Stas Bekman's picture stas
/
tiny-wmt19-en-ru

tiny-wmt19-en-ru / fsmt-make-super-tiny-model.py
1
#!/usr/bin/env python
2
# coding: utf-8
3
# Copyright 2020 The HuggingFace Team. All rights reserved.
4
#
5
# Licensed under the Apache License, Version 2.0 (the "License");
6
# you may not use this file except in compliance with the License.
7
# You may obtain a copy of the License at
8
#
9
#     http://www.apache.org/licenses/LICENSE-2.0
10
#
11
# Unless required by applicable law or agreed to in writing, software
12
# distributed under the License is distributed on an "AS IS" BASIS,
13
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14
# See the License for the specific language governing permissions and
15
# limitations under the License.
16
17
# This script creates a super tiny model that is useful inside tests, when we just want to test that
18
# the machinery works, without needing to the check the quality of the outcomes.
19
#
20
# This version creates a tiny vocab first, and then a tiny model - so the outcome is truly tiny -
21
# all files ~60KB. As compared to taking a full-size model, reducing to the minimum its layers and
22
# emb dimensions, but keeping the full vocab + merges files, leading to ~3MB in total for all files.
23
# The latter is done by `fsmt-make-super-tiny-model.py`.
24
#
25
# It will be used then as "stas/tiny-wmt19-en-ru"
26
27
from pathlib import Path
28
import json
29
import tempfile
30
31
from transformers import FSMTTokenizer, FSMTConfig, FSMTForConditionalGeneration
32
from transformers.models.fsmt.tokenization_fsmt import VOCAB_FILES_NAMES
33
34
mname_tiny = "tiny-wmt19-en-ru"
35
36
# Build
37
38
# borrowed from a test 
39
vocab = [ "l", "o", "w", "e", "r", "s", "t", "i", "d", "n", "w</w>", "r</w>", "t</w>", "lo", "low", "er</w>", "low</w>", "lowest</w>", "newer</w>", "wider</w>", "<unk>", ]
40
vocab_tokens = dict(zip(vocab, range(len(vocab))))
41
merges = ["l o 123", "lo w 1456", "e r</w> 1789", ""]
42
43
with tempfile.TemporaryDirectory() as tmpdirname:
44
    build_dir = Path(tmpdirname)
45
    src_vocab_file = build_dir / VOCAB_FILES_NAMES["src_vocab_file"]
46
    tgt_vocab_file = build_dir / VOCAB_FILES_NAMES["tgt_vocab_file"]
47
    merges_file = build_dir / VOCAB_FILES_NAMES["merges_file"]
48
    with open(src_vocab_file, "w") as fp: fp.write(json.dumps(vocab_tokens))
49
    with open(tgt_vocab_file, "w") as fp: fp.write(json.dumps(vocab_tokens))
50
    with open(merges_file, "w") as fp   : fp.write("\n".join(merges))
51
52
    tokenizer = FSMTTokenizer(
53
        langs=["en", "ru"],
54
        src_vocab_size = len(vocab),
55
        tgt_vocab_size = len(vocab),
56
        src_vocab_file=src_vocab_file,
57
        tgt_vocab_file=tgt_vocab_file,
58
        merges_file=merges_file,
59
    )
60
    
61
config = FSMTConfig(
62
    langs=['ru', 'en'],
63
    src_vocab_size=1000, tgt_vocab_size=1000,
64
    d_model=4,
65
    encoder_layers=1, decoder_layers=1,
66
    encoder_ffn_dim=4, decoder_ffn_dim=4,
67
    encoder_attention_heads=1, decoder_attention_heads=1,
68
)
69
70
tiny_model = FSMTForConditionalGeneration(config)
71
print(f"num of params {tiny_model.num_parameters()}")
72
73
# Test
74
batch = tokenizer(["Making tiny model"], return_tensors="pt")
75
outputs = tiny_model(**batch)
76
77
print("test output:", len(outputs.logits[0]))
78
79
# Save
80
tiny_model.half() # makes it smaller
81
tiny_model.save_pretrained(mname_tiny)
82
tokenizer.save_pretrained(mname_tiny)
83
84
print(f"Generated {mname_tiny}")
85
86
# Upload
87
# transformers-cli upload tiny-wmt19-en-ru
88