Merge pull request #50 from aced-differentiate/example_cleanup

tidying up examples

Project.toml

@@ -19,7 +19,7 @@ Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 
 [compat]
 CSV = "0.7, 0.8"
-ChemistryFeaturization = "0.2"
+ChemistryFeaturization = "0.2.2"
 DataFrames = "0.21, 0.22"
 Flux = "0.11"
 LightGraphs = "1.3"

examples/1_formation_energy/formation_energy.jl

@@ -1,8 +1,6 @@
 #=
  Train a simple network to predict formation energy per atom (downloaded from Materials Project).
 =#
-#using Pkg
-#Pkg.activate("../../")
 using CSV, DataFrames
 using Random, Statistics
 using Flux
@@ -14,7 +12,7 @@ using AtomicGraphNets
 println("Setting things up...")
 
 # data-related options
-num_pts = 100 # how many points to use? Up to 32530 in the formation energy case as of 2020/04/01
+num_pts = 100 # how many points to use?
 train_frac = 0.8 # what fraction for training?
 num_epochs = 5 # how many epochs to train?
 num_train = Int32(round(train_frac * num_pts))
@@ -48,13 +46,8 @@ output = y[indices]
 
 # next, make graphs and build input features (matrices of dimension (# features, # nodes))
 println("Building graphs and feature vectors from structures...")
-#graphs = SimpleWeightedGraph{Int32, Float32}[]
-#element_lists = Array{String}[]
-#inputs = Tuple{Array{Float32,2},SparseArrays.SparseMatrixCSC{Float32,Int64}}[]
 inputs = AtomGraph[]
 
-#TODO: this with bulk processing fcn
-
 for r in eachrow(info)
     cifpath = string(datadir, prop, "_cifs/", r[Symbol(id)], ".cif")
     gr = build_graph(cifpath)
@@ -71,21 +64,15 @@ train_input = inputs[1:num_train]
 test_input = inputs[num_train+1:end]
 train_data = zip(train_input, train_output)
 
-# build the network (basically just copied from CGCNN.py for now): the convolutional layers, a mean pooling function, some dense layers, then fully connected output to one value for prediction
-
+# build the model
 println("Building the network...")
-#model = Chain([AGNConv(num_features=>num_features) for i in 1:num_conv]..., AGNMeanPool(crys_fea_len, 0.1), [Dense(crys_fea_len, crys_fea_len, softplus) for i in 1:num_hidden_layers]..., Dense(crys_fea_len, 1))
 model = Xie_model(num_features, num_conv=num_conv, atom_conv_feature_length=crys_fea_len, num_hidden_layers=1)
 
-# MaxPool might make more sense?
-
-# define loss function
+# define loss function and a callback to monitor progress
 loss(x,y) = Flux.mse(model(x), y)
-# and a callback to see training progress
 evalcb() = @show(mean(loss.(test_input, test_output)))
 evalcb()
 
 # train
 println("Training!")
-#Flux.train!(loss, params(model), train_data, opt)
 @epochs num_epochs Flux.train!(loss, params(model), train_data, opt, cb = Flux.throttle(evalcb, 5))

examples/2_qm9/README.md

@@ -1,3 +1,5 @@
 # Example 2: QM9
 
-In this example we will use the same network architecture but use data from the QM9 dataset. Note that the .xyz files provided within the QM9 dataset are not parsable directly by ASE, you need to remove the last couple lines, which is easy enough to script yourself, but I've included a small set of them here for demonstration purposes.
+In this example we will use the same network architecture but use data from the QM9 dataset. Note that the .xyz files provided within the QM9 dataset are not parsable directly by ASE, you need to remove the last couple lines, which is easy enough to script yourself, but I've included a small set of them here for demonstration purposes.
+
+NB: the actual model performance on QM9 is not that great because we're currently not encoding a variety of important features for organic molecules. This is provided mainly to show the processing of a different dataset and demonstrate batch processing capabilities.