t5-make-very-small-model.py

#!/usr/bin/env python
# coding: utf-8
# Copyright 2020 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 smallish random model, with a few layers to test things like MP/PP, where
# tiny and tiner models are too too small
#
# It will be used then as "stas/t5-very-small-random"

from pathlib import Path
import json
import tempfile

from transformers import T5Tokenizer, T5TokenizerFast, T5Config, T5ForConditionalGeneration
from transformers.models.t5.tokenization_t5 import VOCAB_FILES_NAMES

mname_from = "patrickvonplaten/t5-tiny-random"
mname_very_small = "t5-very-small-random"

tokenizer = T5Tokenizer.from_pretrained(mname_from)
config = T5Config.from_pretrained(mname_from)
tokenizer_fast = T5TokenizerFast.from_pretrained(mname_from)

config.update(dict(
    vocab_size=32128,
    d_model=64,
    d_ff=256,
    d_kv=8,
    num_layers=8,
    num_decoder_layers=8,
    num_heads=4,
    relative_attention_num_buckets=32,
))

very_small_model = T5ForConditionalGeneration(config)
print(f"num of params {very_small_model.num_parameters()}")

# Test
src_texts = ["A long paragraph for summarization.", "Another paragraph for summarization."]
tgt_texts = ["Summary of the text.", "Another summary."]

batch = tokenizer.prepare_seq2seq_batch(src_texts, tgt_texts, return_tensors="pt")
outputs = very_small_model(**batch)

print("test output:", len(outputs.logits[0]))

# Save
very_small_model.half() # makes it smaller
very_small_model.save_pretrained(mname_very_small)
config.save_pretrained(mname_very_small)
tokenizer.save_pretrained(mname_very_small)
tokenizer_fast.save_pretrained(mname_very_small)

print(f"Generated {mname_very_small}")

# Upload
# transformers-cli repo create t5-very-small-random
# clone and add files