See RNN v1.13.5

FEATURES:

• detect_nans now detects np.inf and -np.inf via include_inf=True (default), and prints as e.g. 25% NaN, 10% Inf
• detect_nans now returns text without a newline, e.g. 50% NaN instead of 50%\nNaNs

BREAKING:

• visuals_rnn now passes include_inf=True to detect_nans

MISC:

• Fixed possible redundant newline in _print_nans
• detect_nans names nan NaN instead of NaNs

Se RNN v1.13.4

FEATURE: in visuals_gen and visuals_rnn, configs will now keep values of dicts of defaults, unless specifying the same keys. E.g.:

defaults = {'1': dict(a=1, b=2), '2': dict(c=3, d=4)}
configs  = {'1': dict(a=3, g=5)}
kw = {'1': {'a': 3, 'g': 5, 'b': 2}, '2': {'c': 3, 'd': 4}}  # AFTER
kw = {'1': {'a': 3, 'g': 5},         '2': {'c': 3, 'd': 4}}  # BEFORE

Also will apply deepcopy(configs) to not affect external dict.

See RNN v1.13.3

FEATURE: Add pad_xticks kwarg to features_hists

See RNN v1.13.2

BUGFIXES:

• get_weights(mode='weights', as_dict=True) would incorrectly handle data packing, getting layer names instead of weight names, thus omitting and mis-labeling weights
• get_weights() would ignore learning_phase argument, always defaulting to 1 via _get_grads()

See RNN v1.13.1

BUGFIX: Fix sample_weights handling in get_gradients for case sample_weights is not None

See RNN v1.13

FEATURES:

• Added get_weight_penalties, weight_loss to inspect_gen
• Added as_tensors kwarg to get_weights
• Added share_xy, center_zero kwargs to features_hist and features_hist_v2
• Deprecated equate_axes in visuals_gen in favor of share_xy, directly setting sharex and sharey
• Improved detect_nans performance

BREAKING:

• _detect_nans -> detect_nans (made public)

MISC:

• Added keys check to configs, ensuring keys serve functionality (e.g. 'subplots' does nothing; 'subplot' is correct)
• Improved get_gradients docstring
• Changed get_gradients arg ordering
• Updated README to account for changes

See RNN v1.12

FEATURES:

• Added hist_clipped

BREAKING:

• features_hist and features_hist_v2 now use hist_clipped() instead of .hist()

FIXES:

• Moved set_xlim() inside of axis loop for features_hist, which makes a difference if sharex=False

See RNN v1.11

FEATURES:

• _id='*' now fetches all layers (but input) (get_weights, get_gradients, get_outputs, weights_norm)
• get_weights now supports omit_names

BREAKING:

• omit_weight_names -> omit_names (weights_norm)

See RNN v1.1

FEATURES:

• Universal layer specifier, _id, that is a layer name, layer index, or list of either or of both
• Enabled full customization of various plot aspects via configs dict; see docstrings & code
• pip install support
• Figure saving support on all visuals
• Added weights_norm for computing & tracking layer weight norm statistics
• Added features_hist_v2 for multi-hist grid visualization
• Added features_hist for grid hist visualization
• rnn_histogram and rnn_heatmap now plot bias w/ kernel & recurrent kernel in one figure
• All plots now return fig and axes objects
• All methods now deepcopy passed lists and dicts when mutating internally to not affect originals

BREAKING:

• Several positional & keyword arguments deprecated in favor of configs
• layer_name & layer_idx deprecated; _id now takes place of both, and can be a list containing both
• label_channels in features_1D deprecated in favor of customizable annotations
• show_features_ -> features_ (0D, 1D, 2D)
• get_layer_outputs -> get_outputs
• get_layer_weights -> get_weights
• get_layer_gradients -> get_gradients
• scale_width -> w
• scale_height -> h

MISC:

• Created _backend.py to move repeated declarations and env vars to
• Improved docstrings
• Improved code readability
• Added tests
• get_full_name added to inspect_gen.py for retrieving full layer name given _id
• _filter_duplicates_by_keys added to utils.py
• _save_rnn_figs added to utils.py
• .flatten() -> .ravel(), since former copies array (thus slower)
• (Bugfix) codecs.open -> open, as former's formatting of README.md in setup.py was overflowing package build metadata

See RNN v1.05

New features:

• show_features_1D and show_features_2D now support tight and borderwidth specifications, useful when the number of subplots is very large.

Bug fixes:

• rnn_heatmap now handles IndRNN's recurrent_kernel as a vector, instead of a 2D array