📑   The Data Set of Flower Images

🌐   0. Code Library, Style and Links

Keras

Training Visualization

Getting started with imager

library(IRdisplay)
library(repr)
library(tensorflow)
library(keras)
library(imager)

display_html("<style> 
@import url('https://fonts.googleapis.com/css?family=Orbitron|Roboto');
body {background-color: gainsboro;} 
a {color: #37c9e1; font-family: 'Roboto';} 
h1 {color: #37c9e1; font-family: 'Orbitron'; text-shadow: 4px 4px 4px #ccc;} 
h2, h3 {color: slategray; font-family: 'Orbitron'; text-shadow: 4px 4px 4px #ccc;}
h4 {color: #818286; font-family: 'Roboto';}
span {text-shadow: 4px 4px 4px #ccc;}
div.output_prompt, div.output_area pre {color: slategray;}
div.input_prompt, div.output_subarea {color: #37c9e1;}      
div.output_stderr pre {background-color: gainsboro;}  
div.output_stderr {background-color: slategrey;}     
</style>")

🌐   1. Explore the Data

flowers <- read.csv("/Users/olgabelitskaya/cookbooks/flower_images/flower_labels.csv")
head(flowers)

file	label
0001.png	0
0002.png	0
0003.png	2
0004.png	0
0005.png	0
0006.png	1

flower_targets <- as.matrix(flowers["label"])
flower_targets <- keras::to_categorical(flower_targets, 10)
head(flower_targets)

1	0	0	0	0	0	0	0	0	0
1	0	0	0	0	0	0	0	0	0
0	0	1	0	0	0	0	0	0	0
1	0	0	0	0	0	0	0	0	0
1	0	0	0	0	0	0	0	0	0
0	1	0	0	0	0	0	0	0	0

image_paths <- list.files("/Users/olgabelitskaya/cookbooks/flower_images", 
                          recursive = TRUE, full.names = TRUE)
image_paths <- image_paths[1:(length(image_paths)-1)]
image_paths[1:3]

1. '/Users/olgabelitskaya/cookbooks/flower_images/0001.png'
2. '/Users/olgabelitskaya/cookbooks/flower_images/0002.png'
3. '/Users/olgabelitskaya/cookbooks/flower_images/0003.png'

img_path <- "/Users/olgabelitskaya/cookbooks/flower_images/0001.png"
img <- keras::image_load(img_path, target_size=c(128,128))
img <- image_to_array(img) / 255
img <- array_reshape(img, c(1,128,128,3))
dim(img)

1. 1
2. 128
3. 128
4. 3

gr_img <- array_reshape(grayscale(img), c(128,128))
options(repr.plot.width=4,repr.plot.height=4)
par(mar=c(2,2,2,2))
dim(gr_img)
image(c(1:128),c(1:128),gr_img,col=grey(seq(0,1,length=256)))

1. 128
2. 128

# 'imager'
im <- load.image(img_path)
options(repr.plot.width=4,repr.plot.height=4)
par(mar=c(2,2,2,2))
dim(im)
plot(im)

1. 128
2. 128
3. 1
4. 4

image_loading <- function(image_path) {
    image <- keras::image_load(image_path, target_size=c(128,128))
    image <- image_to_array(image) / 255
    image <- array_reshape(image, c(1, dim(image)))
    return(image)
}

flower_tensors <- lapply(image_paths, image_loading)
flower_tensors <- array_reshape(flower_tensors, c(210,128,128,3))

dim(flower_tensors); dim(flower_targets)

1. 210
2. 128
3. 128
4. 3

1. 210
2. 10

🌐   2. Models

# MLP
mlp_model <- keras_model_sequential()

mlp_model %>%  

layer_dense(128, input_shape=c(128*128*3)) %>%  
layer_activation("relu") %>%  
layer_batch_normalization() %>%  

layer_dense(256) %>%  
layer_activation("relu") %>%  
layer_batch_normalization() %>%

layer_dense(512) %>%  
layer_activation("relu") %>%  
layer_batch_normalization() %>%

layer_dense(1024) %>%  
layer_activation("relu") %>%  
layer_dropout(0.2) %>%
  
layer_dense(10) %>%    
layer_activation("softmax")

mlp_model %>%
  compile(loss="categorical_crossentropy",optimizer="adam",metrics="accuracy")

summary(mlp_model)

________________________________________________________________________________
Layer (type)                        Output Shape                    Param #     
================================================================================
dense_16 (Dense)                    (None, 128)                     6291584     
________________________________________________________________________________
activation_16 (Activation)          (None, 128)                     0           
________________________________________________________________________________
batch_normalization_10 (BatchNormal (None, 128)                     512         
________________________________________________________________________________
dense_17 (Dense)                    (None, 256)                     33024       
________________________________________________________________________________
activation_17 (Activation)          (None, 256)                     0           
________________________________________________________________________________
batch_normalization_11 (BatchNormal (None, 256)                     1024        
________________________________________________________________________________
dense_18 (Dense)                    (None, 512)                     131584      
________________________________________________________________________________
activation_18 (Activation)          (None, 512)                     0           
________________________________________________________________________________
batch_normalization_12 (BatchNormal (None, 512)                     2048        
________________________________________________________________________________
dense_19 (Dense)                    (None, 1024)                    525312      
________________________________________________________________________________
activation_19 (Activation)          (None, 1024)                    0           
________________________________________________________________________________
dropout_4 (Dropout)                 (None, 1024)                    0           
________________________________________________________________________________
dense_20 (Dense)                    (None, 10)                      10250       
________________________________________________________________________________
activation_20 (Activation)          (None, 10)                      0           
================================================================================
Total params: 6,995,338
Trainable params: 6,993,546
Non-trainable params: 1,792
________________________________________________________________________________

mlp_fit <- mlp_model %>%
  fit(
    x=array_reshape(flower_tensors, c(210,128*128*3)),
    y=flower_targets,
    shuffle=T,
    batch_size=64,
    validation_split=0.1,
    epochs=30
  )

options(repr.plot.width=9,repr.plot.height=9)
plot(mlp_fit)

mlp_fit_df <- as.data.frame(mlp_fit)
mlp_fit_df[31:60,1:4]

	epoch	value	metric	data
31	1	0.2380952	acc	validation
32	2	0.2380952	acc	validation
33	3	0.3333333	acc	validation
34	4	0.2857143	acc	validation
35	5	0.3333333	acc	validation
36	6	0.4285714	acc	validation
37	7	0.4761905	acc	validation
38	8	0.4761905	acc	validation
39	9	0.4761905	acc	validation
40	10	0.4761905	acc	validation
41	11	0.4285714	acc	validation
42	12	0.4761905	acc	validation
43	13	0.5238096	acc	validation
44	14	0.4761905	acc	validation
45	15	0.4761905	acc	validation
46	16	0.4761905	acc	validation
47	17	0.4761905	acc	validation
48	18	0.4761905	acc	validation
49	19	0.4761905	acc	validation
50	20	0.4761905	acc	validation
51	21	0.4761905	acc	validation
52	22	0.4761905	acc	validation
53	23	0.4761905	acc	validation
54	24	0.4761905	acc	validation
55	25	0.4761905	acc	validation
56	26	0.4761905	acc	validation
57	27	0.4285714	acc	validation
58	28	0.4285714	acc	validation
59	29	0.4761905	acc	validation
60	30	0.4761905	acc	validation

# CNN
cnn_model <- keras_model_sequential()

cnn_model %>%  
  
layer_conv_2d(filter=32,kernel_size=c(5,5),padding="same",
              input_shape=c(128,128,3) ) %>%  
layer_activation("relu") %>%  
  
layer_max_pooling_2d(pool_size=c(2,2)) %>%  
layer_dropout(0.25) %>%

layer_conv_2d(filter=96,kernel_size=c(5,5),padding="same") %>% 
layer_activation("relu") %>%  

layer_max_pooling_2d(pool_size=c(2,2)) %>%  
layer_dropout(0.25) %>%

layer_global_average_pooling_2d() %>%  

layer_dense(512) %>%  
layer_activation("tanh") %>%  
layer_dropout(0.25) %>%  

layer_dense(128) %>%  
layer_activation("tanh") %>%  
layer_dropout(0.25) %>%
  
layer_dense(10) %>%    
layer_activation("softmax")

cnn_model %>%
  compile(loss="categorical_crossentropy",optimizer="nadam",metrics="accuracy")

summary(cnn_model)

________________________________________________________________________________
Layer (type)                        Output Shape                    Param #     
================================================================================
conv2d_3 (Conv2D)                   (None, 128, 128, 32)            2432        
________________________________________________________________________________
activation_26 (Activation)          (None, 128, 128, 32)            0           
________________________________________________________________________________
max_pooling2d_3 (MaxPooling2D)      (None, 64, 64, 32)              0           
________________________________________________________________________________
dropout_9 (Dropout)                 (None, 64, 64, 32)              0           
________________________________________________________________________________
conv2d_4 (Conv2D)                   (None, 64, 64, 96)              76896       
________________________________________________________________________________
activation_27 (Activation)          (None, 64, 64, 96)              0           
________________________________________________________________________________
max_pooling2d_4 (MaxPooling2D)      (None, 32, 32, 96)              0           
________________________________________________________________________________
dropout_10 (Dropout)                (None, 32, 32, 96)              0           
________________________________________________________________________________
global_average_pooling2d_2 (GlobalA (None, 96)                      0           
________________________________________________________________________________
dense_24 (Dense)                    (None, 512)                     49664       
________________________________________________________________________________
activation_28 (Activation)          (None, 512)                     0           
________________________________________________________________________________
dropout_11 (Dropout)                (None, 512)                     0           
________________________________________________________________________________
dense_25 (Dense)                    (None, 128)                     65664       
________________________________________________________________________________
activation_29 (Activation)          (None, 128)                     0           
________________________________________________________________________________
dropout_12 (Dropout)                (None, 128)                     0           
________________________________________________________________________________
dense_26 (Dense)                    (None, 10)                      1290        
________________________________________________________________________________
activation_30 (Activation)          (None, 10)                      0           
================================================================================
Total params: 195,946
Trainable params: 195,946
Non-trainable params: 0
________________________________________________________________________________

cnn_fit <- cnn_model %>%
  fit(
    x=flower_tensors,
    y=flower_targets,
    shuffle=T,
    batch_size=16,
    validation_split=0.1,
    epochs=30
  )

options(repr.plot.width=9,repr.plot.height=9)
plot(cnn_fit)

cnn_fit_df <- as.data.frame(cnn_fit)
cnn_fit_df[31:60,1:4]

	epoch	value	metric	data
31	1	0.0952381	acc	validation
32	2	0.2857143	acc	validation
33	3	0.1428571	acc	validation
34	4	0.1904762	acc	validation
35	5	0.3333333	acc	validation
36	6	0.4761905	acc	validation
37	7	0.4761905	acc	validation
38	8	0.4761905	acc	validation
39	9	0.4761905	acc	validation
40	10	0.4761905	acc	validation
41	11	0.5714286	acc	validation
42	12	0.5238095	acc	validation
43	13	0.6666667	acc	validation
44	14	0.5714286	acc	validation
45	15	0.4761905	acc	validation
46	16	0.5714286	acc	validation
47	17	0.2857143	acc	validation
48	18	0.5238095	acc	validation
49	19	0.6666667	acc	validation
50	20	0.5238095	acc	validation
51	21	0.5238095	acc	validation
52	22	0.7619048	acc	validation
53	23	0.7619048	acc	validation
54	24	0.6190476	acc	validation
55	25	0.6666667	acc	validation
56	26	0.5238095	acc	validation
57	27	0.7619048	acc	validation
58	28	0.6666667	acc	validation
59	29	0.5714286	acc	validation
60	30	0.7619048	acc	validation

# RNN
rnn_model <- keras_model_sequential()

rnn_model %>%  

layer_lstm(128, return_sequences=T, input_shape=c(1,128*128*3)) %>%  
layer_lstm(128) %>%

layer_dense(512) %>%  
layer_activation("relu") %>%
# layer_dropout(0.2) %>%

layer_dense(10) %>%    
layer_activation("softmax")

rnn_model %>%
  compile(loss="categorical_crossentropy",optimizer="adam",metrics="accuracy")

summary(rnn_model)

________________________________________________________________________________
Layer (type)                        Output Shape                    Param #     
================================================================================
lstm_22 (LSTM)                      (None, 1, 128)                  25231872    
________________________________________________________________________________
lstm_23 (LSTM)                      (None, 128)                     131584      
________________________________________________________________________________
dense_35 (Dense)                    (None, 512)                     66048       
________________________________________________________________________________
activation_39 (Activation)          (None, 512)                     0           
________________________________________________________________________________
dense_36 (Dense)                    (None, 10)                      5130        
________________________________________________________________________________
activation_40 (Activation)          (None, 10)                      0           
================================================================================
Total params: 25,434,634
Trainable params: 25,434,634
Non-trainable params: 0
________________________________________________________________________________

rnn_fit <- rnn_model %>%
  fit(
    x=array_reshape(flower_tensors, c(210,1,128*128*3)),
    y=flower_targets,
    shuffle=T,
    batch_size=64,
    validation_split=0.1,
    epochs=30
  )

options(repr.plot.width=9,repr.plot.height=9)
plot(rnn_fit)

rnn_fit_df <- as.data.frame(rnn_fit)
rnn_fit_df[31:60,1:4]

	epoch	value	metric	data
31	1	0.0952381	acc	validation
32	2	0.0952381	acc	validation
33	3	0.1428571	acc	validation
34	4	0.0952381	acc	validation
35	5	0.0952381	acc	validation
36	6	0.0952381	acc	validation
37	7	0.0952381	acc	validation
38	8	0.1428571	acc	validation
39	9	0.1428571	acc	validation
40	10	0.1428571	acc	validation
41	11	0.1428571	acc	validation
42	12	0.1428571	acc	validation
43	13	0.1428571	acc	validation
44	14	0.0952381	acc	validation
45	15	0.2380952	acc	validation
46	16	0.1904762	acc	validation
47	17	0.3333333	acc	validation
48	18	0.1904762	acc	validation
49	19	0.2857143	acc	validation
50	20	0.3333333	acc	validation
51	21	0.2857143	acc	validation
52	22	0.1428571	acc	validation
53	23	0.2857143	acc	validation
54	24	0.2857143	acc	validation
55	25	0.3333333	acc	validation
56	26	0.3333333	acc	validation
57	27	0.2857143	acc	validation
58	28	0.2380952	acc	validation
59	29	0.3333333	acc	validation
60	30	0.3333333	acc	validation