#!/usr/bin/env python | |
# coding: utf-8 | |
# Copyright 2021 The HuggingFace Team. All rights reserved. | |
# | |
# Licensed under the Apache License, Version 2.0 (the "License"); | |
# you may not use this file except in compliance with the License. | |
# You may obtain a copy of the License at | |
# | |
# http://www.apache.org/licenses/LICENSE-2.0 | |
# | |
# Unless required by applicable law or agreed to in writing, software | |
# distributed under the License is distributed on an "AS IS" BASIS, | |
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | |
# See the License for the specific language governing permissions and | |
# limitations under the License. | |
# This script creates a tiny random model | |
# | |
# It will be used then as "hf-internal-testing/tiny-xlm-roberta" | |
# ***To build from scratch*** | |
# | |
# 1. clone sentencepiece into a parent dir | |
# git clone https://github.com/google/sentencepiece | |
# | |
# 2. create a new repo at https://huggingface.co/new | |
# make sure to choose 'hf-internal-testing' as the Owner | |
# | |
# 3. clone | |
# git clone https://huggingface.co/hf-internal-testing/tiny-xlm-roberta | |
# cd tiny-xlm-roberta | |
# 4. start with some pre-existing script from one of the https://huggingface.co/hf-internal-testing/ tiny model repos, e.g. | |
# wget https://huggingface.co/hf-internal-testing/tiny-xlm-roberta/raw/main/make-tiny-xlm-roberta.py | |
# chmod a+x ./make-tiny-xlm-roberta.py | |
# mv ./make-tiny-xlm-roberta.py ./make-tiny-xlm-roberta.py | |
# | |
# 5. automatically rename things from the old names to new ones | |
# perl -pi -e 's|MT5|XLMRoberta|g' make-tiny-xlm-roberta.py | |
# perl -pi -e 's|mt5|xlm-roberta|g' make-tiny-xlm-roberta.py | |
# | |
# 6. edit and re-run this script while fixing it up | |
# ./make-tiny-xlm-roberta.py | |
# | |
# 7. add/commit/push | |
# git add * | |
# git commit -m "new tiny model" | |
# git push | |
# ***To update*** | |
# | |
# 1. clone the existing repo | |
# git clone https://huggingface.co/hf-internal-testing/tiny-xlm-roberta | |
# cd tiny-xlm-roberta | |
# | |
# 2. edit and re-run this script after doing whatever changes are needed | |
# ./make-tiny-xlm-roberta.py | |
# | |
# 3. commit/push | |
# git commit -m "new tiny model" | |
# git push | |
import sys | |
import os | |
from transformers import XLMRobertaTokenizer, XLMRobertaTokenizerFast, XLMRobertaConfig, XLMRobertaForCausalLM | |
mname_orig = "xlm-roberta-base" | |
mname_tiny = "tiny-xlm-roberta" | |
### Tokenizer | |
# Shrink the orig vocab to keep things small | |
vocab_keep_items = 5000 | |
tmp_dir = f"/tmp/{mname_tiny}" | |
vocab_orig_path = f"{tmp_dir}/sentencepiece.bpe.model" | |
vocab_short_path = f"{tmp_dir}/spiece-short.model" | |
# workaround for fast tokenizer protobuf issue, and it's much faster too! | |
os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python" | |
if 1: # set to 0 to skip this after running once to speed things up during tune up | |
# HACK: need the sentencepiece source to get sentencepiece_model_pb2, as it doesn't get installed | |
sys.path.append("../sentencepiece/python/src/sentencepiece") | |
import sentencepiece_model_pb2 as model | |
tokenizer_orig = XLMRobertaTokenizer.from_pretrained(mname_orig) | |
tokenizer_orig.save_pretrained(tmp_dir) | |
with open(vocab_orig_path, 'rb') as f: data = f.read() | |
# adapted from https://blog.ceshine.net/post/trim-down-sentencepiece-vocabulary/ | |
m = model.ModelProto() | |
m.ParseFromString(data) | |
print(f"Shrinking vocab from original {len(m.pieces)} dict items") | |
for i in range(len(m.pieces) - vocab_keep_items): _ = m.pieces.pop() | |
print(f"new dict {len(m.pieces)}") | |
with open(vocab_short_path, 'wb') as f: f.write(m.SerializeToString()) | |
m = None | |
tokenizer_fast_tiny = XLMRobertaTokenizerFast(vocab_file=vocab_short_path) | |
tokenizer_tiny = XLMRobertaTokenizer(vocab_file=vocab_short_path) | |
### Config | |
config_tiny = XLMRobertaConfig.from_pretrained(mname_orig) | |
# remember to update this to the actual config as each model is different and then shrink the numbers | |
config_tiny.update(dict( | |
vocab_size=vocab_keep_items+12, | |
d_ff=256, | |
d_kv=8, | |
d_model=64, | |
hidden_size=256, | |
intermediate_size=256, | |
max_position_embeddings=64, | |
num_attention_heads=2, | |
num_decoder_layers=2, | |
num_heads=2, | |
num_hidden_layers=2, | |
num_layers=2, | |
relative_attention_num_buckets=32, | |
)) | |
print("New config", config_tiny) | |
### Model | |
model_tiny = XLMRobertaForCausalLM(config_tiny) | |
print(f"{mname_tiny}: num of params {model_tiny.num_parameters()}") | |
model_tiny.resize_token_embeddings(len(tokenizer_tiny)) | |
inputs = tokenizer_tiny("hello", return_tensors="pt") | |
outputs = model_tiny(**inputs) | |
print("Test with normal tokenizer:", len(outputs.logits[0])) | |
inputs = tokenizer_fast_tiny("hello", return_tensors="pt") | |
outputs = model_tiny(**inputs) | |
print("Test with fast tokenizer:", len(outputs.logits[0])) | |
# Save | |
model_tiny.half() # makes it smaller | |
model_tiny.save_pretrained(".") | |
tokenizer_tiny.save_pretrained(".") | |
tokenizer_fast_tiny.save_pretrained(".") | |
readme = "README.md" | |
if not os.path.exists(readme): | |
with open(readme, "w") as f: | |
f.write(f"This is a {mname_tiny} random model to be used for basic testing.\n") | |
print(f"Generated {mname_tiny}") | |