scripts/nn/examples/Example-EfficientNet.dml

#-------------------------------------------------------------
#
# Licensed to the Apache Software Foundation (ASF) under one
# or more contributor license agreements.  See the NOTICE file
# distributed with this work for additional information
# regarding copyright ownership.  The ASF licenses this file
# to you under the Apache License, Version 2.0 (the
# "License"); you may not use this file except in compliance
# with the License.  You may obtain a copy of the License at
#
#   http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing,
# software distributed under the License is distributed on an
# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
# KIND, either express or implied.  See the License for the
# specific language governing permissions and limitations
# under the License.
#
#-------------------------------------------------------------


/*
* The MNIST Data can be downloaded as follows:
* mkdir -p data/mnist/
* cd data/mnist/
* curl -O https://pjreddie.com/media/files/mnist_train.csv
* curl -O https://pjreddie.com/media/files/mnist_test.csv
*/

# TODO add tests in functions/builtin, applications/nn

# This script trains a minified version of the EfficientNet-B0 model
# with a single MBConv layer. This model heavily overfits on a simple
# MNist dataset since it was originally developed on the ImageNet dataset
# Thus layer outputs and other factors are too large for normal MNist.
# Therefore we only train once on the Mnist Train ds and print out its Accuracy
# Import required methods
source("nn/examples/efficientNet.dml") as eff

# Read training data
data = read("data/mnist/mnist_test.csv", format="csv")
N = nrow(data)

# Extract images and labels
images = data[,2:ncol(data)]
labels = data[,1]

# Scale images to [0,1], and one-hot encode the labels
images = images / 255.0
labels = table(seq(1, N), labels+1, N, 10)

model = eff::initNetwork(1, 10, -1)

# Train
epochs = 1
batch_size = 256
model = eff::netTrain(model, images, 1, 28, 28, labels, epochs, batch_size, 0.025, 0.9, TRUE)


# Also Predict in Batches since otherwise we can run into Memory Issues
# Could be unnecessary on more powerful machines :)
iters = ceil(N / batch_size)
accuracy = 0.0
for(i in 1:iters) {
  beg = ((i-1) * batch_size) %% N + 1
  end = min(N, beg + batch_size - 1)
  X_batch = images[beg:end,]
  y_batch = labels[beg:end,]

  pred = eff::netPredict(X_batch, model, 1, 28, 28)
  partial_acc = mean(rowIndexMax(pred) == rowIndexMax(y_batch))
  accuracy = accuracy + partial_acc
}

print("Total Accuracy: " + (accuracy / iters))