TensorFlow notes

Tensor Ranks, Shapes and Types. The None element in a shape corresponds to a variable-sized dimension.
The API documentation for Tensor ops such as tf.reshape, tf.transpose, etc.
RNN reference: code for BasicRNNCell.
In generating data for an RNN or LSTM a common tensor shape is (batch_size, n_steps, input_size) so that for an input tensor T, the scalar T[i,j,k] is the kth coefficient of the vector which occurs as the jth time-step of the ith training sample in the batch. That is, the ith training sample is the sequence of vectors T[i,0], T[i,1], ..., T[i, n_steps-1]. See for example here or here. Confusion: arrays are 0-indexed but it's not clear to me in TF shapes whether the shape [9] means that it has 9 entries, or indices 0,...,9.
The foundational stuff in Oreilly "Hello, Tensorflow!" is quite good, on e.g. the graph
The documentation g3doc/get_started/basic_usage.md is pretty clear
ALso see g3doc/api_docs/python/math_ops.md

Provide feedback