Benchmarking CNTK on Keras: is it Better at Deep Learning than TensorFlow?

This R Notebook is the complement to my blog post Benchmarking CNTK on Keras: is it Better at Deep Learning than TensorFlow?.

This notebook is licensed under the MIT License. If you use the code or data visualization designs contained within this notebook, it would be greatly appreciated if proper attribution is given back to this notebook and/or myself. Thanks! :)

library(readr)
library(dplyr)

Attaching package: ‘dplyr’

The following objects are masked from ‘package:stats’:

    filter, lag

The following objects are masked from ‘package:base’:

    intersect, setdiff, setequal, union

library(ggplot2)

Stackoverflow is a great place to get help: http://stackoverflow.com/tags/ggplot2.

library(plotly)

Attaching package: ‘plotly’

The following object is masked from ‘package:ggplot2’:

    last_plot

The following object is masked from ‘package:stats’:

    filter

The following object is masked from ‘package:graphics’:

    layout

library(boot)
library(scales)

Attaching package: ‘scales’

The following object is masked from ‘package:readr’:

    col_factor

library(htmlwidgets)
sessionInfo()

R version 3.4.0 (2017-04-21)
Platform: x86_64-apple-darwin15.6.0 (64-bit)
Running under: macOS Sierra 10.12.5

Matrix products: default
BLAS: /System/Library/Frameworks/Accelerate.framework/Versions/A/Frameworks/vecLib.framework/Versions/A/libBLAS.dylib
LAPACK: /Library/Frameworks/R.framework/Versions/3.4/Resources/lib/libRlapack.dylib

locale:
[1] en_US.UTF-8/en_US.UTF-8/en_US.UTF-8/C/en_US.UTF-8/en_US.UTF-8

attached base packages:
[1] stats     graphics  grDevices utils     datasets  methods   base     

other attached packages:
[1] htmlwidgets_0.8    scales_0.4.1       boot_1.3-19        plotly_4.7.0       ggplot2_2.2.1.9000
[6] dplyr_0.5.0        readr_1.1.1       

loaded via a namespace (and not attached):
 [1] Rcpp_0.12.11      knitr_1.16        magrittr_1.5      hms_0.3           munsell_0.4.3    
 [6] viridisLite_0.2.0 colorspace_1.3-2  R6_2.2.1          rlang_0.1.1       httr_1.2.1       
[11] plyr_1.8.4        tools_3.4.0       grid_3.4.0        data.table_1.10.4 gtable_0.2.0     
[16] DBI_0.6-1         htmltools_0.3.6   lazyeval_0.2.0    assertthat_0.2.0  digest_0.6.12    
[21] tibble_1.3.3      tidyr_0.6.3       purrr_0.2.2.2     compiler_3.4.0    jsonlite_1.5     

file_path <- "/Users/maxwoolf/Dropbox/PythonProjects/keras-cntk-benchmark/logs/"

CNTK is blue, TensorFlow is red.

framework_colors <- c(CNTK="#2980b9",TensorFlow="#c0392b")

1 IMDb

1.1 imdb_bidirectional_lstm

df_imdb_bi_cntk <- read_csv(paste0(file_path, "imdb_bidirectional_lstm_cntk.csv"))

Parsed with column specification:
cols(
  epoch = col_integer(),
  elapsed = col_double(),
  loss = col_double(),
  acc = col_double(),
  val_loss = col_double(),
  val_acc = col_double()
)

df_imdb_bi_tf <- read_csv(paste0(file_path, "imdb_bidirectional_lstm_tensorflow.csv"))

Parsed with column specification:
cols(
  epoch = col_integer(),
  elapsed = col_double(),
  loss = col_double(),
  acc = col_double(),
  val_loss = col_double(),
  val_acc = col_double()
)

Test the bootstrapping functionality via boot.

elapsed_mean <- function(data, indices){
  d <- data[indices,]
  mean(d$elapsed)
}
(df_imdb_bi_cntk %>% boot(elapsed_mean, 1000) %>% boot.ci(type="bca"))$bca[4:5]

[1] 149.4228 156.4935

1.1.1 Test Acc

df <- df_imdb_bi_cntk %>% mutate(framework = 'CNTK', epoch = epoch + 1) %>% union(
        df_imdb_bi_tf %>% mutate(framework = 'TensorFlow', epoch = epoch + 1)
)
df

plot <- ggplot(df, aes(epoch, val_acc, color=framework)) +
          geom_line() +
          geom_point(size=2) + 
          theme_minimal(base_family="Source Sans Pro", base_size=14) +
          scale_y_continuous(labels = percent) +
          scale_color_manual(values=framework_colors) +
          labs(title = "Performance of Bidirectional LSTM Approach on IMDb Data",
               x = "Epoch",
               y = "Test Accuracy (Higher is Better)",
               color = "")
plot

plot %>% ggplotly(width="100%", height=400)

plot %>% ggplotly(width="100%", height=400) %>%
    saveWidget("keras-1.html", selfcontained=F, libdir="plotly")

1.1.2 Speed

df_speed <- df %>% group_by(framework) %>%
  do({
    boot <- boot.ci(boot(., elapsed_mean, 1000), type="bca")$bca
    data.frame(
      mean = mean(.$elapsed),
      low_ci = boot[4],
      high_ci = boot[5]
    )
    }) %>%
  ungroup() %>%
  mutate(framework = factor(framework))
df_speed

plot <- ggplot(df_speed, aes(x=framework, y=mean, ymin=low_ci, ymax=high_ci, fill=framework)) +
          geom_bar(stat="identity", width=0.25) +
          geom_errorbar(width = 0.25) +
          theme_minimal(base_family="Source Sans Pro", base_size=14) +
          scale_fill_manual(values=framework_colors) +
          labs(title = "Speed of Bidirectional LSTM Approach on IMDb Data",
               x = "Keras Backend",
               y = "Average Epoch Runtime (seconds)",
               fill = "")
plot

plot %>% ggplotly(tooltip=c("x", "y"), width="100%")

plot %>% ggplotly(tooltip=c("x", "y"), width="100%", height=400) %>%
    saveWidget("keras-2.html", selfcontained=F, libdir="plotly")

speed_imdb_bi = (df_speed$mean[2] / df_speed$mean[1]) %>% round(2)

CNTK runs at 1.82x the runtime of TensorFlow.

1.2 FastText

df_imdb_ft_cntk <- read_csv(paste0(file_path, "imdb_fasttext_cntk.csv"))

Parsed with column specification:
cols(
  epoch = col_integer(),
  elapsed = col_double(),
  loss = col_double(),
  acc = col_double(),
  val_loss = col_double(),
  val_acc = col_double()
)

df_imdb_ft_tf <- read_csv(paste0(file_path, "imdb_fasttext_tensorflow.csv"))

Parsed with column specification:
cols(
  epoch = col_integer(),
  elapsed = col_double(),
  loss = col_double(),
  acc = col_double(),
  val_loss = col_double(),
  val_acc = col_double()
)

1.2.1 Test Acc

df <- df_imdb_ft_cntk %>% mutate(framework = 'CNTK', epoch = epoch + 1) %>% union(
        df_imdb_ft_tf %>% mutate(framework = 'TensorFlow', epoch = epoch + 1)
)
df

plot <- ggplot(df, aes(epoch, val_acc, color=framework)) +
          geom_line() +
          geom_point(size=2) + 
          theme_minimal(base_family="Source Sans Pro", base_size=14) +
          scale_y_continuous(labels = percent) +
          scale_color_manual(values=framework_colors) +
          labs(title = "Performance of fasttext Approach on IMDb Data",
               x = "Epoch",
               y = "Test Accuracy (Higher is Better)",
               color= "")
plot %>% ggplotly()

plot %>% ggplotly(width="100%", height=400) %>%
    saveWidget("keras-3.html", selfcontained=F, libdir="plotly")

1.2.2 Speed

df_speed <- df %>% group_by(framework) %>%
  do({
    boot <- boot.ci(boot(., elapsed_mean, 1000), type="bca")$bca
    data.frame(
      mean = mean(.$elapsed),
      low_ci = boot[4],
      high_ci = boot[5]
    )
    }) %>%
  ungroup() %>%
  mutate(framework = factor(framework))
df_speed

plot <- ggplot(df_speed, aes(x=framework, y=mean, ymin=low_ci, ymax=high_ci, fill=framework)) +
          geom_bar(stat="identity", width=0.5) +
          geom_errorbar(width = 0.25) +
          theme_minimal(base_family="Source Sans Pro", base_size=14) +
          scale_fill_manual(values=framework_colors) +
          labs(title = "Speed of fasttext Approach on IMDb Data",
               x = "Keras Backend",
               y = "Average Epoch Runtime (seconds)",
               fill="")
plot %>% ggplotly(tooltip=c("x", "y"))

plot %>% ggplotly(tooltip=c("x", "y"), width="100%", height=400) %>%
    saveWidget("keras-4.html", selfcontained=F, libdir="plotly")

speed_imdb_fasttext = (df_speed$mean[2] / df_speed$mean[1]) %>% round(2)

CNTK runs at 0.84x the runtime of TensorFlow.

2 MNIST

2.1 MLP

df_mnist_mlp_cntk <- read_csv(paste0(file_path, "mnist_mlp_cntk.csv"))

Parsed with column specification:
cols(
  epoch = col_integer(),
  elapsed = col_double(),
  loss = col_double(),
  acc = col_double(),
  val_loss = col_double(),
  val_acc = col_double()
)

df_mnist_mlp_tf <- read_csv(paste0(file_path, "mnist_mlp_tensorflow.csv"))

Parsed with column specification:
cols(
  epoch = col_integer(),
  elapsed = col_double(),
  loss = col_double(),
  acc = col_double(),
  val_loss = col_double(),
  val_acc = col_double()
)

2.1.1 Test Acc

df <- df_mnist_mlp_cntk %>% mutate(framework = 'CNTK', epoch = epoch + 1) %>% union(
        df_mnist_mlp_tf %>% mutate(framework = 'TensorFlow', epoch = epoch + 1)
)
df

plot <- ggplot(df, aes(epoch, val_acc, color=framework)) +
          geom_line() +
          geom_point(size=2) + 
          theme_minimal(base_family="Source Sans Pro", base_size=14) +
          scale_color_manual(values=framework_colors) +
          scale_y_continuous(labels = percent) +
          labs(title = "Performance of MLP Approach on MNIST Data",
               x = "Epoch",
               y = "Test Accuracy (Higher is Better)",
               color="")
plot %>% ggplotly()

plot %>% ggplotly(width="100%", height=400) %>%
    saveWidget("keras-5.html", selfcontained=F, libdir="plotly")

2.1.2 Speed

df_speed <- df %>% group_by(framework) %>%
  do({
    boot <- boot.ci(boot(., elapsed_mean, 1000), type="bca")$bca
    data.frame(
      mean = mean(.$elapsed),
      low_ci = boot[4],
      high_ci = boot[5]
    )
    }) %>%
  ungroup() %>%
  mutate(framework = factor(framework))

|============================================================================|100% ~0 s remaining     

df_speed

plot <- ggplot(df_speed, aes(x=framework, y=mean, ymin=low_ci, ymax=high_ci, fill=framework)) +
          geom_bar(stat="identity", width=0.5) +
          geom_errorbar(width = 0.25) +
          theme_minimal(base_family="Source Sans Pro", base_size=14) +
          scale_fill_manual(values=framework_colors) +
          labs(title = "Speed of MLP Approach on MNIST Data",
               x = "Keras Backend",
               y = "Average Epoch Runtime (seconds)",
               fill="")
plot %>% ggplotly(tooltip=c("x", "y"))

plot %>% ggplotly(tooltip=c("x", "y"), width="100%", height=400) %>%
    saveWidget("keras-6.html", selfcontained=F, libdir="plotly")

speed_mnist_mlp = (df_speed$mean[2] / df_speed$mean[1]) %>% round(2)

CNTK runs at 1.22x the runtime of TensorFlow.

2.2 CNN

df_mnist_cnn_cntk <- read_csv(paste0(file_path, "mnist_cnn_cntk.csv"))

Parsed with column specification:
cols(
  epoch = col_integer(),
  elapsed = col_double(),
  loss = col_double(),
  acc = col_double(),
  val_loss = col_double(),
  val_acc = col_double()
)

df_mnist_cnn_tf <- read_csv(paste0(file_path, "mnist_cnn_tensorflow.csv"))

Parsed with column specification:
cols(
  epoch = col_integer(),
  elapsed = col_double(),
  loss = col_double(),
  acc = col_double(),
  val_loss = col_double(),
  val_acc = col_double()
)

2.2.1 Test Acc

df <- df_mnist_cnn_cntk %>% mutate(framework = 'CNTK', epoch = epoch + 1) %>% union(
        df_mnist_cnn_tf %>% mutate(framework = 'TensorFlow', epoch = epoch + 1)
)
df

plot <- ggplot(df, aes(epoch, val_acc, color=framework)) +
          geom_line() +
          geom_point(size=2) + 
          theme_minimal(base_family="Source Sans Pro", base_size=14) +
          scale_y_continuous(labels = percent) +
          scale_color_manual(values=framework_colors) +
          labs(title = "Performance of CNN Approach on MNIST Data",
               x = "Epoch",
               y = "Test Accuracy (Higher is Better)",
               color="")
plot %>% ggplotly()

plot %>% ggplotly(width="100%", height=400) %>%
    saveWidget("keras-7.html", selfcontained=F, libdir="plotly")

2.2.2 Speed

df_speed <- df %>% group_by(framework) %>%
  do({
    boot <- boot.ci(boot(., elapsed_mean, 1000), type="bca")$bca
    data.frame(
      mean = mean(.$elapsed),
      low_ci = boot[4],
      high_ci = boot[5]
    )
    }) %>%
  ungroup() %>%
  mutate(framework = factor(framework))

extreme order statistics used as endpoints

df_speed

plot <- ggplot(df_speed, aes(x=framework, y=mean, ymin=low_ci, ymax=high_ci, fill=framework)) +
          geom_bar(stat="identity", width=0.5) +
          geom_errorbar(width = 0.25) +
          theme_minimal(base_family="Source Sans Pro", base_size=14) +
          scale_fill_manual(values=framework_colors) +
          labs(title = "Speed of CNN Approach on MNIST Data",
               x = "Keras Backend",
               y = "Average Epoch Runtime (seconds)",
               fill = "")
plot %>% ggplotly(tooltip=c("x", "y"))

plot %>% ggplotly(tooltip=c("x", "y"), width="100%", height=400) %>%
    saveWidget("keras-8.html", selfcontained=F, libdir="plotly")

speed_mnist_cnn = (df_speed$mean[2] / df_speed$mean[1]) %>% round(2)

CNTK runs at 0.7x the runtime of TensorFlow.

3 CIFAR:

3.1 CNN

df_cifar_cnn_cntk <- read_csv(paste0(file_path, "cifar10_cnn_cntk.csv"))

Parsed with column specification:
cols(
  epoch = col_integer(),
  elapsed = col_double(),
  loss = col_double(),
  acc = col_double(),
  val_loss = col_double(),
  val_acc = col_double()
)

df_cifar_cnn_tf <- read_csv(paste0(file_path, "cifar10_cnn_tensorflow.csv"))

Parsed with column specification:
cols(
  epoch = col_integer(),
  elapsed = col_double(),
  loss = col_double(),
  acc = col_double(),
  val_loss = col_double(),
  val_acc = col_double()
)

3.1.1 Test Acc

df <- df_cifar_cnn_cntk %>% mutate(framework = 'CNTK', epoch = epoch + 1) %>% union(
        df_cifar_cnn_tf %>% mutate(framework = 'TensorFlow', epoch = epoch + 1)
)
df

plot <- ggplot(df, aes(epoch, val_acc, color=framework)) +
          geom_line() +
          geom_point(size=2) + 
          theme_minimal(base_family="Source Sans Pro", base_size=14) +
          scale_y_continuous(labels = percent) +
          scale_color_manual(values=framework_colors) +
          labs(title = "Performance of CNN Approach on CIFAR-10 Data",
               x = "Epoch",
               y = "Test Accuracy (Higher is Better)",
               color="")
plot %>% ggplotly()

plot %>% ggplotly(width="100%", height=400) %>%
    saveWidget("keras-9.html", selfcontained=F, libdir="plotly")

3.1.2 Speed

df_speed <- df %>% group_by(framework) %>%
  do({
    boot <- boot.ci(boot(., elapsed_mean, 1000), type="bca")$bca
    data.frame(
      mean = mean(.$elapsed),
      low_ci = boot[4],
      high_ci = boot[5]
    )
    }) %>%
  ungroup() %>%
  mutate(framework = factor(framework))
df_speed

plot <- ggplot(df_speed, aes(x=framework, y=mean, ymin=low_ci, ymax=high_ci, fill=framework)) +
          geom_bar(stat="identity", width=0.5) +
          geom_errorbar(width = 0.25) +
          theme_minimal(base_family="Source Sans Pro", base_size=14) +
          scale_fill_manual(values=framework_colors) +
          labs(title = "Speed of CNN Approach on CIFAR-10 Data",
               x = "Keras Backend",
               y = "Average Epoch Runtime (seconds)",
               fill = "")
plot %>% ggplotly(tooltip=c("x", "y"))

plot %>% ggplotly(tooltip=c("x", "y"), width="100%", height=400) %>%
    saveWidget("keras-10.html", selfcontained=F, libdir="plotly")

speed_cifar_cnn = (df_speed$mean[2] / df_speed$mean[1]) %>% round(2)

CNTK runs at 0.97x the runtime of TensorFlow.; not statistically significant

4 Text Generation

4.1 CNN

df_gen_lstm_cntk <- read_csv(paste0(file_path, "lstm_text_generation_cntk.csv"))

Parsed with column specification:
cols(
  epoch = col_integer(),
  elapsed = col_double(),
  loss = col_double(),
  acc = col_character(),
  val_loss = col_character(),
  val_acc = col_character()
)

df_gen_lstm_tf <- read_csv(paste0(file_path, "lstm_text_generation_tensorflow.csv"))

Parsed with column specification:
cols(
  epoch = col_integer(),
  elapsed = col_double(),
  loss = col_double(),
  acc = col_character(),
  val_loss = col_character(),
  val_acc = col_character()
)

4.1.1 Test Acc

df <- df_gen_lstm_cntk %>% mutate(framework = 'CNTK', epoch = epoch + 1) %>% union(
        df_gen_lstm_tf %>% mutate(framework = 'TensorFlow', epoch = epoch + 1)
)
df

plot <- ggplot(df, aes(epoch, loss, color=framework)) +
          geom_line() +
          geom_point(size=2) + 
          theme_minimal(base_family="Source Sans Pro", base_size=14) +
          scale_color_manual(values=framework_colors) +
          labs(title = "Performance of Text Generation via LSTM",
               x = "Epoch",
               y = "Loss (Lower is Better)",
               color="")
plot %>% ggplotly()

plot %>% ggplotly(width="100%", height=400) %>%
    saveWidget("keras-11.html", selfcontained=F, libdir="plotly")

4.1.2 Speed

df_speed <- df %>% group_by(framework) %>%
  do({
    boot <- boot.ci(boot(., elapsed_mean, 1000), type="bca")$bca
    data.frame(
      mean = mean(.$elapsed),
      low_ci = boot[4],
      high_ci = boot[5]
    )
    }) %>%
  ungroup() %>%
  mutate(framework = factor(framework))
df_speed

plot <- ggplot(df_speed, aes(x=framework, y=mean, ymin=low_ci, ymax=high_ci, fill=framework)) +
          geom_bar(stat="identity", width=0.5) +
          geom_errorbar(width = 0.25) +
          theme_minimal(base_family="Source Sans Pro", base_size=14) +
          scale_fill_manual(values=framework_colors) +
          labs(title = "Speed of CNN of Text Generation via LSTM",
               x = "Keras Backend",
               y = "Average Epoch Runtime (seconds)",
               fill = "")
plot %>% ggplotly(tooltip=c("x", "y"))

plot %>% ggplotly(tooltip=c("x", "y"), width="100%", height=400) %>%
    saveWidget("keras-12.html", selfcontained=F, libdir="plotly")

speed_lstm_gen = (df_speed$mean[2] / df_speed$mean[1]) %>% round(2)

CNTK runs at 1.89x the runtime of TensorFlow.

4.2 Custom Keras

df_gen_keras_cntk <- read_csv(paste0(file_path, "text_generator_keras_cntk.csv"))

Parsed with column specification:
cols(
  epoch = col_integer(),
  elapsed = col_double(),
  loss = col_double()
)

df_gen_keras_tf <- read_csv(paste0(file_path, "text_generator_keras_tensorflow.csv"))

Parsed with column specification:
cols(
  epoch = col_integer(),
  elapsed = col_double(),
  loss = col_double()
)

4.2.1 Test Acc

df <- df_gen_keras_cntk %>% mutate(framework = 'CNTK', epoch = epoch + 1) %>% union(
        df_gen_keras_tf %>% mutate(framework = 'TensorFlow', epoch = epoch + 1)
)
df

plot <- ggplot(df, aes(epoch, loss, color=framework)) +
          geom_line() +
          geom_point(size=2) + 
          theme_minimal(base_family="Source Sans Pro", base_size=14) +
          scale_color_manual(values=framework_colors) +
          labs(title = "Performance of Text Generation via Custom Keras Model",
               x = "Epoch",
               y = "Loss (Lower is Better)",
               color="")
plot %>% ggplotly()

plot %>% ggplotly(width="100%", height=400) %>%
    saveWidget("keras-13.html", selfcontained=F, libdir="plotly")

4.2.2 Speed

df_speed <- df %>% group_by(framework) %>%
  do({
    boot <- boot.ci(boot(., elapsed_mean, 1000), type="bca")$bca
    data.frame(
      mean = mean(.$elapsed),
      low_ci = boot[4],
      high_ci = boot[5]
    )
    }) %>%
  ungroup() %>%
  mutate(framework = factor(framework))
df_speed

plot <- ggplot(df_speed, aes(x=framework, y=mean, ymin=low_ci, ymax=high_ci, fill=framework)) +
          geom_bar(stat="identity", width=0.5) +
          geom_errorbar(width = 0.25) +
          theme_minimal(base_family="Source Sans Pro", base_size=14) +
          scale_fill_manual(values=framework_colors) +
          labs(title = "Speed of Text Generation via Custom Keras Model",
               x = "Keras Backend",
               y = "Average Epoch Runtime (seconds)",
               fill = "")
plot %>% ggplotly(tooltip=c("x", "y"))

plot %>% ggplotly(tooltip=c("x", "y"), width="100%", height=400) %>%
    saveWidget("keras-14.html", selfcontained=F, libdir="plotly")

speed_keras_gen = (df_speed$mean[2] / df_speed$mean[1]) %>% round(2)

CNTK runs at 1.08x the runtime of TensorFlow.

4.3 LICENSE

MIT License

Copyright (c) 2017 Max Woolf

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.