make-tiny-layoutlm.py

#!/usr/bin/env python


import sys
import os

from transformers import LayoutLMTokenizer, LayoutLMTokenizerFast, LayoutLMConfig, LayoutLMForMaskedLM

mname_orig = "microsoft/layoutlm-base-uncased"
mname_tiny = "tiny-layoutlm"

### 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 = LayoutLMConfig.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,
    hidden_size=32,
    intermediate_size=64,
    max_position_embeddings=512,
    max_2d_position_embeddings=128,
    num_attention_heads=2,
    num_hidden_layers=2,
))
print("New config", config_tiny)



### Model

model_tiny = LayoutLMForMaskedLM(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}")