nnlm | The simplest Neural Network Language Model , tensorflow | Machine Learning library

Mentioning the Answer in this (Answer) Section even though it is present in the Comments Section, for the benefit of the community.

Adding the import statement: import tensorflow_hub as hub and then using a custom layer with custom_objects={'KerasLayer': hub.KerasLayer} in the model_from_json() statement has resolved the error.

Complete working code is shown below:

Mentioning the Answer in this (Answer) section even though it is already present in the Comments Section, for the benefit of the Community.

Passing Raw Text Values instead of the Tokens (generated using Tokenizer) has resolved the issue.

Example code is shown below:

I walked through the same error and this is how I solved it;

My error was:

The choice of 16 units in the hidden layer is not a uniquely determined magic value. Like Shubham commented, it's all about experimenting and finding values that work well for your problem. Here is some folklore to guide your experimentation:

• The usual range for the number of units in hidden layers is tens to thousands.
• Powers of two may utilize specific hardware (like GPUs) more effectively.
• Simple feed-forward networks like the one above often decrease the number of units between successive layers. A commonly cited intuition is to progress from many basic features to fewer, more abstract ones. (Hidden layers tend to produce dense representations like embeddings, not discrete features, but the reasoning applies analogously to the dimension of the feature space.)
• The code snippet above does not show regularization. When trying whether more hidden units help, watch out for the gap between training and validation quality. A widening gap may indicate the need to regularize more.

Upgraded to tensorflow 2.0 and worked fine...

You are recreating the whole model on each iteration, so the TensorFlow graph is growing constantly. You should instead have a single model with a placeholder for your input, then feed the different paragraphs.

Answered in Github. Following is the link, https://github.com/tensorflow/transform/issues/125#issuecomment-514558533.

Posting the answer here for the benefit of community.

Adding tftransform_output.load_transform_graph() to train_input_fn will resolve the issue. This relates to the way tf.Learn works. In your serving graph, it tries to read from the training checkpoint, but because you are using materialized data, your training graph doesn't contain the embedding.

Below is the code for the same:

For TensorFlow 2, TF Hub will switch to shipping TF2's native object-based SavedModels [doc, RFC]. These are loaded by tf.saved_model.load() if already on your filesystem, or hub.load() with optional download from a URL. That gives you a restored Trackable object with a __call__ member that behaves like a @tf.function, meaning it has one or more concrete functions, each backed by a TF graph, and dispatches between them based on Tensor shapes/dtypes and non-Tensor arguments.

With the current alpha version of TF2, if you know the permissible TensorSpec for inputs, you can drill down to the outputs like:

I just tried it out and it works for me when I remove "input_shape = [None],". So this code should work: