The ruta's discuss from fdavidcl

apply_filter does not support higher dimensions

Error in data + term : arreglos de dimensón no compatibles

More control over verbosity

evaluate functions are always verbose. keras docs

Feature Request: KATE

Hi - your package works really smooth for image data. I wonder whether you could add KATE: K-Competitive Autoencoder for Text to your models.

Simon

Machine-readable parameters and arguments

Provide an API or similar which specifies all possible argument values for each parameter.

Add strides to convolutional layer parameters

Refactoring

Reconstruction error name in losses
Loss names
object, x, and other parameter names in generics
Allow for more parameters in layers (e.g. dropout)

Tensorflow error when training denoising AE with validation data

ValueError: Output of generator should be a tuple `(x, y, sample_weight)` or `(x, y)`. Found: [array([[0, 0, 0, 1, 0, 0, 0, 1, 1, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0],

Allow use of tied weights

Allow to tie encoder's and decoder's weights

Evaluation methods

Implement evaluation methods for performance assessment

Decoding fails after loading saved model

Calculate correct layer for encoder in variational models

Seen in #32 that ruta's encoder does not have all necessary layers.

x_train_mat <- matrix(1:200, nrow = 10)

network <-
  input() +
  dense(256, "elu") +
  variational_block(4, seed = 42, epsilon_std = .5) +
  dense(256, "elu") +
  output("sigmoid")

learner <- autoencoder_variational(network, loss = "binary_crossentropy")
model <- learner %>% train(x_train_mat, epochs = 5)

latent_model <- model$models$encoder

## Correct:
inputs <- keras::get_layer(model$models[[1]], index = 1)$input
latent_space <- keras::get_layer(model$models[[1]], index = 7)$output

latent_model <- keras::keras_model(
  inputs = inputs,
  outputs = latent_space
)

Parametrize optimizer in training

And batch_size

Name metric in evaluate functions

Currently, the evaluated metric shows up as [[2]] in the result.

Support for differently shaped data (e.g. CIFAR10)

Hello, there!

I want to use this awesome software to learning CIFAR10 structured data. Could you support them?

Best,
Seongho

Tensorflow error when using variational AE with 1-unit encoding

Sampling contractive autoencoder

https://www.icml.cc/2012/papers/910.pdf

to_keras(loss, learner) is not a valid scope name

ValueError when calling evaluate_... in a clean conda+TF 1.14 installation

Interpret latent space

Hi - in general, I'd like to use your package for topicmodeling on large scale text collections. I managed to get document level vectors with this code.

network <-
  input() +
  dense(256, "elu") +
  variational_block(4, seed = 42, epsilon_std = .5) +
  dense(256, "elu") +
  output("sigmoid")

learner <- autoencoder_variational(network, loss = "binary_crossentropy")
model <- learner %>% train(x_train_mat, epochs = 5)

# summary(learner$network)

inputs <- get_layer(model$models[[1]], index = 1)$input
latent_space <- get_layer(model$models[[1]], index = 7)$output

latent_model <- keras_model(
  inputs = inputs,
  outputs = latent_space
)

Via PCA the output looks like this:

Now I have two questions:

Do you know whether there is a better way of interpreting the latent space? LDA outputs topic proportions. I thought maybe a sigmoid or (softmax for sparsity) activation at the bottleneck. But this way the model does not learn anything useful.
Do you know how I could get word-level features that are associated with the each latent dimension?

Thanks in advance!

Best Simon

(PS: if you prefer a fully reproducible example, please let me know).

Variational autoencoder is broken

Variational autoencoder tutorial gives the following error when it's run:

stop(structure(list(message = "TypeError: in user code:\n\n File \"/home/david/.local/share/r-miniconda/envs/r-reticulate/lib/python3.8/site-packages/keras/engine/training.py\", line 1021, in train_function *\n return step_function(self, iterator)\n File \"/home/david/.local/share/r-miniconda/envs/r-reticulate/lib/python3.8/site-packages/keras/engine/training.py\", line 1010, in step_function **\n outputs = model.distribute_strategy.run(run_step, args=(data,))\n File \"/home/david/.local/share/r-miniconda/envs/r-reticulate/lib/python3.8/site-packages/keras/engine/training.py\", line 1000, in run_step **\n outputs = model.train_step(data)\n File \"/home/david/.local/share/r-miniconda/envs/r-reticulate/lib/python3.8/site-packages/keras/engine/training.py\", line 860, in train_step\n loss = self.compute_loss(x, y, y_pred, sample_weight)\n File \"/home/david/.local/share/r-miniconda/envs/r-reticulate/lib/python3.8/site-packages/keras/engine/training.py\", line 918, in compute_loss\n return self.compiled_loss(\n File \"/home/david/.local/share/r-miniconda/envs/r-reticulate/lib/python3.8/site-packages/keras/engine/compile_utils.py\", line 239, in __call__\n self._loss_metric.update_state(\n File \"/home/david/.local/share/r-miniconda/envs/r-reticulate/lib/python3.8/site-packages/keras/utils/metrics_utils.py\", line 70, in decorated\n update_op = update_state_fn(*args, **kwargs)\n File \"/home/david/.local/share/r-miniconda/envs/r-reticulate/lib/python3.8/site-packages/keras/metrics.py\", line 178, in update_state_fn\n return ag_update_state(*args, **kwargs)\n File \"/home/david/.local/share/r-miniconda/envs/r-reticulate/lib/python3.8/site-packages/keras/metrics.py\", line 455, in update_state **\n sample_weight = tf.__internal__.ops.broadcast_weights(\n File \"/home/david/.local/share/r-miniconda/envs/r-reticulate/lib/python3.8/site-packages/keras/engine/keras_tensor.py\", line 254, in __array__\n raise TypeError(\n\n TypeError: You are passing KerasTensor(type_spec=TensorSpec(shape=(), dtype=tf.float32, name=None), name='Placeholder:0', description=\"created by layer 'tf.cast_5'\"), an intermediate Keras symbolic input/output, to a TF API that does not allow registering custom dispatchers, such as `tf.cond`, `tf.function`, gradient tapes, or `tf.map_fn`. Keras Functional model construction only supports TF API calls that *do* support dispatching, such as `tf.math.add` or `tf.reshape`. Other APIs cannot be called directly on symbolic Kerasinputs/outputs. You can work around this limitation by putting the operation in a custom Keras layer `call` and calling that layer on this symbolic input/output.\n", 
call = py_call_impl(callable, dots$args, dots$keywords), 
cppstack = structure(list(file = "", line = -1L, stack = c("/home/david/R/x86_64-pc-linux-gnu-library/4.1/reticulate/libs/reticulate.so(Rcpp::exception::exception(char const*, bool)+0x74) [0x7f06a41c1524]", 
"/home/david/R/x86_64-pc-linux-gnu-library/4.1/reticulate/libs/reticulate.so(Rcpp::stop(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&)+0x29) [0x7f06a41b0bc4]", ...

Until this is fixed, users should implement variational autoencoders directly in Keras: https://keras.rstudio.com/articles/examples/variational_autoencoder.html. Sorry for any inconvenience.

Add examples to web documentation

Use knitr commands like in keras

Allow training/prediction/evaluation with generators

Tensorflow error when using softsign activation in contractive autoencoder

LookupError: No gradient defined for operation 'loss_186/activation_281_loss/loop_body/gradients/activation_278/Softsign_grad/SoftsignGrad/pfor/SoftsignGrad' (op type: SoftsignGrad)

Parametrize prior in variational autoencoder

Global weight decay

Maybe with global defaults for a whole network? Regularizers, activations and so on

Refactoring

Check which methods are exported or whether to export them at all
Method consistency
Clean code

Add deconvolution layers

Current workaround:

deconv <- function(filters, kernel_size, ...) {
  layer_keras("conv_2d_transpose", filters = filters, kernel_size = kernel_size, ...)
}

Save and load autoencoder models

Support for data augmentation

Use fit_generator to dynamically augment data

Variational autoencoder error (ruta)

I tried to run the code for VAE - just copy pasting in RStudio - and I got the following error when executing the following command.

model <- learner %>% train(x_train, epochs = 5)

Error in py_call_impl(callable, dots$args, dots$keywords) :
_SymbolicException: Inputs to eager execution function cannot be Keras symbolic tensors, but found [<tf.Tensor 'z_mean_3/Identity:0' shape=(None, 3) dtype=float32>, <tf.Tensor 'z_log_var_3/Identity:0' shape=(None, 3) dtype=float32>]

Any clue on what I am doing wrong? Thanks.

PS: Figured out from more reading. Needed the following lines:
if (tensorflow::tf$executing_eagerly())
tensorflow::tf$compat$v1$disable_eager_execution()

Working now.

Thanks.

fdavidcl / ruta Goto Github PK

ruta's Issues

Recommend Projects

Recommend Topics

Recommend Org