Question

我已经成功地为带有张量流服务的模型提供了服务，并从中获得了预测。但是，为了提高效率，我想尝试输入图像的base64编码。我尝试了以下方法，但无法使其正常工作。有人能指出我正确的方向吗？

我加载一个已经训练好的模型，并准备将其与TF Serve一起使用：

import keras
import tensorflow
from keras import backend as K

K.set_learning_phase(0)

# model = load_model()

if fixed_shape is not None:
    shape = [None, *fixed_shape]
else:
    shape = [None, None, None, 3]

string_inp = tensorflow.placeholder(
    tensorflow.string, shape=(None,))  # string input for the base64 encoded image
imgs_map = tensorflow.map_fn(
    tensorflow.image.decode_image,
    string_inp,
    dtype=tensorflow.float32
)  # decode the jpeg
imgs_map.set_shape(shape)
pred = model(imgs_map)

model_inputs = list(string_inp)
model_outputs = list(pred)
inputs = {i.name + "_bytes": i for i in model_inputs}
outputs = {f"output {i}:" + o.name: o for i, o in enumerate(model_outputs)}
signature = tensorflow.saved_model.signature_def_utils.predict_signature_def(
    inputs=inputs, outputs=outputs
)

builder = tensorflow.saved_model.builder.SavedModelBuilder("out_dir")
builder.add_meta_graph_and_variables(
    sess=K.get_session(),
    tags=[tensorflow.saved_model.tag_constants.SERVING],
    signature_def_map={
        tensorflow.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY:
            signature
    })
builder.save()

然后为我使用的客户端

image_byte = base64.b64encode(batch[0]).decode("utf-8")
payload = '{"instances" : [{"b64": "%s"}]}' % image_byte
response = requests.post(predict_url, data=payload)

，结果为requests.exceptions.HTTPError: 400 Client Error: Bad Request for url:。但是，如果我在邮递员中尝试相同的请求，则会得到：

{
    "error": "2 root error(s) found.\n  (0) Invalid argument: assertion failed: [Unable to decode bytes as JPEG, PNG, GIF, or BMP]\n\t [[{{node map_5/while/decode_image/cond_jpeg/cond_png/cond_gif/Assert_1/Assert}}]]\n\t [[GroupCrossDeviceControlEdges_0/map_5/while/decode_image/cond_jpeg/cond_png/cond_gif/DecodeBmp/_4702]]\n  (1) Invalid argument: assertion failed: [Unable to decode bytes as JPEG, PNG, GIF, or BMP]\n\t [[{{node map_5/while/decode_image/cond_jpeg/cond_png/cond_gif/Assert_1/Assert}}]]\n0 successful operations.\n0 derived errors ignored."
}

TF服务：如何对以base64编码的图像进行预测？

0 个答案: