Based off of: https://github.com/stwunsch/tmva_mnist/blob/master/train.py and modified for use with a resnet model. (See https://gist.github.com/ashlaban/6be0ddc58d940d6ab1783ac0dbab19cc)

train.py

#
# NOTE: Requires that you have run `python create_dataset.py`
#       from https://github.com/stwunsch/tmva_mnist first.
#

import ROOT

from resnet import ResnetBuilder

# Setup TMVA
ROOT.TMVA.Tools.Instance()
ROOT.TMVA.PyMethodBase.PyInitialize()

output = ROOT.TFile.Open("TMVA.root", "RECREATE")
factory = ROOT.TMVA.Factory(
    "MNIST", output,
    "!V:!Silent:Color:DrawProgressBar:AnalysisType=multiclass")

# Load data
dataloader = ROOT.TMVA.DataLoader("dataset")

data = ROOT.TFile.Open("mnist.root")
tree_digit = []
for i in range(10):
    tree_digit.append(data.Get("train_digits/train_digit{}".format(i)))
    dataloader.AddTree(tree_digit[i], "digit{}".format(i))

for i in range(28 * 28 * 1):
    dataloader.AddVariable("x[{}]".format(i), "x_{}".format(i), "")

dataloader.PrepareTrainingAndTestTree(
    ROOT.TCut(""), "SplitMode=Random:NormMode=None:!V")

# Define model
model = ResnetBuilder.build_resnet_18((1, 28, 28), 10)

model.compile(
    loss="categorical_crossentropy", optimizer="adam", metrics=["accuracy"])

# Store model to file
model.save("model.h5")
model.summary()

# Book methods
factory.BookMethod(dataloader, ROOT.TMVA.Types.kPyKeras, "PyKeras",
                   "H:!V:FilenameModel=model.h5:NumEpochs=10:BatchSize=100")

# Run training, test and evaluation
factory.TrainAllMethods()
factory.TestAllMethods()
factory.EvaluateAllMethods()