I'm migrating an old TensorFlow 1.x training script and I got some problem with hub.text_embedding_column function. At the moment, the following code does not work
# Python 3.6.9
import tensorflow as tf # tf.__version__ 2.1.0
import tensorflow_hub as hub # hub.__version__ 0.7.0
module_spec = hub.load_module_spec('https://tfhub.dev/google/universal-sentence-encoder/4')
text_column = hub.text_embedding_column(key='test_col', module_spec=module_spec)
The error that I got is:
RuntimeError: Missing implementation that supports: loader(*('/tmp/tfhub_modules/29abffb443cb0a0ca9c72e8e3863b76d85028490',), **{})
I tried help(hub.text_embedding_column) and the help returns me
TODO(b/131678043): This does not work yet with TF2.
Do you know any workaround to use text_embedding_column with TF2? I'm able to load the model using hub.load('https://tfhub.dev/google/universal-sentence-encoder/4') but then I don't know what to do with it.
Thank you all (:
Following this github issue I was able to solve my problem, at least for now. It's still based on tf.compact.v1.placeholder but at least I can use hub.text_embedding_column.
# Python 3.6.9
import tensorflow as tf # tf.__version__ 2.1.0
import tensorflow_hub as hub # hub.__version__ 0.7.0
def build_module_fn(model_url):
def module_fn():
text_input = tf.compat.v1.placeholder(dtype=tf.string, shape=[None])
embed_layer = hub.KerasLayer(
model_url,
input_shape=[], # Expects a tensor of shape [batch_size] as input.
dtype=tf.string) # Expects a tf.string input tensor
embeddings = embed_layer(text_input)
hub.add_signature(inputs=text_input, outputs=embeddings)
return module_fn
module_spec = hub.create_module_spec(build_module_fn('https://tfhub.dev/google/universal-sentence-encoder/4'))
hub.text_embedding_column(key='text', module_spec=module_spec)
Hope that this can help other people!
Related
I'm trying to make encoder decoder like model where I'm getting the following error at model.fit
AttributeError: module 'tensorflow.python.framework.ops' has no attribute '_TensorLike'
model.fit(train_dataloader,
validation_data = test_dataloader,
steps_per_epoch=len(train_dataset)//8,
epochs=10)
I'm using keras 2.3.1 and segmentaion-model
How to resolve it?
This issue is fixed in latest keras version 2.6.0
Workaround for older Keras version
Change is_tensor in file keras/backend/tensorflow_backend.py, as of keras with tensorflow 2.3.0
from tensorflow.python.framework import tensor_util
def is_tensor(x):
return tensor_util.is_tensor(x)
#return isinstance(x, tf_ops._TensorLike) or tf_ops.is_dense_tensor_like(x)
I'm trying to calculate the F1 score using tf.contrib.metrics.f1_score, but it gives me an error. I know how to calculate it using precision and recall but i want to use this function.
I have tried it on ubuntu 16.04 LTS with tensorflow version 1.9.0 with gpu suport and no gpu suport
from tensorflow.contrib.metrics import f1_score as ms
i get this error:
ImportError: Traceback (most recent call last)
<ipython-input-6-627f14191ea2> in <module>()----> 1 from tensorflow.contrib.metrics import f1_score as ms
ImportError: cannot import name 'f1_score'
AND
from tensorflow.contrib import metrics as ms
ms.f1_score
I get this error:
AttributeError Traceback (most recent call last)
<ipython-input-8-c19f57465581> in <module>()
1 from tensorflow.contrib import metrics as ms
----> 2 ms.f1_score
AttributeError: module 'tensorflow.contrib.metrics' has no attribute 'f1_score'
I expect ms.f1_score would load
If you are sure that you have tf.contrib available and this doesn't work for you, maybe you will need to reinstall tensorflow use pip install -U tensorflow or use the -GPU if you are using that version.
If it fails, go to the place where tensorflow is installed and manually check if it is available or not, if it is available, make sure that you don't have a file in the same directory (Current working directory) named as tensorflow.py or tf.py
After that you should get
Update: As pointed by User #grwlf
Since TensorFlow 2.0, tf.contrib modules were moved to the Addons repo. See github.com/tensorflow/addons. There, F1 mesure is available as F1Score from tensorflow_addons.metrics import F1Score
You can find the documentation of f1_score here
Since it is a function, maybe you can try out:
from tensorflow.contrib import metrics as ms
ms.f1_score(labels,predictions)
Which will return a scalar tensor of the best f1 scores across different thresholds.
Example from tensorflow docs:
def model_fn(features, labels, mode):
predictions = make_predictions(features)
loss = make_loss(predictions, labels)
train_op = tf.contrib.training.create_train_op( total_loss=loss, optimizer='Adam')
eval_metric_ops = {'f1': f1_score(labels, predictions)}
return tf.estimator.EstimatorSpec( mode=mode, predictions=predictions, loss=loss, train_op=train_op, eval_metric_ops=eval_metric_ops, export_outputs=export_outputs)
estimator = tf.estimator.Estimator(model_fn=model_fn)
Hope this answers your question.
That's probably totally noob question which has something to do with python module importing, but I can't understand why the following is valid:
> import tensorflow as tf
> f = tf.train.Feature()
> from tensorflow import train
> f = train.Feature()
But the following statement causes an error:
> from tensorflow.train import Feature
ModuleNotFoundError: No module named 'tensorflow.train'
Can please somebody explain me why it doesn't work this way? My goal is to use more short notation in the code like this:
> example = Example(
features=Features(feature={
'x1': Feature(float_list=FloatList(value=feature_x1.ravel())),
'x2': Feature(float_list=FloatList(value=feature_x2.ravel())),
'y': Feature(int64_list=Int64List(value=label))
})
)
tensorflow version is 1.7.0
Solution
Replace
from tensorflow.train import Feature
with
from tensorflow.core.example.feature_pb2 import Feature
Explanation
Remarks about TensorFlow's Aliases
In general, you have to remember that, for example:
from tensorflow import train
is actually an alias for
from tensorflow.python.training import training
You can easily check the real module name by printing the module. For the current example you will get:
from tensorflow import train
print (train)
<module 'tensorflow.python.training.training' from ....
Your Problem
In Tensorflow 1.7, you can't use from tensorflow.train import Feature, because the from clause needs an actual module name (and not an alias). Given train is an alias, you will get an ImportError.
By doing
from tensorflow import train
print (train.Feature)
<class 'tensorflow.core.example.feature_pb2.Feature'>
you'll get the complete path of train. Now, you can use the import path as shown above in the solution above.
Note
In TensorFlow 1.9.0, from tensorflow.train import Feature will work, because tensorflow.train is an actual package, which you can therefore import. (This is what I see in my installed Tensorflow 1.9.0, as well as in the documentation, but not in the Github repository. It must be generated somewhere.)
Info about the path of the modules
You can find the complete module path in the docs. Every module has a "Defined in" section. See image below (taken from Module: tf.train):
I would advise against importing Feature (or any other object) from the non-public API, which is inconvenient (you have to figure out where Feature is actually defined), verbose, and subject to change in future versions.
I would suggest as an alternative to simply define
import tensorflow as tf
Feature = tf.train.Feature
I'm building a CNN text classifier using TensorFlow which I want to load in tensorflow-serving and query using the serving apis. When I call the Predict() method on the grcp stub I receive this error: AttributeError: 'grpc._cython.cygrpc.Channel' object has no attribute 'unary_unary'
What I've done to date:
I have successfully trained and exported a model suitable for serving (i.e., the signatures are verified and using tf.Saver I can successfully return a prediction). I can also load the model in tensorflow_model_server without error.
Here is a snippet of the client code (simplified for readability):
with tf.Session() as sess:
host = FLAGS.server
channel = grpc.insecure_channel('localhost:9001')
stub = prediction_service_pb2.beta_create_PredictionService_stub(channel)
request = predict_pb2.PredictRequest()
request.model_spec.name = 'predict_text'
request.model_spec.signature_name = 'predict_text'
x_text = ["space"]
# restore vocab processor
# then create a ndarray with transform_fit using the vocabulary
vocab = learn.preprocessing.VocabularyProcessor.restore('/some_path/model_export/1/assets/vocab')
x = np.array(list(vocab.fit_transform(x_text)))
# data
temp_data = tf.contrib.util.make_tensor_proto(x, shape=[1, 15], verify_shape=True)
request.inputs['input'].CopyFrom(tf.contrib.util.make_tensor_proto(x, shape=[1, 15], verify_shape=True))
# get classification prediction
result = stub.Predict(request, 5.0)
Where I'm bending the rules: I am using tensorflow-serving-apis in Python 3.5.3 when pip install is not officially supported. Various posts (example: https://github.com/tensorflow/serving/issues/581) have reported that using tensorflow-serving with Python 3 has been successful. I have downloaded tensorflow-serving-apis package from pypi (https://pypi.python.org/pypi/tensorflow-serving-api/1.5.0)and manually pasted into the environment.
Versions: tensorflow: 1.5.0, tensorflow-serving-apis: 1.5.0, grpcio: 1.9.0rc3, grcpio-tools: 1.9.0rcs, protobuf: 3.5.1 (all other dependency version have been verified but are not included for brevity -- happy to add if they have utility)
Environment: Linux Mint 17 Qiana; x64, Python 3.5.3
Investigations:
A github issue (https://github.com/GoogleCloudPlatform/google-cloud-python/issues/2258) indicated that a historical package triggered this error was related to grpc beta.
What data or learning or implementation am I missing?
beta_create_PredictionService_stub() is deprecated. Try this:
from tensorflow_serving.apis import prediction_service_pb2_grpc
...
stub = prediction_service_pb2_grpc.PredictionServiceStub(channel)
Try to use grpc.beta.implementations.insecure_channel instead of grpc.insecure_channel.
See example code here.
I'm using CNTK as the backend for Keras. I'm trying to use my model which I have trained using Keras in C++.
I have trained and saved my model using Keras which is in HDF5. How do I now use CNTK API to save it in their model-v2 format?
I tried this:
model = load_model('model2.h5')
cntk.ops.functions.Function.save(model, 'CNTK_model2.pb')
but i got the following error:
TypeError: save() missing 1 required positional argument: 'filename'
If tensorflow were the backend I would have done this:
model = load_model('model2.h5')
sess = K.get_session()
tf_saver = tf.train.Saver()
tf_saver.save(sess=sess, save_path=checkpoint_path)
How can I achieve the same thing?
As per the comments here, I was able to use this:
import cntk as C
import keras.backend as K
keras_model = K.load_model('my_keras_model.h5')
C.combine(keras_model.model.outputs).save('my_cntk_model')
cntk_model = C.load_model('my_cntk_model')
You can do something like this
model.outputs[0].save('CNTK_model2.pb')
I'm assuming here you have called model.compile (i.e. that's the only case I have tried :-)
The reason you see this error is because keras' cntk backend use a user defined function to do reshape on batch axis, which can't been serialized. We have fixed this issue in CNTK v2.2. Please upgrade your cntk to v2.2, and upgrade keras to last master.
Please see this pull request:
https://github.com/fchollet/keras/pull/7907