NEWS.md

In tpq/caress: Caress, a Gentler Introduction to Keras

caress 0.1.8

• Update set_layer_weights
  • Force unfreeze when freeze = FALSE

caress 0.1.7

• Update build
  • Make it easier to pass through custom loss functions

caress 0.1.6

• Add helper functions for gradients
  • New get_layer_loss to get the loss for a layer
  • New get_layer_gradient to get the gradient for a layer
  • New ascent to wrap gradient ascent

caress 0.1.5

• Update build
  • Pass monitor argument to early stopping callback
  • Add support for reduce LR on plateau callback

caress 0.1.3

• Add learnable arrays
  • New layer_learnable_array initializes a layer of learnable weights
• Add backend wrappers
  • New layer_kernel_dot wraps k_dot
  • New layer_kernel_conv2d wraps k_conv2d

caress 0.1.2

• Add pairwise layers
  • New layer_pairwise_residual calculates residuals for the final dimension of a tensor
  • New layer_pairwise_rmse calculates RMSE based on layer_pairwise_residual

caress 0.1.1

• Add tryCatch for deepviz call because of problems with lambda layer

caress 0.1.0

• Update constraints
  • Fix 0 / 0 NaN problem with the unit sum constraints

caress 0.0.9

• Update constraints
  • Fix 0 / 0 NaN problem with the unit sum constraints

caress 0.0.8

• Update stereo methods
  • Rename layer_to_dense_stereo_and_mono to layer_to_dense_stereo_and_add
  • Add layer_to_dense_stereo_and_cat to concatenate layers
  • Add layer_to_dense_stereo_and_diff to subtract layers
• Improve front-end
  • Fix to_output to correctly handle arrays
  • sample_random now uses min-max instead of mean-sd normalization
  • New model_decode wrapper helps decode latent space
  • Now pass loss_weights to compile()

caress 0.0.7

• Update stereo methods
  • Rename layer_to_stereo to layer_to_dense_stereo, etc.
  • Use layer_add instead of layer_concatenate to make the mono channel
  • Add explicit units and activation arguments
• Improve front-end
  • New layer_to_dense_DeepTRIAGE adds a DeepTRIAGE layer
  • from_input and to_output now correctly accept arrays
  • layer_pseudo_embed should now accept arrays
  • Now provide metrics as a named list

caress 0.0.6

• Add more helper functions
  • set_model_weights sets weights for a layer
  • model_mirror sets many weights based on a pre-trained "reference" model
  • constraint_runif uniformly randomizes weights each cycle
• Add new layer functions
  • layer_to_stereo splits a model into two parallel layers
  • layer_to_stereo_and_mono splits a model into three parallel layers
  • layer_pseudo_embedding creates an embedding matrix layer
• Improve back-end
  • Add callback_early_stopping callback to build by default
  • If all output is 0, treat as a regression model
  • build now chooses a suitable metric

caress 0.0.5

• Add more helper functions
  • layer_orthogonal_to equals output when orthogonal to output
• Improve front-end
  • Fix get_layer_output to handle lists of input
  • from_input and to_output now accept a vector of names
  • from_input and to_ouput now return a named list
  • Fix <<- WARNINGS

caress 0.0.4

• Improve front-end
  • from_input now supports lists of inputs
  • to_output now supports lists of outputs
  • prepare now supports lists of lists
• Improve back-end
  • Simplify type_of_y function to handle 2+D tensors
  • Replace type2loss with to_loss function
  • Have to_output call type_of_y

caress 0.0.3

• Add more helper functions
  • sample_random bulk prepares a list of training and test sets
  • constraint_rows_to_unit_sum constrains row sums of weights to 1
  • constraint_cols_to_unit_sum constrains column sums of weights to 1
  • constraint_all_zeros constrains all weights to 0
  • constraint_all_ones constrains all weights to 1

caress 0.0.2

• Add more helper functions
  • get_layer_names returns all layer names
  • get_layer_output returns layer output for a given input
  • get_layer_weights returns layer weights

caress 0.0.1

• Add some helper functions
  • from_input creates an input layer based on a matrix
  • to_output creates an output layer based on a matrix or vector
  • prepare joins input and output and displays model
  • build automatically compiles and fits model

tpq/caress documentation built on March 11, 2021, 8:03 p.m.