public void PrepareData()
{
// full dataset https://github.com/le-scientifique/torchDatasets/raw/master/dbpedia_csv.tar.gz
var url = "https://raw.githubusercontent.com/SciSharp/TensorFlow.NET/master/data/dbpedia_subset.zip";
Web.Download(url, dataDir, "dbpedia_subset.zip");
Compress.UnZip(Path.Combine(dataDir, "dbpedia_subset.zip"), Path.Combine(dataDir, "dbpedia_csv"));
Console.WriteLine("Building dataset...");
var(x, y) = (new int[0][], new int[0]);
if (ModelName == "char_cnn")
{
(x, y, alphabet_size) = DataHelpers.build_char_dataset(TRAIN_PATH, "char_cnn", CHAR_MAX_LEN);
}
else
{
var word_dict = DataHelpers.build_word_dict(TRAIN_PATH);
vocabulary_size = len(word_dict);
(x, y) = DataHelpers.build_word_dataset(TRAIN_PATH, word_dict, WORD_MAX_LEN);
}
Console.WriteLine("\tDONE ");
var(train_x, valid_x, train_y, valid_y) = train_test_split(x, y, test_size: 0.15f);
Console.WriteLine("Training set size: " + train_x.shape[0]);
Console.WriteLine("Test set size: " + valid_x.shape[0]);
}
protected virtual bool RunWithBuiltGraph(Session session, Graph graph)
{
Console.WriteLine("Building dataset...");
var(x, y, alphabet_size) = DataHelpers.build_char_dataset("train", model_name, CHAR_MAX_LEN, DataLimit);
var(train_x, valid_x, train_y, valid_y) = train_test_split(x, y, test_size: 0.15f);
ITextClassificationModel model = null;
switch (model_name) // word_cnn | char_cnn | vd_cnn | word_rnn | att_rnn | rcnn
{
case "word_cnn":
case "char_cnn":
case "word_rnn":
case "att_rnn":
case "rcnn":
throw new NotImplementedException();
break;
case "vd_cnn":
model = new VdCnn(alphabet_size, CHAR_MAX_LEN, NUM_CLASS);
break;
}
// todo train the model
return(false);
}
protected virtual bool RunWithImportedGraph(Session sess, Graph graph)
{
var stopwatch = Stopwatch.StartNew();
Console.WriteLine("Building dataset...");
var path = UseSubset ? SUBSET_PATH : TRAIN_PATH;
int[][] x = null;
int[] y = null;
int alphabet_size = 0;
int vocabulary_size = 0;
if (model_name == "vd_cnn")
{
(x, y, alphabet_size) = DataHelpers.build_char_dataset(path, model_name, CHAR_MAX_LEN, DataLimit = null, shuffle: !UseSubset);
}
else
{
var word_dict = DataHelpers.build_word_dict(TRAIN_PATH);
vocabulary_size = len(word_dict);
(x, y) = DataHelpers.build_word_dataset(TRAIN_PATH, word_dict, WORD_MAX_LEN);
}
Console.WriteLine("\tDONE ");
var(train_x, valid_x, train_y, valid_y) = train_test_split(x, y, test_size: 0.15f);
Console.WriteLine("Training set size: " + train_x.len);
Console.WriteLine("Test set size: " + valid_x.len);
Console.WriteLine("Import graph...");
var meta_file = model_name + ".meta";
tf.train.import_meta_graph(Path.Join("graph", meta_file));
Console.WriteLine("\tDONE " + stopwatch.Elapsed);
sess.run(tf.global_variables_initializer());
var saver = tf.train.Saver(tf.global_variables());
var train_batches = batch_iter(train_x, train_y, BATCH_SIZE, NUM_EPOCHS);
var num_batches_per_epoch = (len(train_x) - 1) / BATCH_SIZE + 1;
double max_accuracy = 0;
Tensor is_training = graph.OperationByName("is_training");
Tensor model_x = graph.OperationByName("x");
Tensor model_y = graph.OperationByName("y");
Tensor loss = graph.OperationByName("loss/Mean"); // word_cnn
Operation optimizer = graph.OperationByName("loss/Adam"); // word_cnn
Tensor global_step = graph.OperationByName("Variable");
Tensor accuracy = graph.OperationByName("accuracy/accuracy");
stopwatch = Stopwatch.StartNew();
int i = 0;
foreach (var(x_batch, y_batch, total) in train_batches)
{
i++;
var train_feed_dict = new FeedDict
{
[model_x] = x_batch,
[model_y] = y_batch,
[is_training] = true,
};
//Console.WriteLine("x: " + x_batch.ToString() + "\n");
//Console.WriteLine("y: " + y_batch.ToString());
// original python:
//_, step, loss = sess.run([model.optimizer, model.global_step, model.loss], feed_dict = train_feed_dict)
var result = sess.run(new ITensorOrOperation[] { optimizer, global_step, loss }, train_feed_dict);
loss_value = result[2];
var step = (int)result[1];
if (step % 10 == 0)
{
var estimate = TimeSpan.FromSeconds((stopwatch.Elapsed.TotalSeconds / i) * total);
Console.WriteLine($"Training on batch {i}/{total} loss: {loss_value}. Estimated training time: {estimate}");
}
if (step % 100 == 0)
{
// # Test accuracy with validation data for each epoch.
var valid_batches = batch_iter(valid_x, valid_y, BATCH_SIZE, 1);
var(sum_accuracy, cnt) = (0.0f, 0);
foreach (var(valid_x_batch, valid_y_batch, total_validation_batches) in valid_batches)
{
var valid_feed_dict = new FeedDict
{
[model_x] = valid_x_batch,
[model_y] = valid_y_batch,
[is_training] = false
};
var result1 = sess.run(accuracy, valid_feed_dict);
float accuracy_value = result1;
sum_accuracy += accuracy_value;
cnt += 1;
}
var valid_accuracy = sum_accuracy / cnt;
print($"\nValidation Accuracy = {valid_accuracy}\n");
// # Save model
if (valid_accuracy > max_accuracy)
{
max_accuracy = valid_accuracy;
// saver.save(sess, $"{dataDir}/{model_name}.ckpt", global_step: step.ToString());
print("Model is saved.\n");
}
}
}
return(false);
}
