Spaces:
Running
Running
AILaborant
commited on
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,122 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import tensorflow as tf
|
2 |
+
from tensorflow.keras.layers import TextVectorization, Embedding, MultiHeadAttention, LayerNormalization, Dense, Dropout
|
3 |
+
from tensorflow.keras.models import Model
|
4 |
+
import gradio as gr
|
5 |
+
import json
|
6 |
+
|
7 |
+
START_TOKEN = '<start>'
|
8 |
+
END_TOKEN = '<end>'
|
9 |
+
|
10 |
+
class TransformerBlock(tf.keras.layers.Layer):
|
11 |
+
def __init__(self, embed_dim, num_heads, ff_dim, rate=0.2, **kwargs):
|
12 |
+
super().__init__(**kwargs)
|
13 |
+
self.embed_dim = embed_dim
|
14 |
+
self.num_heads = num_heads
|
15 |
+
self.ff_dim = ff_dim
|
16 |
+
self.rate = rate
|
17 |
+
|
18 |
+
self.att = MultiHeadAttention(num_heads=num_heads, key_dim=embed_dim)
|
19 |
+
self.ffn = tf.keras.Sequential([
|
20 |
+
Dense(ff_dim, activation='relu'),
|
21 |
+
Dense(embed_dim),
|
22 |
+
])
|
23 |
+
|
24 |
+
self.layernorm1 = LayerNormalization(epsilon=1e-5)
|
25 |
+
self.layernorm2 = LayerNormalization(epsilon=1e-5)
|
26 |
+
self.dropout1 = Dropout(rate)
|
27 |
+
self.dropout2 = Dropout(rate)
|
28 |
+
|
29 |
+
def call(self, inputs, training=None):
|
30 |
+
attn_output = self.att(inputs, inputs)
|
31 |
+
attn_output = self.dropout1(attn_output, training=training)
|
32 |
+
out1 = self.layernorm1(inputs + attn_output)
|
33 |
+
|
34 |
+
ffn_output = self.ffn(out1)
|
35 |
+
ffn_output = self.dropout2(ffn_output, training=training)
|
36 |
+
return self.layernorm2(out1 + ffn_output)
|
37 |
+
|
38 |
+
def get_config(self):
|
39 |
+
config = super().get_config()
|
40 |
+
config.update({
|
41 |
+
'embed_dim': self.embed_dim,
|
42 |
+
'num_heads': self.num_heads,
|
43 |
+
'ff_dim': self.ff_dim,
|
44 |
+
'rate': self.rate,
|
45 |
+
})
|
46 |
+
return config
|
47 |
+
|
48 |
+
class TokenAndPositionEmbedding(tf.keras.layers.Layer):
|
49 |
+
def __init__(self, maxlen, vocab_size, embed_dim, **kwargs):
|
50 |
+
super().__init__(**kwargs)
|
51 |
+
self.maxlen = maxlen
|
52 |
+
self.vocab_size = vocab_size
|
53 |
+
self.embed_dim = embed_dim
|
54 |
+
|
55 |
+
self.token_emb = Embedding(input_dim=vocab_size, output_dim=embed_dim)
|
56 |
+
self.pos_emb = Embedding(input_dim=maxlen, output_dim=embed_dim)
|
57 |
+
|
58 |
+
def call(self, x):
|
59 |
+
maxlen = tf.shape(x)[-1]
|
60 |
+
positions = tf.range(start=0, limit=maxlen, delta=1)
|
61 |
+
positions = self.pos_emb(positions)
|
62 |
+
x = self.token_emb(x)
|
63 |
+
return x + positions
|
64 |
+
|
65 |
+
def get_config(self):
|
66 |
+
config = super().get_config()
|
67 |
+
config.update({
|
68 |
+
'maxlen': self.maxlen,
|
69 |
+
'vocab_size': self.vocab_size,
|
70 |
+
'embed_dim': self.embed_dim,
|
71 |
+
})
|
72 |
+
return config
|
73 |
+
|
74 |
+
def load_model(filename="tg-medium"):
|
75 |
+
model = tf.keras.models.load_model(f'{filename}.h5', custom_objects={
|
76 |
+
'TokenAndPositionEmbedding': TokenAndPositionEmbedding,
|
77 |
+
'TransformerBlock': TransformerBlock
|
78 |
+
})
|
79 |
+
with open(f'{filename}.json', 'r', encoding='utf-8') as f:
|
80 |
+
vocab = json.load(f)
|
81 |
+
vectorizer = TextVectorization(
|
82 |
+
max_tokens=96000,
|
83 |
+
output_sequence_length=100,
|
84 |
+
standardize=None,
|
85 |
+
vocabulary=vocab
|
86 |
+
)
|
87 |
+
return model, vectorizer
|
88 |
+
|
89 |
+
def generate_text(model, vectorizer, prompt):
|
90 |
+
prompt = START_TOKEN + ' ' + prompt + ' ' + END_TOKEN
|
91 |
+
input_seq = vectorizer([prompt])
|
92 |
+
input_seq = input_seq[:, :-1]
|
93 |
+
predictions = model.predict(input_seq)
|
94 |
+
predicted_tokens = tf.argmax(predictions[0], axis=-1)
|
95 |
+
vocab = vectorizer.get_vocabulary()
|
96 |
+
output_tokens = [vocab[idx] for idx in predicted_tokens.numpy()]
|
97 |
+
if END_TOKEN in output_tokens:
|
98 |
+
end_index = output_tokens.index(END_TOKEN)
|
99 |
+
output_tokens = output_tokens[:end_index]
|
100 |
+
if START_TOKEN in output_tokens:
|
101 |
+
output_tokens.remove(START_TOKEN)
|
102 |
+
output = ' '.join(output_tokens)
|
103 |
+
return output
|
104 |
+
|
105 |
+
def main():
|
106 |
+
model, vectorizer = load_model()
|
107 |
+
|
108 |
+
def generate_response(prompt):
|
109 |
+
return generate_text(model, vectorizer, prompt)
|
110 |
+
|
111 |
+
iface = gr.Interface(
|
112 |
+
fn=generate_response,
|
113 |
+
inputs=gr.Textbox(lines=2, placeholder="Start your conversation."),
|
114 |
+
outputs="text",
|
115 |
+
title="tg-medium",
|
116 |
+
description="Interference API. (russian only)"
|
117 |
+
)
|
118 |
+
|
119 |
+
iface.launch()
|
120 |
+
|
121 |
+
if __name__ == "__main__":
|
122 |
+
main()
|