make-tiny-albert.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-xlm-roberta
# chmod a+x ./make-tiny-xlm-roberta.py
# mv ./make-tiny-xlm-roberta.py ./make-tiny-albert.py
#
# 5. automatically rename things from the old names to new ones
# perl -pi -e 's|XLMRoberta|Albert|g' make-*
# perl -pi -e 's|xlm-roberta|albert|g' make-*
#
# 6. edit and re-run this script while fixing it up
# ./make-tiny-albert.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-albert
# cd tiny-albert
#
# 2. edit and re-run this script after doing whatever changes are needed
# ./make-tiny-albert.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 AlbertTokenizer, AlbertTokenizerFast, AlbertConfig, AlbertForMaskedLM

mname_orig = "albert-base-v1"
mname_tiny = "tiny-albert"
model_max_length = 256

### 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}/spiece.model"
vocab_short_path = f"{tmp_dir}/spiece-short.model"
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 = AlbertTokenizer.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

# albert breaks without having tokenizer.model_max_length match config.max_position_embeddings
# these values are hardcoded in the source for official models, so we have to explicitly set those here
tokenizer_fast_tiny = AlbertTokenizerFast(vocab_file=vocab_short_path,
                                          model_max_length=model_max_length)


### Config

config_tiny = AlbertConfig.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=64,
    hidden_size=32,
    intermediate_size=128,
    max_position_embeddings=model_max_length,
    num_attention_heads=2,
    num_hidden_groups=1,
    num_hidden_layers=2,
))
print("New config", config_tiny)

### Model

model_tiny = AlbertForMaskedLM(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}")