Folks,
I have some problem when try resuming h2o deep learning in R from a checkpointed model with validation frame provided. It says "Validation dataset must be the same as for the check pointed model", which I believe I do have the same validation datasets. If I leave validation_frame blank, checkpointing model works fine. I attach my code below:
localh2o <- h2o.init(nthreads = -1)
train_image.hex <- read.csv("mnist_train.csv",header=FALSE)
train_image.hex[,785] <- factor(train_image.hex[,785])
train_image.hex <- as.h2o(train_image.hex)
test_image.hex <- read.csv("mnist_test.csv",header=FALSE)
test_image.hex[,785] <- factor(test_image.hex[,785])
test_image.hex <- as.h2o(test_image.hex)
mnist_model <- h2o.deeplearning(x=1:784, y = 785,
training_frame= train_image.hex,
validation_frame = test_image.hex,
activation = "RectifierWithDropout", hidden = c(500,1000),
input_dropout_ratio = 0.2,
hidden_dropout_ratios = c(0.5,0.5), adaptive_rate=TRUE,
rho=0.98, epsilon = 1e-7,
l1 = 1e-8, l2 = 1e-7, max_w2 = 10,
epochs = 10, export_weights_and_biases = TRUE,
variable_importances = FALSE
)
h2o.saveModel(mnist_model, path="/tmp",force=TRUE)
Then I shut down the h2o, quit R and restart h2o in R to resume training, where h2o errors out:
localh2o <- h2o.init(nthreads = -1)
train_image.hex <- read.csv("mnist_train.csv",header=FALSE)
train_image.hex[,785] <- factor(train_image.hex[,785])
train_image.hex <- as.h2o(train_image.hex)
test_image.hex <- read.csv("mnist_test.csv",header=FALSE)
test_image.hex[,785] <- factor(test_image.hex[,785])
test_image.hex <- as.h2o(test_image.hex)
startmodel <- h2o.loadModel("/tmp/DeepLearning_model_R_1443812402059_20", localh2o)
mnist_model <- h2o.deeplearning(x=1:784, y = 785,
checkpoint = startmodel@model_id,
training_frame= train_image.hex,
validation_frame = test_image.hex,
activation = "RectifierWithDropout", hidden = c(500,1000),
input_dropout_ratio = 0.2,
hidden_dropout_ratios = c(0.5,0.5), adaptive_rate=TRUE,
rho=0.98, epsilon = 1e-7,
l1 = 1e-8, l2 = 1e-7, max_w2 = 10,
epochs = 10, export_weights_and_biases = TRUE,
variable_importances = FALSE
)
