jaeriver opened a new issue, #21069:
URL: https://github.com/apache/incubator-mxnet/issues/21069
## Description
I'm testing nlp Bert performance using mxnet (including gluonnlp) in aws
lambda based on arm64 (graviton2 processor).
These scripts that I used
### save Bert model with Gluonnlp
```
model_name_ = "bert_12_768_12"
dataset = "book_corpus_wiki_en_uncased"
model, _ = nlp.model.get_model(
name=model_name_,
dataset_name=dataset,
pretrained=True,
use_pooler=True,
use_decoder=False,
use_classifier=False,
)
model = nlp.model.BERTClassifier(model, dropout=0.1, num_classes=2)
model.initialize(ctx=ctx)
model.hybridize(static_alloc=True)
mx_out = model(inputs_nd, token_types_nd, valid_length_nd)
mx_out.wait_to_read()
target_path = f"./{model_name}_{batch_size}"
from pathlib import Path
Path(target_path).mkdir(parents=True, exist_ok=True)
model.export(f'{model_name}_{batch_size}/model')
```
### load Bert in AWS Lambda based on ARM64
```
import mxnet as mx
import mxnet.ndarray as nd
from mxnet import nd, gluon
model_json, model_params = model_path + '/model-symbol.json', model_path +
'/model-0000.params'
model = gluon.nn.SymbolBlock.imports(model_json, ['data0', 'data1',
'data2'], model_params, ctx=ctx)
```
Then, I faced this error
```
[ERROR] MXNetError: MXNetError: Invalid Parameter format for output_dim
expect int but value='768', in operator
Embedding(name="bertmodel0_word_embed_embedding0_fwd", sparse_grad="False",
output_dim="768", dtype="float32", input_dim="30522")
```
I checked these scripts work without any issue in AWS Lambda based on X86
(Intel processor) and AWS EC2 Graviton2 environment.
## Environment
- AWS Lambda based on ARM64
- python 3.8
- mxnet 1.9.1
Could you help me to solve this issue? Thank you !
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
