
|
198 |
+
attention_mask=padding_mask,
|
199 |
+
)
|
200 |
+
out_2 = self.layernorm_2(out_1 + attention_output_2)
|
201 |
+
|
202 |
+
proj_output = self.dense_proj(out_2)
|
203 |
+
return self.layernorm_3(out_2 + proj_output)
|
204 |
+
|
205 |
+
def get_causal_attention_mask(self, inputs):
|
206 |
+
input_shape = tf.shape(inputs)
|
207 |
+
batch_size, sequence_length = input_shape[0], input_shape[1]
|
208 |
+
i = tf.range(sequence_length)[:, tf.newaxis]
|
209 |
+
j = tf.range(sequence_length)
|
210 |
+
mask = tf.cast(i >= j, dtype="int32")
|
211 |
+
mask = tf.reshape(mask, (1, input_shape[1], input_shape[1]))
|
212 |
+
mult = tf.concat(
|
213 |
+
[tf.expand_dims(batch_size, -1), tf.constant([1, 1], dtype=tf.int32)],
|
214 |
+
axis=0,
|
215 |
+
)
|
216 |
+
return tf.tile(mask, mult)
|
217 |
+
def get_config(self):
|
218 |
+
config = super().get_config()
|
219 |
+
config.update({
|
220 |
+
"embed_dim": self.embed_dim,
|
221 |
+
"latent_dim": self.latent_dim,
|
222 |
+
"num_heads": self.num_heads,
|
223 |
+
})
|
224 |
+
return config
|
225 |
+
|
226 |
+
url = "https://drive.google.com/uc?id=1a4eTAL4sLUi42P28Veihrv-fVPwFymTa"
|
227 |
+
gdown.download(url, quiet=False)
|
228 |
+
|
229 |
+
custom_objects = {"TransformerEncoder": TransformerEncoder, "PositionalEmbedding": PositionalEmbedding, "TransformerDecoder": TransformerDecoder}
|
230 |
+
with keras.utils.custom_object_scope(custom_objects):
|
231 |
+
transformer = tf.keras.models.load_model('re-model.h5')
|
232 |
+
|
233 |
+
cmn_vocab = cmn_vectorization.get_vocabulary()
|
234 |
+
cmn_index_lookup = dict(zip(range(len(cmn_vocab)), cmn_vocab))
|
235 |
+
max_decoded_sentence_length = 20
|
236 |
+
|
237 |
+
|
238 |
+
def decode_sequence_chinese(input_sentence):
|
239 |
+
tokenized_input_sentence = eng_vectorization([input_sentence])
|
240 |
+
decoded_sentence = "[start]"
|
241 |
+
for i in range(max_decoded_sentence_length):
|
242 |
+
tokenized_target_sentence = cmn_vectorization([decoded_sentence])[:, :-1]
|
243 |
+
predictions = transformer([tokenized_input_sentence, tokenized_target_sentence])
|
244 |
+
|
245 |
+
sampled_token_index = np.argmax(predictions[0, i, :])
|
246 |
+
sampled_token = cmn_index_lookup[sampled_token_index]
|
247 |
+
decoded_sentence += " " + sampled_token
|
248 |
+
|
249 |
+
if sampled_token == "[end]":
|
250 |
+
break
|
251 |
+
return decoded_sentence
|
252 |
+
|
app.py
CHANGED
@@ -2,20 +2,28 @@
|
|
2 |
import streamlit as st
|
3 |
from Autocorrect.autocorrectreal import edit
|
4 |
from TestTranslation.translation import *
|
5 |
-
|
6 |
|
7 |
|
8 |
st.title("Translation model test")
|
9 |
|
10 |
option = st.selectbox("Select input type:", ("text input", "audio input"))
|
|
|
11 |
if option == "text input":
|
12 |
input_sentence = st.text_input("Enter input sentence:")
|
13 |
if input_sentence is not None and len(input_sentence) > 0:
|
14 |
-
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
19 |
else:
|
20 |
wav_sentence = st.file_uploader("Upload a wav file:")
|
21 |
st.button("Submit wav file")
|
|
|
2 |
import streamlit as st
|
3 |
from Autocorrect.autocorrectreal import edit
|
4 |
from TestTranslation.translation import *
|
5 |
+
from TestTranslationChinese.translation_model import decode_sequence_chinese
|
6 |
|
7 |
|
8 |
st.title("Translation model test")
|
9 |
|
10 |
option = st.selectbox("Select input type:", ("text input", "audio input"))
|
11 |
+
option2 = st.selectbox("Select translation language:", ("Spanish", "Chinese"))
|
12 |
if option == "text input":
|
13 |
input_sentence = st.text_input("Enter input sentence:")
|
14 |
if input_sentence is not None and len(input_sentence) > 0:
|
15 |
+
if option2 == "Spanish":
|
16 |
+
edited = edit(input_sentence)
|
17 |
+
st.write("Autocorrected sentence: " + edited)
|
18 |
+
translated = decode_sequence(edited)[8:-5]
|
19 |
+
st.write(translated)
|
20 |
+
input_sentence = None
|
21 |
+
else:
|
22 |
+
edited = edit(input_sentence)
|
23 |
+
st.write("Autocorrected sentence: " + edited)
|
24 |
+
translated = decode_sequence_chinese(edited)[8:]
|
25 |
+
st.write(translated)
|
26 |
+
input_sentence = None
|
27 |
else:
|
28 |
wav_sentence = st.file_uploader("Upload a wav file:")
|
29 |
st.button("Submit wav file")
|