[409433]: / vignettes / using_tb.Rmd

Download this file

166 lines (121 with data), 4.8 kB

---
title: "Using Tensorboard"
output: rmarkdown::html_vignette
vignette: >
  %\VignetteIndexEntry{Using Tensorboard}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

```{r, echo=FALSE, warning=FALSE, message=FALSE}

if (!reticulate::py_module_available("tensorflow")) {
  knitr::opts_chunk$set(eval = FALSE)
} else {
  knitr::opts_chunk$set(eval = TRUE)
}
```

```{r, message=FALSE}
library(deepG)
library(keras)
library(magrittr)
```

```{r, echo=FALSE, warning=FALSE, message=FALSE}
options(rmarkdown.html_vignette.check_title = FALSE)
```

```{css, echo=FALSE}
mark.in {
  background-color: CornflowerBlue;
}

mark.out {
  background-color: IndianRed;
}

```


Tensorflow offers the <a href="https://www.tensorflow.org/tensorboard">Tensorboard</a> application to visualize the training process of our networks. DeepG expands on some of the default Tensorboard options and implements some custom settings.
We train again a model that can differentiate sequences based on the GC content, as described in the <a href="getting_started.html">Getting started tutorial</a>.

We start by creating our data. To show the difference between accuracy and balanced accuracy, we create 3 times more data with high GC content. 

```{r warning = FALSE}
set.seed(123)
vocabulary <- c("A", "C", "G", "T")

data_type <- c("train", "validation")

for (i in 1:length(data_type)) {
  
  temp_file <- tempfile()
  assign(paste0(data_type[i], "_dir"), temp_file)
  dir.create(temp_file)
  
  for (j in 1:6) {
    
    if (j %% 2 == 1) {
      header <- "high_gc"
      prob <- c(0.1, 0.4, 0.4, 0.1)
    } else {
      header <- "equal_dist"
      prob <- rep(0.25, 4)
    }
    fasta_name_start <- paste0(header, "_", data_type[i])
    
    create_dummy_data(file_path = temp_file,
                      num_files = 2,
                      seq_length = 100, 
                      num_seq = ifelse(j %% 2 == 1, 6, 2), # create more sequences for high GC content
                      header = header,
                      prob = prob,
                      fasta_name_start = fasta_name_start,
                      vocabulary = vocabulary)
  }
  
}
```

```{r warning = FALSE}
list.files(train_dir)
list.files(validation_dir)
```

To use tensorboard, we first need to create a folder where we store our tensorboard logs. 

```{r warning = FALSE}
# create folder for tensorboard logs
tb_dir <- tempfile()
dir.create(tb_dir)
```

When creating our model, we can add some additional metrics to observe like AUC, F1 and balanced accuracy.

```{r warning = FALSE}
maxlen <- 50
model <- create_model_lstm_cnn(maxlen = maxlen,
                               filters = c(8),
                               kernel_size = c(12),
                               pool_size = c(3),
                               layer_lstm = 8,
                               auc_metric = TRUE,
                               f1_metric = TRUE,
                               bal_acc = TRUE,
                               layer_dense = c(4, 2),
                               model_seed = 3)
```

Finally we can train the model.

```{r warning = FALSE, eval=FALSE}
hist <- train_model(model,
                    train_type = "label_header",
                    run_name = "gc_model_1",
                    path = train_dir,
                    path_val = validation_dir,
                    epochs = 5,
                    steps_per_epoch = 2, # 20
                    batch_size = 64,
                    step = 50, 
                    path_tensorboard = tb_dir, # path to tensorboard logs
                    tb_images = F, # show confusion matrix in tensorboard
                    vocabulary_label = c("high_gc", "equal_dist"))
                    
plot(hist)
```

Use the following command to open tensorboard in a browser:

```{r warning = FALSE, eval=FALSE}
keras::tensorboard(tb_dir)
```

We can observe the scores for loss, accuracy, balanced accuracy, F1 and AUC under the "SCALARS" tab. 

<img src="tb_images/loss.png" height= "300" />

<img src="tb_images/acc.png" height= "300" />

<img src="tb_images/bal_acc.png" height= "300" />

![](tb_images/f1.png)

<img src="tb_images/auc.png" height= "300" />

We can also observe how the learning rate might have changed 

<img src="tb_images/lr.png" height= "300" />

In the "training files seen" window, we can observe how often we iterated over the training files.

<img src="tb_images/files_seen.png" height= "300" />

In the "IMAGES" tab we can see a confusion matrix for the train and validation scores for every epoch

<img src="tb_images/cm_train.png" height= "300" />

<img src="tb_images/cm_val.png" height= "300" />


In the "HPARAM" tab you can see hyper parameters of each run 

<img src="tb_images/hparam.png" height= "300" />