Please make sure that this is a bug. As per our
GitHub Policy,
we only address code/doc bugs, performance issues, feature requests and
build/installation issues on GitHub. tag:bug_template

System information

• Have I written custom code (as opposed to using a stock
  example script provided in TensorFlow): Yes

In Keras documentation, glorot_uniform says that the initializer is using Glorot Uniform from this paper. However, the Keras implementation is totally different from the equation on the paper. Also, there are some arguments such as mode ='fan_avg' is the default. It should be same as the referenced paper. 'fan_sum'. Golort uniform is shown

https://github.com/aymericdamien/TensorFlow-Examples/blob/9e1bb504f5d0f209d000997ce2ad95bb891798ab/examples/3_NeuralNetworks/bidirectional_rnn.py#L81

I think "outputs [-1]" and "outputs [0]" are equivalent (reversed) in this line of code, but the former (89%) works better than the latter (86%). Why?

Consider this code that downloads models and tokenizers to disk and then uses BertTokenizer.from_pretrained to load the tokenizer from disk.

ISSUE: BertTokenizer.from_pretrained() does not seem to be compatible with Python's native pathlib module.

# -*- coding: utf-8 -*-
"""
Created on: 25-04-2020
Author: MacwanJ

ISSUE:

When you look at the variables in the pretrained base uncased BERT the varibles look like list 1. When you do the training from scratch, 2 additional variables per layer are introduced, with suffixes adam_m and adam_v. It would be nice for someone to explain what these variables are? and what is their significance to the process of training?
If one were to manually initialize variables from a pri

Alexnet implementation in tensorflow has incomplete architecture where 2 convolution neural layers are missing. This issue is in reference to the python notebook mentioned below.

https://github.com/donnemartin/data-science-ipython-notebooks/blob/master/deep-learning/tensor-flow-examples/notebooks/3_neural_networks/alexnet.ipynb

Several parts of the op sec like the main op description, attributes, input and output descriptions become part of the binary that consumes ONNX e.g. onnxruntime causing an increase in its size due to strings that take no part in the execution of the model or its verification.

Setting __ONNX_NO_DOC_STRINGS doesn't really help here since (1) it's not used in the SetDoc(string) overload (s