doc
Browse files- make-tiny-xlm-roberta.py +14 -17
make-tiny-xlm-roberta.py
CHANGED
@@ -29,15 +29,15 @@
|
|
29 |
# 3. clone
|
30 |
# git clone https://huggingface.co/hf-internal-testing/tiny-xlm-roberta
|
31 |
# cd tiny-xlm-roberta
|
32 |
-
|
33 |
# 4. start with some pre-existing script from one of the https://huggingface.co/hf-internal-testing/ tiny model repos, e.g.
|
34 |
-
# wget https://huggingface.co/hf-internal-testing/tiny-
|
35 |
-
# chmod a+x ./make-tiny-
|
36 |
-
# mv ./make-tiny-
|
37 |
#
|
38 |
# 5. automatically rename things from the old names to new ones
|
39 |
-
# perl -pi -e 's|
|
40 |
-
# perl -pi -e 's|
|
41 |
#
|
42 |
# 6. edit and re-run this script while fixing it up
|
43 |
# ./make-tiny-xlm-roberta.py
|
@@ -63,7 +63,10 @@
|
|
63 |
import sys
|
64 |
import os
|
65 |
|
66 |
-
|
|
|
|
|
|
|
67 |
|
68 |
mname_orig = "xlm-roberta-base"
|
69 |
mname_tiny = "tiny-xlm-roberta"
|
@@ -75,13 +78,11 @@ vocab_keep_items = 5000
|
|
75 |
tmp_dir = f"/tmp/{mname_tiny}"
|
76 |
vocab_orig_path = f"{tmp_dir}/sentencepiece.bpe.model"
|
77 |
vocab_short_path = f"{tmp_dir}/spiece-short.model"
|
78 |
-
# workaround for fast tokenizer protobuf issue, and it's much faster too!
|
79 |
-
os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
|
80 |
if 1: # set to 0 to skip this after running once to speed things up during tune up
|
81 |
# HACK: need the sentencepiece source to get sentencepiece_model_pb2, as it doesn't get installed
|
82 |
sys.path.append("../sentencepiece/python/src/sentencepiece")
|
83 |
import sentencepiece_model_pb2 as model
|
84 |
-
tokenizer_orig =
|
85 |
tokenizer_orig.save_pretrained(tmp_dir)
|
86 |
with open(vocab_orig_path, 'rb') as f: data = f.read()
|
87 |
# adapted from https://blog.ceshine.net/post/trim-down-sentencepiece-vocabulary/
|
@@ -94,11 +95,11 @@ if 1: # set to 0 to skip this after running once to speed things up during tune
|
|
94 |
m = None
|
95 |
|
96 |
tokenizer_fast_tiny = XLMRobertaTokenizerFast(vocab_file=vocab_short_path)
|
97 |
-
tokenizer_tiny = XLMRobertaTokenizer(vocab_file=vocab_short_path)
|
98 |
|
99 |
### Config
|
100 |
|
101 |
config_tiny = XLMRobertaConfig.from_pretrained(mname_orig)
|
|
|
102 |
# remember to update this to the actual config as each model is different and then shrink the numbers
|
103 |
config_tiny.update(dict(
|
104 |
vocab_size=vocab_keep_items+12,
|
@@ -121,12 +122,9 @@ print("New config", config_tiny)
|
|
121 |
|
122 |
model_tiny = XLMRobertaForCausalLM(config_tiny)
|
123 |
print(f"{mname_tiny}: num of params {model_tiny.num_parameters()}")
|
124 |
-
model_tiny.resize_token_embeddings(len(
|
125 |
-
|
126 |
-
inputs = tokenizer_tiny("hello", return_tensors="pt")
|
127 |
-
outputs = model_tiny(**inputs)
|
128 |
-
print("Test with normal tokenizer:", len(outputs.logits[0]))
|
129 |
|
|
|
130 |
inputs = tokenizer_fast_tiny("hello", return_tensors="pt")
|
131 |
outputs = model_tiny(**inputs)
|
132 |
print("Test with fast tokenizer:", len(outputs.logits[0]))
|
@@ -134,7 +132,6 @@ print("Test with fast tokenizer:", len(outputs.logits[0]))
|
|
134 |
# Save
|
135 |
model_tiny.half() # makes it smaller
|
136 |
model_tiny.save_pretrained(".")
|
137 |
-
tokenizer_tiny.save_pretrained(".")
|
138 |
tokenizer_fast_tiny.save_pretrained(".")
|
139 |
|
140 |
readme = "README.md"
|
29 |
# 3. clone
|
30 |
# git clone https://huggingface.co/hf-internal-testing/tiny-xlm-roberta
|
31 |
# cd tiny-xlm-roberta
|
32 |
+
#
|
33 |
# 4. start with some pre-existing script from one of the https://huggingface.co/hf-internal-testing/ tiny model repos, e.g.
|
34 |
+
# wget https://huggingface.co/hf-internal-testing/tiny-albert/raw/main/make-tiny-albert.py
|
35 |
+
# chmod a+x ./make-tiny-albert.py
|
36 |
+
# mv ./make-tiny-albert.py ./make-tiny-xlm-roberta.py
|
37 |
#
|
38 |
# 5. automatically rename things from the old names to new ones
|
39 |
+
# perl -pi -e 's|Albert|XLMRoberta|g' make-tiny-xlm-roberta.py
|
40 |
+
# perl -pi -e 's|albert|xlm-roberta|g' make-tiny-xlm-roberta.py
|
41 |
#
|
42 |
# 6. edit and re-run this script while fixing it up
|
43 |
# ./make-tiny-xlm-roberta.py
|
63 |
import sys
|
64 |
import os
|
65 |
|
66 |
+
# workaround for fast tokenizer protobuf issue, and it's much faster too!
|
67 |
+
os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
|
68 |
+
|
69 |
+
from transformers import XLMRobertaTokenizerFast, XLMRobertaConfig, XLMRobertaForCausalLM
|
70 |
|
71 |
mname_orig = "xlm-roberta-base"
|
72 |
mname_tiny = "tiny-xlm-roberta"
|
78 |
tmp_dir = f"/tmp/{mname_tiny}"
|
79 |
vocab_orig_path = f"{tmp_dir}/sentencepiece.bpe.model"
|
80 |
vocab_short_path = f"{tmp_dir}/spiece-short.model"
|
|
|
|
|
81 |
if 1: # set to 0 to skip this after running once to speed things up during tune up
|
82 |
# HACK: need the sentencepiece source to get sentencepiece_model_pb2, as it doesn't get installed
|
83 |
sys.path.append("../sentencepiece/python/src/sentencepiece")
|
84 |
import sentencepiece_model_pb2 as model
|
85 |
+
tokenizer_orig = XLMRobertaTokenizerFast.from_pretrained(mname_orig)
|
86 |
tokenizer_orig.save_pretrained(tmp_dir)
|
87 |
with open(vocab_orig_path, 'rb') as f: data = f.read()
|
88 |
# adapted from https://blog.ceshine.net/post/trim-down-sentencepiece-vocabulary/
|
95 |
m = None
|
96 |
|
97 |
tokenizer_fast_tiny = XLMRobertaTokenizerFast(vocab_file=vocab_short_path)
|
|
|
98 |
|
99 |
### Config
|
100 |
|
101 |
config_tiny = XLMRobertaConfig.from_pretrained(mname_orig)
|
102 |
+
print(config_tiny)
|
103 |
# remember to update this to the actual config as each model is different and then shrink the numbers
|
104 |
config_tiny.update(dict(
|
105 |
vocab_size=vocab_keep_items+12,
|
122 |
|
123 |
model_tiny = XLMRobertaForCausalLM(config_tiny)
|
124 |
print(f"{mname_tiny}: num of params {model_tiny.num_parameters()}")
|
125 |
+
model_tiny.resize_token_embeddings(len(tokenizer_fast_tiny))
|
|
|
|
|
|
|
|
|
126 |
|
127 |
+
# Test
|
128 |
inputs = tokenizer_fast_tiny("hello", return_tensors="pt")
|
129 |
outputs = model_tiny(**inputs)
|
130 |
print("Test with fast tokenizer:", len(outputs.logits[0]))
|
132 |
# Save
|
133 |
model_tiny.half() # makes it smaller
|
134 |
model_tiny.save_pretrained(".")
|
|
|
135 |
tokenizer_fast_tiny.save_pretrained(".")
|
136 |
|
137 |
readme = "README.md"
|