stas commited on
Commit
2b8bbe4
1 Parent(s): 7b9f4b2

add the script

Browse files
Files changed (1) hide show
  1. fsmt-make-tiny-model.py +59 -0
fsmt-make-tiny-model.py ADDED
@@ -0,0 +1,59 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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 model through reduction of a normal pre-trained model, but keeping the
21
+ # full vocab, merges file, and thus also resulting in a larger model due to a large vocab size.
22
+ # This gives ~3MB in total for all files.
23
+ #
24
+ # If you want a 50 times smaller than this see `fsmt-make-super-tiny-model.py`, which is slightly more complicated
25
+ #
26
+ #
27
+ # It will be used then as "stas/tiny-wmt19-en-de"
28
+
29
+ # Build
30
+ from transformers import FSMTTokenizer, FSMTConfig, FSMTForConditionalGeneration
31
+ mname = "facebook/wmt19-en-de"
32
+ tokenizer = FSMTTokenizer.from_pretrained(mname)
33
+ # get the correct vocab sizes, etc. from the master model
34
+ config = FSMTConfig.from_pretrained(mname)
35
+ config.update(dict(
36
+ d_model=4,
37
+ encoder_layers=1, decoder_layers=1,
38
+ encoder_ffn_dim=4, decoder_ffn_dim=4,
39
+ encoder_attention_heads=1, decoder_attention_heads=1))
40
+
41
+ tiny_model = FSMTForConditionalGeneration(config)
42
+ print(f"num of params {tiny_model.num_parameters()}")
43
+
44
+ # Test
45
+ batch = tokenizer(["Making tiny model"], return_tensors="pt")
46
+ outputs = tiny_model(**batch)
47
+
48
+ print("test output:", len(outputs.logits[0]))
49
+
50
+ # Save
51
+ mname_tiny = "tiny-wmt19-en-de"
52
+ tiny_model.half() # makes it smaller
53
+ tiny_model.save_pretrained(mname_tiny)
54
+ tokenizer.save_pretrained(mname_tiny)
55
+
56
+ print(f"Generated {mname_tiny}")
57
+
58
+ # Upload
59
+ # transformers-cli upload tiny-wmt19-en-de