File size: 4,513 Bytes
ce1c4dd 6318241 1cf16a5 ce1c4dd 8210be8 ce1c4dd 8210be8 ce1c4dd 0101c12 2e7cfe8 0101c12 899179c 340a52c 899179c 340a52c 0101c12 340a52c 0101c12 340a52c 0101c12 340a52c 6547184 340a52c 0101c12 2e7cfe8 0101c12 9e210c8 0101c12 9e210c8 0101c12 4918bed 9e210c8 0dcd084 9e210c8 0101c12 ce1c4dd |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 |
import gradio as gr
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline
import torch
device = "cuda:0" if torch.cuda.is_available() else "cpu"
model = AutoModelForSeq2SeqLM.from_pretrained("Jayyydyyy/m2m100_418m_tokipona").to(device)
tokenizer = AutoTokenizer.from_pretrained("facebook/m2m100_418M")
LANG_CODES = {
"English":"en",
"toki pona":"tl"
}
def translate(text, src_lang, tgt_lang, candidates:int):
"""
Translate the text from source lang to target lang
"""
src = LANG_CODES.get(src_lang)
tgt = LANG_CODES.get(tgt_lang)
tokenizer.src_lang = src
tokenizer.tgt_lang = tgt
ins = tokenizer(text, return_tensors='pt').to(device)
gen_args = {
'return_dict_in_generate': True,
'output_scores': True,
'output_hidden_states': True,
'length_penalty': 0.0, # don't encourage longer or shorter output,
'num_return_sequences': candidates,
'num_beams':candidates,
'forced_bos_token_id': tokenizer.lang_code_to_id[tgt]
}
outs = model.generate(**{**ins, **gen_args})
output = tokenizer.batch_decode(outs.sequences, skip_special_tokens=True)
return '\n'.join(output)
with gr.Blocks() as app:
markdown="""
# An English / toki pona Neural Machine Translation App!
### toki a! 💬
This is an english to toki pona / toki pona to english neural machine translation app.
Input your text to translate, a source language and target language, and desired number of return sequences!
### Grammar Regularization
An interesting quirk of training a many-to-many translation model is that pseudo-grammar correction
can be achieved by translating *from* **language A** *to* **language A**
Remember, this can ***approximate*** grammaticality, but it isn't always the best.
For example, "mi li toki e toki pona" (Source Language: toki pona & Target Language: toki pona) will result in:
- ['mi toki e toki pona.', 'mi toki pona.', 'mi toki e toki pona']
- (Thus, the ungrammatical "li" is dropped)
### Model and Data
This app utilizes a fine-tuned version of Facebook/Meta AI's M2M100 418M param model.
By leveraging the pretrained weights of the massively multilingual M2M100 model,
we can jumpstart our transfer learning to accomplish machine translation for toki pona!
The model was fine-tuned on the English/toki pona bitexts found at [https://tatoeba.org/](https://tatoeba.org/)
### This app is a work in progress and obviously not all translations will be perfect.
In addition to parameter quantity and the hyper-parameters used while training,
the *quality of data* found on Tatoeba directly influences the perfomance of projects like this!
If you wish to contribute, please add high quality and diverse translations to Tatoeba!
"""
with gr.Row():
gr.Markdown(markdown)
with gr.Column():
input_text = gr.components.Textbox(label="Input Text", value="Toad (Pit Crew) is a fun character you can try in Mario Kart Tour! Wow!")
source_lang = gr.components.Dropdown(label="Source Language", value="English", choices=list(LANG_CODES.keys()))
target_lang = gr.components.Dropdown(label="Target Language", value="toki pona", choices=list(LANG_CODES.keys()))
return_seqs = gr.Slider(label="Number of return sequences", value=3, minimum=1, maximum=24, step=1)
inputs=[input_text, source_lang, target_lang, return_seqs]
outputs = gr.Textbox()
translate_btn = gr.Button("Translate! | o ante toki!")
translate_btn.click(translate, inputs=inputs, outputs=outputs)
gr.Examples(
[
["Hello! How are you?", "English", "toki pona", 3],
["toki a! ilo pi ante toki ni li pona!", "toki pona", "English", 3],
["mi li toki e toki pona", "toki pona", "toki pona", 3],
["It's a good music generated by AI.", "English", "English", 3],
["I love this tool!", "English", "toki pona", 3],
["toki pona li toki pona.", "toki pona", "English", 3],
["pona toki a", "toki pona", "toki pona", 3],
["I want some bread and rice.", "English", "English", 3],
],
inputs=inputs
)
app.launch() |