TFJS 使用 headers 将模型保存到 http

Question

我正在尝试使用 https://www.tensorflow.org/js/guide/save_load 处的指南和后端来保存和上传带有附加 headers（对于 class 名称）的 tfjs 模型从 https://gist.github.com/dsmilkov/1b6046fd6132d7408d5257b0976f7864 复制。但是遵循指南并不能按指南中的预期和说明进行操作。我在哪里犯了错误？谢谢

我的浏览器代码是：

const saveResult = await model.save(tf.io.http('http://localhost:5000/upload', {method: 'POST', headers: {'class': 'Dog'}}));

服务器代码是：

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import io

from flask import Flask, Response, request
from flask_cors import CORS, cross_origin
import tensorflow as tf
import tensorflowjs as tfjs
import werkzeug.formparser

class ModelReceiver(object):

  def __init__(self):
    self._model = None
    self._model_json_bytes = None
    self._model_json_writer = None
    self._weight_bytes = None
    self._weight_writer = None

  @property
  def model(self):
    self._model_json_writer.flush()
    self._weight_writer.flush()
    self._model_json_writer.seek(0)
    self._weight_writer.seek(0)

    json_content = self._model_json_bytes.read()
    weights_content = self._weight_bytes.read()
    return tfjs.converters.deserialize_keras_model(
        json_content,
        weight_data=[weights_content],
        use_unique_name_scope=True)

  def stream_factory(self,
                     total_content_length,
                     content_type,
                     filename,
                     content_length=None):
    # Note: this example code isnot* thread-safe.
    if filename == 'model.json':
      self._model_json_bytes = io.BytesIO()
      self._model_json_writer = io.BufferedWriter(self._model_json_bytes)
      return self._model_json_writer
    elif filename == 'model.weights.bin':
      self._weight_bytes = io.BytesIO()
      self._weight_writer = io.BufferedWriter(self._weight_bytes)
      return self._weight_writer


def main():
  app = Flask('model-server')
  CORS(app)
  app.config['CORS_HEADER'] = 'Content-Type'

  model_receiver = ModelReceiver()

  @app.route('/upload', methods=['POST'])
  @cross_origin()
  def upload():
    print('headers are:')
    print(request.headers)
    print('Handling request...')
    werkzeug.formparser.parse_form_data(
        request.environ, stream_factory=model_receiver.stream_factory)
    print('Received model:')
    with tf.Graph().as_default(), tf.Session():
      model = model_receiver.model
      model.summary()
      # You can perform `model.predict()`, `model.fit()`,
      # `model.evaluate()` etc. here.
    return Response(status=200)

  app.run('localhost', 5000)


if __name__ == '__main__':
  main()

Answer 1

问题出在客户端，headers应该这样使用

var myInit = { method: 'POST',
               headers: {
                   'classes': ['class1', 'class2']
               }};

const saveResult = await model.save(tf.io.http(
    'http://localhost:5000/upload', loadOptions={requestInit: myInit}));

Answer 2

model.save 参数为 url 和配置。该配置包含 Request api 所需的 requestInit，在后台使用。

请求将如下所示

model.save(url, {requestInit: {method: 'POST', headers: {'class': 'Dog'}}))

Answer 3

如果你的目标是用模型存储一些辅助信息（比如class标签），TensorFlow.js中有tf.LayersModel相对little-known的特性这将使您的生活更轻松。它比使用 header.

更简单

是setUserDefinedMetadata()和getUserDefinedMetadata()方法。

在JavaScript这边，做：

// The argument to setUserDefinedMetadata() can be any serializable JSON
// object of a reasonable size.
myModel.setUserDefinedMetadata({outputClassLabels: ['Cat', 'Dog', 'Turtle']});

// The user metadata is stored with the model itself. No need to specify 
// additional headers.
await model.save('http://localhost:5000/upload');

接收模型工件的服务器只需检查请求中 JSON 负载的 'userDefinedMetadata' 字段即可。

TFJS 使用 headers 将模型保存到 http

TFJS save model to http with headers

python

flask

tensorflow.js