transformers/benchmark/run.py

import os
import torch
import onnxruntime
from transformers.onnx.features import FeaturesManager
from transformers.onnx import validate_model_outputs

import towhee
from towhee import ops
from pymilvus import connections, DataType, FieldSchema, Collection, CollectionSchema, utility
from datasets import load_dataset

from statistics import mode
import argparse

import transformers

transformers.logging.set_verbosity_error()

parser = argparse.ArgumentParser()
parser.add_argument('--model', required=True, type=str)
parser.add_argument('--dataset', type=str, default='imdb')
parser.add_argument('--insert_size', type=int, default=1000)
parser.add_argument('--query_size', type=int, default=100)
parser.add_argument('--topk', type=int, default=10)
parser.add_argument('--collection_name', type=str, default=None)
parser.add_argument('--format', type=str, required=True)
parser.add_argument('--onnx_dir', type=str, default='../saved/onnx')

args = parser.parse_args()
model_name = args.model
onnx_path = os.path.join(args.onnx_dir, model_name.replace('/', '-') + '.onnx')
dataset_name = args.dataset
insert_size = args.insert_size
query_size = args.query_size
topk = args.topk
collection_name = args.collection_name if args.collection_name else model_name.replace('-', '_').replace('/', '_')

device = 'cpu'
host = 'localhost'
port = '19530'
index_type = 'FLAT'
metric_type = 'L2'

data = load_dataset(dataset_name).shuffle(seed=32)
assert insert_size <= len(data['train']), 'There is no enough data. Please decrease insert size.'
assert insert_size <= len(data['test']), 'There is no enough data. Please decrease query size.'

insert_data = data['train']
insert_data = insert_data[:insert_size] if insert_size else insert_data
query_data = data['test']
query_data = query_data[:query_size] if query_size else query_data

# Warm up
print('Warming up...')
op = ops.text_embedding.transformers(model_name=model_name, device=device).get_op()
dim = op('This is test.').shape[-1]
print(f'output dim: {dim}')

# Prepare Milvus
print('Connecting milvus ...')
connections.connect(host=host, port=port)


def create_milvus(collection_name):
    print('Creating collection ...')
    fields = [
        FieldSchema(name='id', dtype=DataType.INT64, description='embedding id', is_primary=True, auto_id=True),
        FieldSchema(name='embedding', dtype=DataType.FLOAT_VECTOR, description='text embedding', dim=dim),
        FieldSchema(name='label', dtype=DataType.VARCHAR, description='label', max_length=500)
    ]
    schema = CollectionSchema(fields=fields, description=f'text embeddings for {model_name} on {dataset_name}')
    if utility.has_collection(collection_name):
        print(f'drop old collection: {collection_name}')
        collection = Collection(collection_name)
        collection.drop()
    collection = Collection(name=collection_name, schema=schema)
    print(f'A new collection is created: {collection_name}.')
    return collection


if args.format == 'pytorch':
    collection_name = collection_name + '_pytorch'

    def insert(model_name, collection_name):
        (
            towhee.dc['text', 'label'](zip(insert_data['text'], insert_data['label'])).stream()
                  .runas_op['text', 'text'](lambda s: s[:1024])
                  .text_embedding.transformers['text', 'emb'](model_name=model_name, device=device)
                  .runas_op['emb', 'emb'](lambda x: x[0])
                  .runas_op['label', 'label'](lambda y: str(y))
                  .ann_insert.milvus[('emb', 'label'), 'miluvs_insert'](
                        uri=f'tcp://{host}:{port}/{collection_name}'
                        )
                  .show(3)
        )
        collection = Collection(collection_name)
        return collection.num_entities


    def query(model_name, collection_name):
        benchmark = (
            towhee.dc['text', 'gt'](zip(query_data['text'], query_data['label'])).stream()
                  .runas_op['text', 'text'](lambda s: s[:1024])
                  .text_embedding.transformers['text', 'emb'](model_name=model_name, device=device)
                  .runas_op['emb', 'emb'](lambda x: x[0])
                  .runas_op['gt', 'gt'](lambda y: str(y))
                  .ann_search.milvus['emb', 'milvus_res'](
                        uri=f'tcp://{host}:{port}/{collection_name}',
                        metric_type=metric_type,
                        limit=topk,
                        output_fields=['label']
                        )
                  .runas_op['milvus_res', 'preds'](lambda x: [y.label for y in x]).unstream()
                  .runas_op['preds', 'pred1'](lambda x: mode(x[:1]))
                  .runas_op['preds', 'pred5'](lambda x: mode(x[:5]))
                  .runas_op['preds', 'pred10'](lambda x: mode(x[:10]))
                  .with_metrics(['accuracy'])
                  .evaluate['gt', 'pred1']('pred1')
                  .evaluate['gt', 'pred5']('pred5')
                  .evaluate['gt', 'pred10']('pred10')
                  .report()
        )
        return benchmark
elif args.format == 'onnx':
    collection_name = collection_name + '_onnx'
    saved_name = model_name.replace('/', '-')
    if not os.path.exists(onnx_path):
        inputs = op.tokenizer('This is test.', return_tensors='pt')
        input_names = list(inputs.keys())
        dynamic_axes = {}
        for i_n in input_names:
            dynamic_axes[i_n] = {0: 'batch_size', 1: 'sequence_length'}
        output_names = ['last_hidden_state']
        for o_n in output_names:
            dynamic_axes[o_n] = {0: 'batch_size', 1: 'sequence_length'}
        torch.onnx.export(
            op.model.model,
            tuple(inputs.values()),
            onnx_path,
            input_names=input_names,
            output_names=output_names,
            dynamic_axes=dynamic_axes,
            opset_version=14,
            do_constant_folding=True,
            )
    sess = onnxruntime.InferenceSession(onnx_path,
                                        providers=onnxruntime.get_available_providers())

    @towhee.register
    def run_onnx(txt):
        inputs = op.tokenizer(txt, return_tensors='np')
        onnx_inputs = [x.name for x in sess.get_inputs()]
        new_inputs = {}
        for k in onnx_inputs:
            new_inputs[k] = inputs[k]
        outs = sess.run(output_names=['last_hidden_state'], input_feed=dict(new_inputs))
        return outs[0].squeeze(0)

    def insert(model_name, collection_name):
        (
            towhee.dc['text', 'label'](zip(insert_data['text'], insert_data['label'])).stream()
                  .runas_op['text', 'text'](lambda s: s[:1024])
                  .run_onnx['text', 'emb']()
                  .runas_op['emb', 'emb'](lambda x: x[0])
                  .runas_op['label', 'label'](lambda y: str(y))
                  .ann_insert.milvus[('emb', 'label'), 'miluvs_insert'](
                        uri=f'tcp://{host}:{port}/{collection_name}'
                        )
                  .show(3)
        )
        collection = Collection(collection_name)
        return collection.num_entities

    def query(model_name, collection_name):
        benchmark = (
            towhee.dc['text', 'gt'](zip(query_data['text'], query_data['label'])).stream()
                  .runas_op['text', 'text'](lambda s: s[:1024])
                  .run_onnx['text', 'emb']()
                  .runas_op['emb', 'emb'](lambda x: x[0])
                  .runas_op['gt', 'gt'](lambda y: str(y))
                  .ann_search.milvus['emb', 'milvus_res'](
                           uri=f'tcp://{host}:{port}/{collection_name}',
                           metric_type=metric_type,
                           limit=topk,
                           output_fields=['label']
                           )
                  .runas_op['milvus_res', 'preds'](lambda x: [y.label for y in x]).unstream()
                  .runas_op['preds', 'pred1'](lambda x: mode(x[:1]))
                  .runas_op['preds', 'pred5'](lambda x: mode(x[:5]))
                  .runas_op['preds', 'pred10'](lambda x: mode(x[:10]))
                  .with_metrics(['accuracy'])
                  .evaluate['gt', 'pred1']('pred1')
                  .evaluate['gt', 'pred5']('pred5')
                  .evaluate['gt', 'pred10']('pred10')
                  .report()
        )
        return benchmark
else:
    raise AttributeError('Only support "pytorch" and "onnx" as format.')

collection = create_milvus(collection_name)
insert_count = insert(model_name, collection_name)
print('Total data inserted:', insert_count)
benchmark = query(model_name, collection_name)