我已经按照本教程构建了一个用于通过tensorflow服务进行图像分类的服务器/客户端演示 https://github.com/tmlabonte/tendies/blob/master/minimum_working_example/tendies-basic-tutorial.ipynb
客户
它接受图像作为输入,将其转换为Base64,然后使用JSON将其传递给服务器
input_image = open(image, "rb").read()
print("Raw bitstring: " + str(input_image[:10]) + " ... " + str(input_image[-10:]))
# Encode image in b64
encoded_input_string = base64.b64encode(input_image)
input_string = encoded_input_string.decode("utf-8")
print("Base64 encoded string: " + input_string[:10] + " ... " + input_string[-10:])
# Wrap bitstring in JSON
instance = [{"images": input_string}]
data = json.dumps({"instances": instance})
print(data[:30] + " ... " + data[-10:])
r = requests.post('http://localhost:9000/v1/models/cnn:predict', data=data)
#json.loads(r.content)
print(r.text)
服务器
将模型加载为.h5后,服务器必须另存为SavedModel。 该映像必须作为Base64编码的字符串从客户端传递到服务器。
model=tf.keras.models.load_model('./model.h5')
input_bytes = tf.placeholder(tf.string, shape=[], name="input_bytes")
# input_bytes = tf.reshape(input_bytes, [])
# Transform bitstring to uint8 tensor
input_tensor = tf.image.decode_jpeg(input_bytes, channels=3)
# Convert to float32 tensor
input_tensor = tf.image.convert_image_dtype(input_tensor, dtype=tf.float32)
input_tensor = input_tensor / 127.5 - 1.0
# Ensure tensor has correct shape
input_tensor = tf.reshape(input_tensor, [64, 64, 3])
# CycleGAN's inference function accepts a batch of images
# So expand the single tensor into a batch of 1
input_tensor = tf.expand_dims(input_tensor, 0)
# x = model.input
y = model(input_tensor)
然后input_bytes成为SavedModel中predition_signature的输入
tensor_info_x = tf.saved_model.utils.build_tensor_info(input_bytes)
最后,服务器结果如下:
§ saved_model_cli show --dir ./ --all
signature_def['predict']:
The given SavedModel SignatureDef contains the following input(s):
inputs['images'] tensor_info:
dtype: DT_STRING
shape: ()
name: input_bytes:0
The given SavedModel SignatureDef contains the following output(s):
outputs['scores'] tensor_info:
dtype: DT_FLOAT
shape: (1, 4)
name: sequential_1/dense_2/Softmax:0
Method name is: tensorflow/serving/predict
发送图像
当我发送图像base64时,我从服务器收到一个关于输入形状的运行时错误,该输入似乎不是标量的:
Using TensorFlow backend.
Raw bitstring: b'\xff\xd8\xff\xe0\x00\x10JFIF' ... b'0;s\xcfJ(\xa0h\xff\xd9'
Base64 encoded string: /9j/4AAQSk ... 9KKKBo/9k=
{"instances": [{"images": "/9j ... Bo/9k="}]}
{ "error": "contents must be scalar, got shape [1]\n\t [[{{node DecodeJpeg}} = DecodeJpeg[_output_shapes=[[?,?,3]], acceptable_fraction=1, channels=3, dct_method=\"\", fancy_upscaling=true, ratio=1, try_recover_truncated=false, _device=\"/job:localhost/replica:0/task:0/device:CPU:0\"](_arg_input_bytes_0_0)]]" }
从服务器上看到的input_bytes
与shape=[]
标量一样,我也曾尝试用tf.reshape(input_bytes, [])
重塑它,但没有办法,我总是遇到相同的错误。
我没有在互联网上找到任何解决方案,也没有在Stackoverflow上找到有关此错误的解决方案。您能建议如何解决吗?
谢谢!
答案 0 :(得分:0)
我已解决该问题,我想评论一下如何才能从该解决方案中受益!
当您发送这样的json时:
{"instances": [{"images": "/9j ... Bo/9k="}]}
实际上,您在发送[]时发送的是大小为1的数组 如果您想发送2张图片,您应该这样写
{"instances": [{"images": "/9j ... Bo/9k="}, {"images": "/9j ... Bo/9k="}]}
这里的大小是2(形状= [2])
所以解决方案是在占位符中声明要接受shape = [None]的任何类型的尺寸
input_bytes = tf.placeholder(tf.string, shape=[None], name="input_bytes")
然后,如果您仅发送1张图像,则可以通过以下方式将向量1转换为标量:
input_scalar = tf.reshape(input_bytes, [])
我的代码中还有另一个错误,我不认为在tensorflow / serving中有通过在json中显式声明'b64'来解码base64的功能,请参考RESTful API Encoding binary values,所以如果您发送
{"instances": [{"images": {"b64": "/9j ... Bo/9k="}}]}
服务器将自动解码base64输入,正确的比特流将到达tf.image.decode_jpeg