File size: 3,536 Bytes
8f43b41 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 |
#!/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 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/mt5-tiny-random"
# To build:
# 1. clone sentencepiece into this dir
# git clone https://github.com/google/sentencepiece
#
# 2. run this script
from pathlib import Path
import json
import tempfile
from transformers import PegasusTokenizer, PegasusTokenizerFast, PegasusConfig, PegasusForConditionalGeneration
#from transformers.models.t5.tokenization_t5 import VOCAB_FILES_NAMES
mname_from = "google/pegasus-cnn_dailymail"
mname_very_small = "pegasus-cnn_dailymail-tiny-random"
tokenizer = PegasusTokenizer.from_pretrained(mname_from)
config = PegasusConfig.from_pretrained(mname_from)
#tokenizer_fast = PegasusTokenizerFast.from_pretrained(mname_from)
# Shrink the vocab of orig
import sys
# HACK: need the sentencepiece source to get sentencepiece_model_pb2, as it doesn't get installed
# git clone https://github.com/google/sentencepiece
sys.path.append("./sentencepiece/python/src/sentencepiece")
import sentencepiece_model_pb2 as model
tmp_dir = "/tmp/pegasus-tiny"
tokenizer.save_pretrained(tmp_dir)
file = tmp_dir + "/spiece.model"
with open(file, 'rb') as f: data = f.read()
# adapted from https://blog.ceshine.net/post/trim-down-sentencepiece-vocabulary/
m = model.ModelProto()
m.ParseFromString(data)
keep_items = 5000
print("Shrinking vocab")
print(f"original dict {len(m.pieces)}")
for i in range(len(m.pieces)-keep_items): _ = m.pieces.pop()
print(f"new dict {len(m.pieces)}")
with open(tmp_dir + "/spiece-short.model", 'wb') as f:
f.write(m.SerializeToString())
tokenizer = PegasusTokenizer(vocab_file=tmp_dir + "/spiece-short.model")
config.update(dict(
vocab_size=keep_items+12,
d_model=64,
decoder_attention_heads=2,
decoder_ffn_dim=64,
decoder_layers=2,
encoder_attention_heads=16,
encoder_ffn_dim=64,
encoder_layers=2,
num_hidden_layers=2,
))
print("new config", config)
very_small_model = PegasusForConditionalGeneration(config)
print(f"num of params {very_small_model.num_parameters()}")
very_small_model.resize_token_embeddings(len(tokenizer))
# 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 pegasus-cnn_dailymail-tiny-random
# clone and add files
|