Here, we apply CNNs to classify Hindi Handwritten Characters and Digits !
Train-Validation-Test Split
| Original Data | Total | Train | Validation | Test |
|---|---|---|---|---|
| character_1 | 2000 | 1400 | 300 | 300 |
| character_2 | 2000 | 1400 | 300 | 300 |
| … | … | … | … | … |
| digit_9 | 2000 | 1400 | 300 | 300 |
| Total | 92,000 | 63,000 | 13,500 | 13,500 |
There are 36 characters and 9 digits resulting in 45 classes. This is a Balanced Multi-class Classification Problem. Based on the above split, we can use batch_size = 250. This results in steps_per_epoch = 252 for training and 54 for validation and testing.
Setting Directory
original_dataset_dir = "C:/Users/KUNAL/Downloads/#R coding/#Books/covered/#Book - Manning - Deep Learning with R and Keras/## Article"
base_dir = "C:/Users/KUNAL/Downloads/#R coding/#Books/covered/#Book - Manning - Deep Learning with R and Keras/## Article/HindiCharacter"
train_dir = file.path(base_dir,"Training")
validation_dir = file.path(base_dir,"Validation")
testing_dir = file.path(base_dir,"Test")# Defining Hyper parameter
img_height = 32
img_width = 32
batch_size = 250
num_classes = 45
library(keras)
datagen = image_data_generator(rescale = 1/255)
train_generator <- flow_images_from_directory(train_dir,
datagen,target_size = c(img_height, img_width),
batch_size = batch_size,
class_mode = "categorical",
color_mode = "grayscale")
# 63000 images belonging to 45 classes
val_generator <- flow_images_from_directory(validation_dir,datagen,
target_size = c(img_height, img_width),
batch_size = batch_size,
class_mode = "categorical",
color_mode = "grayscale")
# 13500 images belonging to 45 classes
test_generator <- flow_images_from_directory(testing_dir,datagen,
target_size = c(img_height, img_width),
batch_size = batch_size,
class_mode = "categorical",
shuffle = F,
color_mode = "grayscale")
# 13500 images belonging to 45 classes
# Model Structure
model <- keras_model_sequential() %>%
layer_conv_2d(filters = 32, kernel_size = c(3, 3), activation = "relu",
input_shape = c(img_height, img_width, 1)) %>%
layer_max_pooling_2d(pool_size = c(2, 2),strides = c(2,2),padding = "same") %>%
layer_conv_2d(filters = 32, kernel_size = c(3, 3), activation = "relu") %>%
layer_max_pooling_2d(pool_size = c(2, 2),strides = c(2,2),padding = "same") %>%
layer_conv_2d(filters = 64, kernel_size = c(3, 3), activation = "relu") %>%
layer_max_pooling_2d(pool_size = c(2, 2),strides = c(2,2),padding = "same") %>%
layer_flatten() %>%
layer_dropout(rate = 0.5) %>%
layer_dense(units = 64, activation = "relu") %>%
layer_dense(units = num_classes, activation = "softmax")
summary(model)
# Compile
model %>% compile(loss="categorical_crossentropy",
optimizer=optimizer_adam(learning_rate = 0.001),
metrics=c("acc"))
# Callback
filepath = file.path(original_dataset_dir,"Hindi_final_model.h5")
check = callback_model_checkpoint(filepath,monitor = 'val_acc',
verbose = 1,
save_best_only = T,
mode = "max")
early = callback_early_stopping(monitor = "val_loss",
mode = "min",
patience = 3)
lr_red = callback_reduce_lr_on_plateau(monitor = "val_loss",
patience = 2,
verbose = 1,
factor = 0.3,
min_lr = 0.000001)
callback_list = list(early,lr_red,check)
# Train
history <- model %>% fit_generator(train_generator,
steps_per_epoch = 252,
epochs = 40,
callbacks = callback_list,
validation_data = val_generator,
validation_steps = 54)Callbacks used -
# Evaluate
model %>% evaluate(test_generator, steps = 54)
# 54/54 [==============================] - 7s
# 138ms/step - loss: 0.0737 - acc: 0.9784
As we can see that the first layer seems to act as an ‘edge-detector’ picking up the structure of characters and digits. As we move towards the higher layers, the representations start becoming more and more abstract. We also see that there are spaces where there were no activations at all - indicating the absence of certain filters.
| Sl. | Structures | Training | Validation | Testing | Remarks |
|---|---|---|---|---|---|
| 1 | (32,64 pool: 2,5 stride: 2,5) without dropout layer |
97.66% | 91.24% | 95.49% | Overfitting |
| 2 | (32,64 pool: 2,5 stride: 2,5) with 50% dropout |
91.18% | 93.48% | 96.41% | epochs = 10/40, lr = 0.0001 |
| 3 | (32,64 pool: 2,5 stride: 2,5) with reduced learning rate |
94.36% | 94.87% | 97.39% | epochs = 12/40, lr = 0.00009 |
| 4 | (32,32,64 pool: 2,2,5 stride: 2,2,5) with 50% dropout |
93.25% | 94.62% | 97.2% | epochs = 25/40, lr = 0.0003 |
| 5 | (32,32,64 pool: 2,2,2 stride: 2,2,2) with 50% dropout |
96.28% | 94.33% | 97.59% | epochs = 19/40, lr = 0.000027 |
| 6 | (32,64 pool: 2,2,2 stride: 2,2,2) with 50% dropout |
97.70% | 93.19% | 96.86% | epochs = 16/40, lr = 0.00009 |
| 7 | 32,32,64 - kernel 3,3,3 - pool 2,2,2 - stride 2,2,2 - classifier 64 - 50% dropout *** | 96.64% | 95.07% | 97.84% | epochs = 22/40, lr = 0.000027 |
| 8 | 32,32,64 - kernel 5,5,5 - pool 2,2,5 - stride 2,2,5 - classifier 64 - 50% dropout | 88.31% | 92.44% | 95.68% | epochs = 17/40, lr = 0.00009 |
We see that Model 7 with 3 convolution layers with filters 32,32,64 kernel 3x3, pool size 2x2 with strides of 2x2 for all layers and 50% dropout layer followed by a dense classifier with 64 neurons is the best performing model with Test Accuracy = 97.84%.
7th Model achieves a 97.84 % Test Accuracy !