make-tiny-deberta.py

#!/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-albert"

# ***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-albert
# cd tiny-albert

# 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-albert/raw/main/make-tiny-albert.py
# chmod a+x ./make-tiny-albert.py
# mv ./make-tiny-albert.py ./make-tiny-albert.py
#
# 5. automatically rename things from the old names to new ones
# perl -pi -e 's|Deberta|Deberta|g' make-*
# perl -pi -e 's|deberta|deberta|g' make-*
#
# 6. edit and re-run this script while fixing it up
# ./make-tiny-deberta.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-deberta
# cd tiny-deberta
#
# 2. edit and re-run this script after doing whatever changes are needed
# ./make-tiny-deberta.py
#
# 3. commit/push
# git commit -m "new tiny model"
# git push

import sys
import os

# workaround for fast tokenizer protobuf issue, and it's much faster too!
os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"

from transformers import DebertaTokenizer, DebertaTokenizerFast, DebertaConfig, DebertaForMaskedLM

mname_orig = "microsoft/deberta-base"
mname_tiny = "tiny-deberta"


### Tokenizer


# XXX: can't figure out how to shrink this tokenizer's vocab! Help?

# # Shrink the orig vocab to keep things small (just enough to tokenize any word, so letters+symbols)
# # DebertaTokenizerFast is fully defined by a tokenizer.json, which contains the vocab and the ids, so we just need to truncate it wisely
# import subprocess
# tokenizer_fast = DebertaTokenizerFast.from_pretrained(mname_orig)
# vocab_keep_items = 50265
# tmp_dir = f"/tmp/{mname_tiny}"
# tokenizer_fast.save_pretrained(tmp_dir)

# # resize tokenizer.json (vocab.txt will be automatically resized on save_pretrained)
# # perl -pi -e 's|(2999).*|$1}}}|' tokenizer.json # 0-indexed, so vocab_keep_items-1!
# closing_pat = "}}}"
# cmd = (f"perl -pi -e s|({vocab_keep_items-1}).*|$1{closing_pat}| {tmp_dir}/tokenizer.json").split()
# result = subprocess.run(cmd, capture_output=True, text=True)
# # reload with modified tokenizer
# tokenizer_fast_tiny = DebertaTokenizerFast.from_pretrained(tmp_dir)
# # it seems that DebertaTokenizer is not needed and DebertaTokenizerFast does the job


# # Shrink the orig vocab to keep things small (just enough to tokenize any word, so letters+symbols)
# # ElectraTokenizerFast is fully defined by a tokenizer.json, which contains the vocab and the ids, so we just need to truncate it wisely
# import subprocess
# tokenizer_fast = DebertaTokenizerFast.from_pretrained(mname_orig)
# vocab_keep_items = 5120
# tmp_dir = f"/tmp/{mname_tiny}"
# vocab_short_path = f"{tmp_dir}/vocab.json"
# tokenizer_fast.save_pretrained(tmp_dir)
# # resize tokenizer.json (vocab.txt will be automatically resized on save_pretrained)
# # perl -pi -e 's|(2999).*|$1}}}|' tokenizer.json # 0-indexed, so vocab_keep_items-1!
# closing_pat = "}"
# cmd = (f"perl -pi -e s|({vocab_keep_items-1}).*|$1{closing_pat}| {tmp_dir}/vocab.json").split()
# result = subprocess.run(cmd, capture_output=True, text=True)
# # reload with modified tokenizer
# #tokenizer_fast_tiny = DebertaTokenizerFast.from_pretrained(tmp_dir, vocab_file=vocab_short_path)
# # it seems that ElectraTokenizer is not needed and ElectraTokenizerFast does the job


# using full tokenizer for now
tokenizer_fast_tiny = DebertaTokenizerFast.from_pretrained(mname_orig)





### Config

config_tiny = DebertaConfig.from_pretrained(mname_orig)
print(config_tiny)
# 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,
    embedding_size=32,
    pooler_size=32,
    hidden_size=32,
    intermediate_size=64,
    max_position_embeddings=128,
    num_attention_heads=2,
    num_hidden_layers=2,
))
print("New config", config_tiny)

### Model

model_tiny = DebertaForMaskedLM(config_tiny)
print(f"{mname_tiny}: num of params {model_tiny.num_parameters()}")
model_tiny.resize_token_embeddings(len(tokenizer_fast_tiny))

# Test
inputs = tokenizer_fast_tiny("The capital of France is [MASK].", return_tensors="pt")
#print(inputs)
outputs = model_tiny(**inputs)
print("Test with normal tokenizer:", len(outputs.logits[0]))

# Save
model_tiny.half() # makes it smaller
model_tiny.save_pretrained(".")
tokenizer_fast_tiny.save_pretrained(".")

#print(model_tiny)

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}")