DOI

Adds a DOI for citation purposes

Drop TF requirement for general use

Only see_rnn.visuals_gen will work without TensorFlow

1.15.0: TF 2.3.0 compatibility; 'SCALEFIG' env flag

FEATURES:

• Compatibility with TensorFlow 2.3.0 (see "TESTS")
• os.environ['SCALEFIG'] (default '1') will scale all drawn figures; specify as tuple (w, h) to scale width & height separately.

BREAKING:

• TF 2.3.0 won't work with keras; this isn't see_rnn, this is breaking changes to some basic TF ops.
• Changed default os.environ['TF_KERAS'] to '1'

TESTS:

• Discontinued support for keras TF2.3.0+; keras TF2.2.0 Graph is still tested, but may discontinue entirely in the future, or reinstate depending on how Keras proceeds
• tf.keras now only tested with TF 2.3.0+ (and still with 1.14.0)

1.14.6: Cleaner axis limits; allow custom title

FEATURES:

• Cleaner axis limits for features_hist and features_hist_v2: xmin and xmax are now limited in characters to avoid clutter in plots
• Docs state title will set to title_mode if not one of 'grads', 'outputs', but this was false; fixed now.
• Created CHANGELOG.md

BREAKING:

• title_mode renamed to `title

1.14.5: Convenience, appearance improvements; bugfixes

FEATURES:

• Added bordercolor kwarg to features_2D which allows setting color of border lines, useful when images are majority-dark
• Improved xaxis annotation handling for pad_xticks kwarg in features_hist and features_hist_v2; behavior configurable via configs['pad_xticks']
• show_xy_ticks can now be an int or bool, which will automatically set to tuple (int, int).
• Changed default kwarg timesteps_xaxis to False in features_2D which would rotate image data

BREAKING:

• pad_xticks is now bool instead of int

BUGFIXES:

• features_2D: moved ndim != 3 check outside of timesteps_xaxis, which would fail to expand_dims(0, ...) for =False w/ 2D input
• weights_norm: the case of one weights per layer would process incorrectly due to iterating over an expected list, where get_weights returns the array itself in case of a single weight

MISC:

• Added test cases to account for fixed bugs

1.14.4: Fix L1-norm case in `weights_norm`

norm_fn=np.abs would compute L1 norm as: np.sqrt(np.sum(np.abs(x))), which is incorrect; the sqrt is redundant. norm_fn=np.abs will now work correctly. L2-norm case always worked correctly.

For L2-norm, set norm_fn = (np.sqrt, np.square) = (outer_fn, inner_fn), which will compute outer_fn(sum(inner_fn(x))). Note that norm_fn=np.square will no longer compute L2-norm correctly.

See RNN v1.14.3

A warning would be thrown even if _id='' or is otherwise falsy, which is redundant.

See RNN v1.14.2

TF2.2-Graph sample_weight bugfix

• Passes None instead of np.ones(len(x)) in get_gradients(sample_weight=None). This is a TF2.2 bug, not See RNN bug.
• Will still bug if sample_weight is not None - nothing to do here except wait for TF 2.3, or nightly when fixed

See RNN v1.14.1

BUGFIXES:

• 'softmax' activation for _id='*' in get_gradients wasn't handled properly
• Added test for softmax; other activations might error, exhaustive list for None gradient yielders undetermined

MISC:

• Moved testing imports to new backend.py
• Changed pathing logic in test_all.py to allow running as __main__
• Added conftest.py to disable plots when Spyder unit-testing, and allow when ran as __main__

See RNN v1.14.0

FEATURES:

• Up to date with TensorFlow 2.2.0
• Support for sample_weight and learning_phase for all backends (TF1, TF2, Eager, Graph, keras, tf.keras)
• Support for multi-input and multi-output networks
• params added to get_gradients; directly get grads of pre-fetched weights & outputs

BREAKING:

• _make_grads_fn no longer supports Eager for tf.keras (but does for keras)
• _get_grads deprecated
• sample_weights -> sample_weight in get_gradients

BUGFIXES:

• _id='*' will now omit 'softmax' activation layers in tf.keras get_gradients, which error with None for gradient
• Corrected gateless architecture detection for _get_cell_weights

MISC:

• Testing moved to TF 2.2, no longer includes TF 2.0 or TF 2.1
• Added _get_grads_eager to inspect_gen.py
• Added _get_params, _layer_of_output to utils.py
• Improved Input layer exclusion for _id='*'
• Added note + tip to get_gradients on performance
• Extended GPU detection method in tests to work with TF2.2