WaveNet genomic

gistfile1.txt

trainMinimalFunctionalAPI <- function(path = "example_files/fasta") {
  library(wavenet)
  message("Initialize model! This can take a few minutes.")

  maxlen <- 1000
  input <- keras::layer_input(batch_shape = c(64, maxlen, 6))
  

  # https://github.com/ibab/tensorflow-wavenet/blob/master/wavenet/ops.py#L46
  first <- keras::layer_conv_1d(
    object = input,
    filters = 32,
    kernel_size = 2,
    padding = "causal",
    use_bias = FALSE
  )
  
  skip_connections <- NULL
  residual_blocks <- 2^rep(1:8, 3)
  if (length(residual_blocks) == 1) {
    dilation_rates <- 2^seq_len(residual_blocks)
  } else {
    dilation_rates <- residual_blocks
  }
  for (i in dilation_rates) {
    out <- layer_wavenet_dilated_causal_convolution_1d(
      first,
      filters = 32,
      kernel_size = 32,
      dilation_rate = i
    )
    x <- out[[1]]
    s <- out[[2]]
    skip_connections <- append(skip_connections, s)
  }
  
  out1 = keras::layer_add(skip_connections) 
  out2 = keras::layer_activation(object = out1,
                                 activation = "relu")
  out3 =  keras::layer_conv_1d(object = out2,
                               filters = 32/2L,
                               kernel_size = 1,
                               activation = "relu",
                               use_bias = FALSE)
 
  pooling = keras::layer_global_max_pooling_1d(object = out3)
  dense = keras::layer_dense(object = pooling, 6)
  output = keras::layer_activation(object = dense, "softmax")
  model <- keras::keras_model(input, output)
  
  summary(model)
  
  model %>% keras::compile(loss = "categorical_crossentropy",
                           optimizer = "adam",
                           metrics = c("acc"))
  

  gen <-
    fastaFileGenerator(
      corpus.dir = path,
      batch.size = 64,
      maxlen = maxlen,
      step = 1,
      randomFiles = FALSE,
      seqStart = "l",
      seqEnd = "l",
      withinFile =  "p",
      vocabulary = c("l", "p", "a", "c", "g", "t")
    )


  message("Start training ...")
  history <-
    model %>% keras::fit_generator(
      generator = gen,
      steps_per_epoch = 200,
      max_queue_size = 200,
      epochs = 50
      )
  
}