def test_overfitting(self):
gen = TrainImageGenerator("../../datasets/micro/annotations.csv",
"../../datasets/micro",
batch_size=1,
augumenter=NoAgumenter())
xs = tf.placeholder(tf.float32, (None, 300, 300, 3))
ys = [
tf.placeholder(
tf.float32,
(None, fm.width, fm.height, len(fm.aspect_ratios), 5))
for fm in FEATURE_MAPS
]
net = ssd(xs)
loss = ssd_total_loss(ys, net)
minimizer = tf.train.AdamOptimizer(1e-4).minimize(loss)
inputs, expected_list = gen.generate_batch([0])
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
writer = tf.summary.FileWriter('../../graphs', sess.graph)
for i in range(500):
calc_loss, _ = sess.run([loss, minimizer], {
xs: inputs,
**dict(zip(ys, expected_list))
})
print(f"{i}: {calc_loss}")
predictions = sess.run(net, {xs: inputs})
visualize_prediction(get_sample_from_batchx(inputs, 0),
get_sample_from_batchy(predictions, 0))
def test_finds_bounding_boxes(self):
trainer = TrainImageGenerator(
annotation_path="../../datasets/micro/annotations.csv",
images_path="../../datasets/micro")
x, y = trainer.generate_sample(0)
found_boxes = interpret_label(y, FEATURE_MAPS)
self.assertTrue(len(found_boxes) > 0)
def test_single_sample_generation(self):
trainer = TrainImageGenerator(
annotation_path="../../datasets/mini/annotations.csv",
images_path="../../datasets/mini")
x, y = trainer.generate_sample(0)
self.assertEqual(np.shape(x), (300, 300, 3))
self.assertEqual(len(y), len(FEATURE_MAPS))
for layer, fm in zip(y, FEATURE_MAPS):
self.assertEqual(fm.width, layer.shape[0])
self.assertEqual(fm.height, layer.shape[1])
self.assertEqual(len(fm.aspect_ratios), layer.shape[2])
self.assertEqual(layer.shape[3], 5)
def test_initalization(self):
trainer = TrainImageGenerator(
annotation_path="../../datasets/mini/annotations.csv",
images_path="../../datasets/mini")
self.assertEqual(trainer.batch_size, 8)
self.assertEqual(trainer.num_samples, 32)
self.assertEqual(trainer.num_batches, 32 / 8)
def fit(self):
gen = TrainImageGenerator(self.annotations_path, self.images_path,
batch_size=self.batch_size, augumenter=NoAgumenter())
print(f"Found {gen.num_samples} samples!")
xs = tf.placeholder(tf.float32, (None, 300, 300, 3))
ys = [tf.placeholder(tf.float32, (None, fm.width, fm.height, len(fm.aspect_ratios), 5)) for fm in FEATURE_MAPS]
net = ssd(xs)
loss = ssd_total_loss(ys, net)
minimizer = tf.train.AdamOptimizer(self.alpha).minimize(loss)
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
training_start_time = datetime.now()
last_epoch_save_time = datetime.now()
for i in range(self.max_epoch):
for batch_x, batch_y in gen.get_batches_in_epoch():
calc_loss, _ = sess.run([loss, minimizer], {xs: batch_x, **dict(zip(ys, batch_y))})
print(f"{i}: {calc_loss}")
if self.save_every_epoch:
saver.save(sess, self.output_model_path + f"-epoch{i}")
now = datetime.now()
# Periodic saving
if (self.save_after_minutes > 0) and (now > last_epoch_save_time + timedelta(minutes=self.save_after_minutes)):
saver.save(sess, self.output_model_path + f"-epoch{i}")
last_epoch_save_time = datetime.now()
# Ending training after time
if (self.kill_after_minutes > 0) and (now > training_start_time + timedelta(minutes=self.kill_after_minutes)):
break
saver.save(sess, self.output_model_path + "-final")
print(f"Model trained sucessfully. Latest version saved as {self.output_model_path}-final")
def test_get_batches_in_epoch(self):
trainer = TrainImageGenerator(
annotation_path="../../datasets/mini/annotations.csv",
images_path="../../datasets/mini")
counter = 0
for x_batch, y_batch in trainer.get_batches_in_epoch():
counter = counter + 1
self.assertEqual((8, 300, 300, 3), np.shape(x_batch))
self.assertEqual(len(y_batch), len(FEATURE_MAPS))
for layer, fm in zip(y_batch, FEATURE_MAPS):
self.assertEqual(8, layer.shape[0])
self.assertEqual(fm.width, layer.shape[1])
self.assertEqual(fm.height, layer.shape[2])
self.assertEqual(len(fm.aspect_ratios), layer.shape[3])
self.assertEqual(layer.shape[4], 5)
self.assertEqual(counter, 4)
from postprocessing.visualization import visualize_prediction
from dataset_generation.data_feeder import TrainImageGenerator
from dataset_generation.augmenter import NoAgumenter
if __name__ == "__main__":
generator = TrainImageGenerator("../../datasets/micro/annotations.csv",
"../../datasets/micro",
batch_size=1,
augumenter=NoAgumenter())
img, label = generator.generate_sample(0)
visualize_prediction(img, label)