import sys
# add the code path
sys.path.append('..')

%load_ext autoreload
%autoreload 2

Example to train Yolo for object detection

from ultrayolo import YoloV3, callbacks
from ultrayolo import datasets
from pathlib import Path
import tensorflow as tf
from ultrayolo import losses
import matplotlib.pyplot as plt

Define the parameters for the run

image_shape = (256,256,3)
batch_shape = 2
max_objects = 100
train_dataset_path = '../tests/data/manifest.txt'
anchors = datasets.load_anchors('../tests/data/yolov3_anchors.txt')
classes = datasets.load_classes('../tests/data/classes.txt')

Create the model

model = YoloV3(image_shape, max_objects,
               anchors=anchors, num_classes=len(classes),
               training=True, backbone='MobileNetV2', base_grid_size=64)

num pooling 1

tf.keras.utils.plot_model(model.model, show_shapes=True)
../_images/4_train_example_8_0.png

Create the dataset

train_dataset = datasets.YoloDatasetMultiFile(
    train_dataset_path, image_shape, max_objects, batch_shape,
    model.anchors, model.masks, 64
)

print('num batches', len(train_dataset))

num batches 2

Make optimizer and loss

optimizer = model.get_optimizer('adam', 1e-4)
model_loss = model.get_loss_function(num_batches = len(train_dataset))

9596 MainThread using adam optimize

model_loss

[yolo_loss_large at 0x165978110,
 yolo_loss_medium at 0x165911550,
 yolo_loss_small at 0x1659a82d0]

compile the model

model.compile(optimizer, model_loss, run_eagerly=True, summary=False)

Create the callbacks

model_callbacks = callbacks.default_callbacks(model,
    run_path='./checkpoints', lrate_mode='exp_range',
    lrate_value=1e-5, verbose=0)

Set the model in transfer mode

model.set_mode_transfer()

9895 MainThread freeze backbone

history = model.fit(train_dataset, train_dataset, 5, callbacks=model_callbacks)

9974 MainThread training for 5 epochs on the dataset /Users/fumarolaf/git/ultrayolo/notebooks/../tests/data

Train for 2 steps, validate for 2 steps
Epoch 1/5
2/2 [==============================] - 8s 4s/step - loss: 1606.3602 - yolo_output_0_loss: 83.3973 - yolo_output_1_loss: 307.9103 - yolo_output_2_loss: 1130.5004 - val_loss: 1491.7465 - val_yolo_output_0_loss: 76.3384 - val_yolo_output_1_loss: 266.1987 - val_yolo_output_2_loss: 1064.6580
Epoch 2/5
2/2 [==============================] - 5s 3s/step - loss: 1580.7247 - yolo_output_0_loss: 77.1867 - yolo_output_1_loss: 300.3980 - yolo_output_2_loss: 1118.5891 - val_loss: 1491.7280 - val_yolo_output_0_loss: 76.3367 - val_yolo_output_1_loss: 266.2234 - val_yolo_output_2_loss: 1064.6182
Epoch 3/5
2/2 [==============================] - 5s 2s/step - loss: 1557.1224 - yolo_output_0_loss: 72.4101 - yolo_output_1_loss: 294.3926 - yolo_output_2_loss: 1105.7705 - val_loss: 1491.6611 - val_yolo_output_0_loss: 76.3197 - val_yolo_output_1_loss: 266.2427 - val_yolo_output_2_loss: 1064.5510
Epoch 4/5
2/2 [==============================] - 5s 2s/step - loss: 1537.5295 - yolo_output_0_loss: 69.5900 - yolo_output_1_loss: 289.4261 - yolo_output_2_loss: 1093.9666 - val_loss: 1491.5289 - val_yolo_output_0_loss: 76.2830 - val_yolo_output_1_loss: 266.2503 - val_yolo_output_2_loss: 1064.4507
Epoch 5/5
2/2 [==============================] - 5s 2s/step - loss: 1517.3444 - yolo_output_0_loss: 65.6196 - yolo_output_1_loss: 285.4909 - yolo_output_2_loss: 1081.6898 - val_loss: 1491.3119 - val_yolo_output_0_loss: 76.2231 - val_yolo_output_1_loss: 266.2384 - val_yolo_output_2_loss: 1064.3088

Evaluate model Loss

loss = history.history['loss']
val_loss = history.history['val_loss']

epochs = range(1, len(loss) + 1)

# "bo" is for "blue dot"
plt.plot(epochs, loss, 'bo', label='Training loss')
# b is for "solid blue line"
plt.plot(epochs, val_loss, 'b', label='Validation loss')
plt.title('Training and validation loss')
plt.xlabel('Epochs')
plt.ylabel('Loss')
plt.legend()

plt.show()
../_images/4_train_example_23_0.png

Yolo loss for large-sized objects

loss = history.history['yolo_output_0_loss']
val_loss = history.history['val_yolo_output_0_loss']

epochs = range(1, len(loss) + 1)

# "bo" is for "blue dot"
plt.plot(epochs, loss, 'bo', label='Training loss large size object')
# b is for "solid blue line"
plt.plot(epochs, val_loss, 'b', label='Validation loss large size object')
plt.title('Training and validation loss')
plt.xlabel('Epochs')
plt.ylabel('Loss')
plt.legend()

plt.show()
../_images/4_train_example_25_0.png

Yolo loss for medium-sized objects

loss = history.history['yolo_output_1_loss']
val_loss = history.history['val_yolo_output_1_loss']

epochs = range(1, len(loss) + 1)

# "bo" is for "blue dot"
plt.plot(epochs, loss, 'bo', label='Training loss medium size object')
# b is for "solid blue line"
plt.plot(epochs, val_loss, 'b', label='Validation loss medium size object')
plt.title('Training and validation loss')
plt.xlabel('Epochs')
plt.ylabel('Loss')
plt.legend()

plt.show()
../_images/4_train_example_27_0.png

Yolo loss for small-sized objects

loss = history.history['yolo_output_2_loss']
val_loss = history.history['val_yolo_output_2_loss']

epochs = range(1, len(loss) + 1)

# "bo" is for "blue dot"
plt.plot(epochs, loss, 'bo', label='Training loss small size object')
# b is for "solid blue line"
plt.plot(epochs, val_loss, 'b', label='Validation loss small size object')
plt.title('Training and validation loss')
plt.xlabel('Epochs')
plt.ylabel('Loss')
plt.legend()

plt.show()
../_images/4_train_example_29_0.png

model.save(‘./save_model/model.h5’)