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

import json
from transformers import AutoTokenizer
from tokenizers import Tokenizer
vocab_keep_items = 5000
tokenizer = AutoTokenizer.from_pretrained(mname_orig, use_fast=True)
assert tokenizer.is_fast, "This only works for fast tokenizers."
tokenizer_json = json.loads(tokenizer._tokenizer.to_str())
vocab = tokenizer_json["model"]["vocab"]
if tokenizer_json["model"]["type"] == "BPE":
    new_vocab = { token: i for token, i in vocab.items() if i < vocab_keep_items }
    merges = tokenizer_json["model"]["merges"]
    new_merges = []
    for i in range(len(merges)):
        a, b = merges[i].split()
        new_token = "".join((a, b))
        if a in new_vocab and b in new_vocab and new_token in new_vocab:
            new_merges.append(merges[i])
    tokenizer_json["model"]["merges"] = new_merges
elif tokenizer_json["model"]["type"] == "Unigram":
    new_vocab = vocab[:vocab_keep_items]
elif tokenizer_json["model"]["type"] == "WordPiece" or tokenizer_json["model"]["type"] == "WordLevel":
    new_vocab = { token: i for token, i in vocab.items() if i < vocab_keep_items }
else:
    raise ValueError(f"don't know how to handle {tokenizer_json['model']['type']}")
tokenizer_json["model"]["vocab"] = new_vocab
tokenizer._tokenizer = Tokenizer.from_str(json.dumps(tokenizer_json))
tokenizer_fast_tiny = tokenizer


### 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}")