library(reticulate)
reticulate::use_condaenv("mlr3keras", required=TRUE)
library(data.table)
devtools::load_all()
# Download data from [fastai/imagenette](https://github.com/fastai/imagenette) and adjust path
dir = "~/Downloads/imagenette2-160/train"
dt = imagepathdf_from_imagenet_dir(dir)
dt$n = runif(nrow(dt))
dt$m = runif(nrow(dt))
dt[, lapply(.SD, as.character)]
g = image_data_generator(validation_split=0.5)
dtgen = make_generator_from_dataframe(dt, c("m","n") ,"class" , g)
imgen = keras::flow_images_from_dataframe(dt, x_col="image", y_col = "class", generator = g, batch_size = as.integer(32), shuffle=TRUE, seed=1L)
cogen = combine_generators(imgen, dtgen)
# Now try the fitting:
base = application_mobilenet(
include_top = FALSE,
weights = "imagenet",
input_tensor = NULL,
input_shape = NULL
)
numeric_input = layer_input(shape=c(2), name="numerics")
out_inter = base$output %>%
layer_global_average_pooling_2d() %>%
layer_dense(
units = 256L,
activation = "relu"
)
model = keras_model(
inputs = list(base$input, numeric_input),
out = layer_concatenate(c(out_inter, numeric_input)) %>%
layer_dense(
units = 10L,
activation = "softmax"
)
)
freeze_weights(base)
unfreeze_weights(base, from = 77)
model$compile(
optimizer = optimizer_rmsprop(),
loss = "categorical_crossentropy"
)
fit_generator(model, generator=cogen, steps_per_epoch=100L)
reticulate::use_condaenv("mlr3keras", required=TRUE)
devtools::load_all()
dt = readRDS("~/Documents/repos/airbnb/munich_clean.RDS")
# For now we build our model only from four features:
x = c("host_is_superhost", "longitude", "latitude", "bedrooms", "price")
dt$bedrooms = as.integer(dt$bedrooms) - 1L
dt$host_is_superhost = as.integer(dt$host_is_superhost)
dt$image = paste0(dt$id, ".jpg")
dt$review_scores_rating = dt$review_scores_rating / 100
# Build the generators
g = image_data_generator(validation_split=0.5)
# All following generators need equal seeds and equal batch_size.
dtgen = make_generator_from_dataframe(data.table(dt), x, "review_scores_rating" , g, shuffle=TRUE, seed=1L, subset="training")
imgen = keras::flow_images_from_dataframe(dt, x_col="image", y_col = "review_scores_rating",
generator = g, batch_size = as.integer(32),
shuffle=TRUE, seed=1L, class_mode="raw", drop_duplicates=NULL, subset="training",
directory="~/Documents/repos/airbnb/data/pictures/32")
cogen = combine_generators(imgen, dtgen)
dtgen2 = make_generator_from_dataframe(data.table(dt), x, "review_scores_rating" , g, shuffle=TRUE, seed=1L, subset="validation")
imgen2 = keras::flow_images_from_dataframe(dt, x_col="image", y_col = "review_scores_rating",
generator = g, batch_size = as.integer(32),
shuffle=TRUE, seed=1L, class_mode="raw", drop_duplicates=NULL, subset="validation",
directory="~/Documents/repos/airbnb/data/pictures/32")
cogen2 = combine_generators(imgen2, dtgen2)
# Build the model
base = application_mobilenet(
include_top = FALSE,
weights = "imagenet",
input_tensor = NULL,
input_shape = NULL
)
numeric_input = layer_input(shape=length(x), name="numerics")
out_inter = base$output %>%
layer_global_average_pooling_2d() %>%
layer_dense(
units = 1024L,
activation = "relu"
) %>%
layer_dense(
units = 1024L,
activation = "relu"
)
model = keras_model(
inputs = list(base$input, numeric_input),
out = layer_concatenate(c(out_inter, numeric_input)) %>%
layer_dense(
units = 1L,
activation = "linear"
)
)
freeze_weights(base)
unfreeze_weights(base, from = 80)
model$compile(
optimizer = optimizer_adam(),
loss = "mean_squared_error"
)
fit_generator(model, generator=cogen, steps_per_epoch=50L, epochs = 10L, validation_data=cogen2, validation_steps=50L)
fit_generator(model, generator=cogen, steps_per_epoch=100L)
fit_generator(model, generator=cogen, steps_per_epoch=100L)
