Tensor name has no shape information ERROR when predicting with tensorflow serving - tensorflow

Question

I'm using tensorflow serving to serve a savedmodel. I have two signatures: 1st outputting keras model.output and the 2nd outputting post processing of model.output. When I try a predict call of the 2nd signature on tensorflow serving it is giving me an error { "error": "Tensor name: prediction has no shape information " }
this is the code to build the savedmodel
shape1 = 92
shape2 = 92
reg=0.000001
learning_rate=0.001
sess = tf.Session()
K.set_session(sess)
K._LEARNING_PHASE = tf.constant(0)
K.set_learning_phase(0)
#preprocessing
x_input = tf.placeholder(tf.string, name='x_input', shape=[None])
reshaped = tf.reshape(x_input, shape=[])
image = tf.image.decode_jpeg(reshaped, channels=3)
image2 = tf.expand_dims(image,0)
resized = tf.image.resize_images(image2, (92,92))
meaned = tf.math.subtract(resized, tf.constant(116.0))
normalized = tf.math.divide(meaned, tf.constant(66.0))
#keras model
model = tf.keras.Sequential()
model.add(InputLayer(input_tensor=normalized))
model.add(Conv2D(32, (3, 3), padding='same', activation='relu', kernel_regularizer=l2(reg)))
model.add(Conv2D(32, (3, 3), padding='same', activation='relu', kernel_regularizer=l2(reg)))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(64, (3, 3), padding='same', activation='relu', kernel_regularizer=l2(reg)))
model.add(Dropout(0.1))
model.add(Conv2D(64, (3, 3), padding='same', activation='relu', kernel_regularizer=l2(reg)))
model.add(Dropout(0.1))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(128, (3, 3), padding='same', activation='relu', kernel_regularizer=l2(reg)))
model.add(Dropout(0.2))
model.add(Conv2D(128, (3, 3), padding='same', activation='relu', kernel_regularizer=l2(reg)))
model.add(Dropout(0.2))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(256, (3, 3), padding='same', activation='relu', kernel_regularizer=l2(reg)))
model.add(Dropout(0.3))
model.add(Conv2D(256, (3, 3), padding='same', activation='relu', kernel_regularizer=l2(reg)))
model.add(Dropout(0.3))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Flatten())
model.add(Dense(256, activation='relu', kernel_regularizer=l2(reg)))
model.add(Dropout(0.5))
model.add(Dense(256, activation='relu', kernel_regularizer=l2(reg)))
model.add(Dropout(0.5))
model.add(Dense(1))
model.add(Activation('sigmoid'))
model.compile(loss='binary_crossentropy',
optimizer=tf.train.RMSPropOptimizer(learning_rate=learning_rate),
metrics=['accuracy'])
#post processing to output label
pred = tf.gather_nd(model.output, (0,0))
label = tf.cond(pred > 0.5, lambda: tf.constant('Dog', shape=[]), lambda: tf.constant('Cat', shape=[]))
model.load_weights(r'./checkpoints/4.ckpt')
export_path = './saved_models/1'
init_op = tf.group(tf.global_variables_initializer(), tf.local_variables_initializer())
sess.run(init_op)
model.load_weights(r'./checkpoints/4.ckpt')
if os.path.isdir(export_path):
print('\nAlready saved a model, cleaning up\n')
print(subprocess.run(['rm', '-r', export_path]))
#first signature(this works)
x_info = tf.saved_model.utils.build_tensor_info(x_input)
y_info = tf.saved_model.utils.build_tensor_info(model.output)
sigmoid_signature = build_signature_def(inputs={"image": x_info}, outputs={"prediction":y_info}, method_name='tensorflow/serving/predict')
#2nd signature(this doesn't work)
x_info = tf.saved_model.utils.build_tensor_info(x_input)
y_info = tf.saved_model.utils.build_tensor_info(label)
label_signature = build_signature_def(inputs={"image": x_info}, outputs={"prediction":y_info}, method_name='tensorflow/serving/predict')
builder = tf.saved_model.builder.SavedModelBuilder(export_path)
legacy_init_op = tf.group(tf.tables_initializer(), name='legacy_init_op')
builder.add_meta_graph_and_variables(sess=sess,
tags=["serve"],
signature_def_map={'sigmoid': sigmoid_signature, 'label': label_signature})
builder.save()
this is code to call tf serving
imgs = ['./Dog/' + img for img in imgs]
img = open('./Dog/3.jpg', 'rb').read()
img = base64.b64encode(img).decode('utf-8')
data = json.dumps(
{"signature_name": "label",
"instances": [
{'image': {'b64': img}}
]
}
)
json_response = requests.post('http://localhost:8501/v1/models/pet:predict', data=data)
print(json_response.text)
Instead of getting a response of {"predictions": "Dog"}, i am getting an error { "error": "Tensor name: prediction has no shape information " }

Answer 1

I managed to fix this. I used tf.reshape on what i wanted to output and passed that into the signature builder.
#post processing to output label
pred = tf.gather_nd(model.output, (0,0))
label = tf.cond(pred > 0.5, lambda: tf.constant('Dog', shape=[]), lambda: tf.constant('Cat', shape=[]))
label_reshaped = tf.reshape(label, [None])
...
#2nd signature(this doesn't work)
x_info = tf.saved_model.utils.build_tensor_info(x_input)
y_info = tf.saved_model.utils.build_tensor_info(label_reshaped)
label_signature = build_signature_def(inputs={"image": x_info}, outputs={"prediction":y_info}, method_name='tensorflow/serving/predict')

Answer 2

Reading the tensorflow serving documentation, you'll see that there are two ways to specify input tensors in your request, the row format (using instances like your example), and the column format (using inputs).
Since the row format requires that all inputs and outputs have the same 0th dimension, if you did not export the model with explicit output shape, you cannot use the row format.
Therefore, in your case (without having to re-export the model with explicit reshaping, like the other answer has provided), you can send this payload instead
data = json.dumps(
{
"signature_name": "label",
"inputs": {'image': {'b64': img}}
}
)
On the other hand, keep in mind that if you do want to send multiple b64 encoded images, your best bet would be to use the row format with multiple instances (such as if you want to run batch predict on multiple images).