I would like to classify images using Keras and Tensorflow in RStudio.
In a normal image classification model - how can I tell Tensorflow to take or NOT to take into account the position within the picture?
For example, if I want the model to say, whether or not there is a dog in the picture versus whether or not there is a dog in the bottom left corner of the picture?
I know that usually to NOT take it into account, I can for example offset my pictures and permutate them so the model learns that objects can be at all positions in the picture.
But how can I get it to do exactly that? To learn that a feature is only valid, if it is at the same position in the picture as it has found it in the training?
Example code (general image classification):
model <- keras_model_sequential()
model %>%
layer_dropout(rate = FLAGS$dropout1) %>%
layer_conv_2d(filters = FLAGS$convol_filters1, kernel_size = c(3,3)
, activation = "relu", padding = "same"
) %>%
layer_conv_2d(filters = FLAGS$convol_filters1, kernel_size = c(3,3)
, activation = "relu"
#, padding = "same"
) %>%
layer_max_pooling_2d(pool_size = c(2,2)) %>%
layer_dropout(rate = FLAGS$dropout1) %>%
layer_conv_2d(filters = FLAGS$convol_filters2, kernel_size = c(3,3)
, activation = "relu", padding = "same"
) %>%
layer_conv_2d(filters = FLAGS$convol_filters2, kernel_size = c(3,3)
, activation = "relu"
#, padding = "same"
) %>%
layer_max_pooling_2d(pool_size = c(2,2)) %>%
layer_dropout(rate = FLAGS$dropout1) %>%
layer_conv_2d(filters = FLAGS$convol_filters3, kernel_size = c(3,3)
, activation = "relu", padding = "same"
) %>%
layer_conv_2d(filters = FLAGS$convol_filters3, kernel_size = c(3,3)
, activation = "relu"
#, padding = "same"
) %>%
layer_max_pooling_2d(pool_size = c(2,2)) %>%
layer_flatten() %>%
layer_dense(units = FLAGS$dense_units1) %>%
layer_activation_relu() %>%
layer_dropout(rate = FLAGS$dropout1) %>%
layer_dense(units = FLAGS$dense_units1) %>%
layer_activation_relu() %>%
layer_dropout(rate = FLAGS$dropout1) %>%
layer_dense(units = FLAGS$dense_units1) %>%
layer_activation_relu() %>%
layer_dense(units = 2, activation = 'softmax')
model %>% compile(loss = 'binary_crossentropy', optimizer = "adadelta", metrics = c("accuracy"))
model %>% fit(data.training, data.trainLabels, epochs = FLAGS$epochs, view_metrics = FALSE,
validation_split = 0.2, shuffle = TRUE)
