def train(self, train_dataset, val_dataset, learning_rate, epochs, layers,
augmentation=None, custom_callbacks=None, no_augmentation_sources=None):
assert self.mode == "training", "Create model in training mode."
if layers in LAYERS_REGEX.keys():
layers = LAYERS_REGEX[layers]
# Data generators
train_generator = data_generator(
train_dataset,
self.config,
shuffle = True,
augmentation = augmentation,
batch_size = self.config.BATCH_SIZE,
no_augmentation_sources = no_augmentation_sources
);
val_generator = data_generator(
val_dataset,
self.config,
shuffle = True,
batch_size = self.config.BATCH_SIZE
);
## Create log_dir if it does not exist:
if not os.path.exists(self.log_dir):
os.makedirs(self.log_dir);
## Implement custom callbacks:
callbacks = [
keras.callbacks.TensorBoard(log_dir=self.log_dir, histogram_freq=0, write_graph=True, write_images=False),
keras.callbacks.ModelCheckpoint(self.checkpoint_path, verbose=1, save_weights_only=True),
];
if custom_callbacks:
callbacks += custom_callbacks;
## Train dense model:
log("\nStarting at epoch {}. LR={}\n".format(self.epoch, learning_rate));
log("Checkpoint Path: {}".format(self.checkpoint_path));
self.set_trainable(layers);
self.compile(learning_rate, self.config.LEARNING_MOMENTUM);
workers = 0 if os.name is 'nt' else multiprocessing.cpu_count();
print("Number of workers to use: %d" % workers)
self.keras_model.fit_generator(
train_generator,
initial_epoch=self.epoch,
epochs=epochs,
steps_per_epoch=self.config.STEPS_PER_EPOCH,
callbacks=callbacks,
validation_data=val_generator,
validation_steps=self.config.VALIDATION_STEPS,
max_queue_size = 100,
workers = 1,
use_multiprocessing = False,
);
self.epoch = max(self.epoch, epochs);
def train(self, train_dataset, val_dataset, learning_rate, epochs, layers,
augmentation=None, custom_callbacks=None, no_augmentation_sources=None):
# Pre-defined layer regular expressions
layer_regex = {
"all": "s_.*",
}
if layers in layer_regex.keys():
layers = layer_regex[layers]
# Data generators
train_generator = modellib.data_generator(train_dataset, self.config, shuffle=True,
augmentation=augmentation,
batch_size=self.config.BATCH_SIZE,
no_augmentation_sources=no_augmentation_sources)
val_generator = modellib.data_generator(val_dataset, self.config, shuffle=True,
batch_size=self.config.BATCH_SIZE)
# Create log_dir if it does not exist
if not os.path.exists(self.log_dir):
os.makedirs(self.log_dir)
# Callbacks
callbacks = [
keras.callbacks.TensorBoard(log_dir=self.log_dir,
histogram_freq=0, write_graph=True, write_images=False),
keras.callbacks.ModelCheckpoint(self.checkpoint_path,
verbose=0, save_weights_only=True),
]
# Add custom callbacks to the list
if custom_callbacks:
callbacks += custom_callbacks
# Train
modellib.log("\nStarting at epoch {}. LR={}\n".format(self.epoch, learning_rate))
modellib.log("Checkpoint Path: {}".format(self.checkpoint_path))
self.set_trainable(layers)
self.compile(learning_rate, self.config.LEARNING_MOMENTUM)
self.keras_model.fit_generator(
train_generator,
initial_epoch=self.epoch,
epochs=epochs,
steps_per_epoch=self.config.STEPS_PER_EPOCH,
callbacks=callbacks,
validation_data=val_generator,
validation_steps=self.config.VALIDATION_STEPS,
max_queue_size=self.config.BATCH_SIZE * 3,
workers=int(multiprocessing.cpu_count() / 8),
use_multiprocessing=True,
)
self.epoch = max(self.epoch, epochs)
def create_data_generator(num_random_rois=2000, detection_targets=True):
# Create data generator
g = modellib.data_generator(dataset,
config,
shuffle=True,
num_random_rois=num_random_rois,
batch_size=4,
detection_targets=detection_targets)
return g, num_random_rois, detection_targets
def mrcnn_generator(self,
shuffle=True,
augment=False,
augmentation=None,
random_rois=0,
batch_size=1,
detection_targets=False,
no_augmentation_sources=None):
from mrcnn.model import data_generator
return data_generator(self, self.config, shuffle, augment,
augmentation, random_rois, batch_size,
detection_targets, no_augmentation_sources)
def test_data_sequence():
config = ShapesConfig()
dataset = FixShapesDataset()
dataset.load_shapes(1, config.IMAGE_SHAPE[0], config.IMAGE_SHAPE[1])
dataset.prepare()
dg = modellib.data_generator(dataset, config)
inputs, output = next(dg)
assert output == []
batch_images, batch_image_meta, batch_rpn_match, batch_rpn_bbox, batch_gt_class_ids, batch_gt_boxes, batch_gt_masks = inputs
assert (batch_gt_boxes[0][:3] == [[44, 78, 87, 121], [0, 28, 56, 87],
[0, 0, 0, 0]]).all()
assert (batch_image_meta[0] == [
0, 128, 128, 3, 128, 128, 3, 0, 0, 128, 128, 1, 1, 1, 1, 1
]).all()
def test_load_image(self):
dataset_train = ShapesDataset()
config = ShapesConfig()
# 500
dataset_train.load_shapes(5, config.IMAGE_SHAPE[0], config.IMAGE_SHAPE[1])
dataset_train.prepare()
ROOT_DIR = os.path.abspath("pretrainings/")
MODEL_DIR = os.path.join(ROOT_DIR, "logs")
# Create model in training mode
model = MaskRCNNAuxFeatures(mode="training", config=config,
model_dir=MODEL_DIR)
train_generator = modellib.data_generator(dataset_train, model.config, shuffle=True,
augmentation=None,
batch_size=model.config.BATCH_SIZE,
no_augmentation_sources=None)
first_image = next(train_generator)
self.assertTrue(first_image)
def check_ratio_of_positive_rois(num_random_rois):
# Check ratio of positive ROIs in a set of images.
limit = 10
temp_g = modellib.data_generator(dataset,
config,
shuffle=True,
num_random_rois=num_random_rois,
batch_size=1,
detection_targets=True)
total = 0
for i in range(limit):
_, [mrcnn_class_ids, _, _] = next(temp_g)
positive_rois = np.sum(mrcnn_class_ids[0] > 0)
total += positive_rois
print("{:5} {:5.2f}".format(positive_rois,
positive_rois / mrcnn_class_ids.shape[1]))
print("Average percent: {:.2f}".format(total /
(limit * mrcnn_class_ids.shape[1])))
center_anchor = anchors_per_cell * (
(center_cell[0] * backbone_shapes[level][1] / config.RPN_ANCHOR_STRIDE**2) \
+ center_cell[1] / config.RPN_ANCHOR_STRIDE)
level_center = int(center_anchor)
# Draw anchors. Brightness show the order in the array, dark to bright.
for i, rect in enumerate(level_anchors[level_center:level_center+anchors_per_cell]):
y1, x1, y2, x2 = rect
p = patches.Rectangle((x1, y1), x2-x1, y2-y1, linewidth=2, facecolor='none',
edgecolor=(i+1)*np.array(colors[level]) / anchors_per_cell)
ax.add_patch(p)
# Create data generator
random_rois = 2000
g = modellib.data_generator(
dataset, config, shuffle=True, random_rois=random_rois,
batch_size=4,
detection_targets=True)
# Get Next Image
if random_rois:
[normalized_images, image_meta, rpn_match, rpn_bbox, gt_class_ids, gt_boxes, gt_masks, rpn_rois, rois], \
[mrcnn_class_ids, mrcnn_bbox, mrcnn_mask] = next(g)
log("rois", rois)
log("mrcnn_class_ids", mrcnn_class_ids)
log("mrcnn_bbox", mrcnn_bbox)
log("mrcnn_mask", mrcnn_mask)
else:
[normalized_images, image_meta, rpn_match, rpn_bbox, gt_boxes, gt_masks], _ = next(g)
log("gt_class_ids", gt_class_ids)
def get_labels(self, labels):
dims = labels.shape
unlabeled_labels = np.zeros((dims[0], dims[1], 1))
building_labels = np.zeros((dims[0], dims[1], 1))
fence_labels = np.zeros((dims[0], dims[1], 1))
other_labels = np.zeros((dims[0], dims[1], 1))
pedestrian_labels = np.zeros((dims[0], dims[1], 1))
pole_labels = np.zeros((dims[0], dims[1], 1))
road_line_labels = np.zeros((dims[0], dims[1], 1))
road_labels = np.zeros((dims[0], dims[1], 1))
sidewalk_labels = np.zeros((dims[0], dims[1], 1))
vegetation_labels = np.zeros((dims[0], dims[1], 1))
car_labels = np.zeros((dims[0], dims[1], 1))
wall_labels = np.zeros((dims[0], dims[1], 1))
traffic_sign_labels = np.zeros((dims[0], dims[1], 1))
unlabeled_index = np.all(labels == (0, 0, 0), axis=-1)
building_index = np.all(labels == (70, 70, 70), axis=-1)
fence_index = np.all(labels == (190, 153, 153), axis=-1)
other_index = np.all(labels == (250, 170, 160), axis=-1)
pedestrian_index = np.all(labels == (220, 20, 60), axis=-1)
pole_index = np.all(labels == (153, 153, 153), axis=-1)
road_line_index = np.all(labels == (157, 234, 50), axis=-1)
road_index = np.all(labels == (128, 64, 128), axis=-1)
sidewalk_index = np.all(labels == (244, 35, 232), axis=-1)
vegetation_index = np.all(labels == (107, 142, 35), axis=-1)
car_index = np.all(labels == (0, 0, 142), axis=-1)
wall_index = np.all(labels == (102, 102, 156), axis=-1)
traffic_sign_index = np.all(labels == (220, 220, 70), axis=-1)
unlabeled_labels[unlabeled_index] = 1
building_labels[building_index] = 10
fence_labels[fence_index] = 10
other_labels[other_index] = 10
pedestrian_labels[pedestrian_index] = 10
pole_labels[pole_index] = 10
road_line_labels[road_line_index] = 10
road_labels[road_index] = 10
sidewalk_labels[sidewalk_index] = 10
vegetation_labels[vegetation_index] = 1
car_labels[car_index] = 10
wall_labels[wall_index] = 10
traffic_sign_labels[traffic_sign_index] = 10
return np.dstack([unlabeled_labels, building_labels, fence_labels,
return np.dstack([unlabeled_labels, building_labels, fence_labels,
other_labels, pedestrian_labels, pole_labels,
road_line_labels, road_labels, sidewalk_labels, vegetation_labels,
car_labels, wall_labels, traffic_sign_labels])
def image_reference(self, image_id):
"""Return the carla data of the image."""
info = self.image_info[image_id]
if info["source"] == "carla":
return info["id"]
else:
super(self.__class__).image_reference(self, image_id)
config = CarlaConfig()
config.STEPS_PER_EPOCH = NUMBER_OF_TRAIN_DATA//config.BATCH_SIZE
config.VALIDATION_STEPS = NUMBER_OF_VAL_DATA//config.BATCH_SIZE
config.display()
dataset = carlaDataset()
dataset.load_images(dir=RGB_TRAIN_DIR, type='train')
# mask, a = train.load_mask(50)
# print(a)
dataset.prepare()
print("Image Count: {}".format(len(dataset.image_ids)))
print("Class Count: {}".format(dataset.num_classes))
for i, info in enumerate(dataset.class_info):
print("{:3}. {:50}".format(i, info['name']))
image_ids = np.random.choice(dataset.image_ids, 4)
for image_id in image_ids:
image = dataset.load_image(image_id)
mask, class_ids = dataset.load_mask(image_id)
visualize.display_top_masks(image, mask, class_ids, dataset.class_names)
# Load random image and mask.
image_id = random.choice(dataset.image_ids)
image = dataset.load_image(image_id)
mask, class_ids = dataset.load_mask(image_id)
# Compute Bounding box
bbox = utils.extract_bboxes(mask)
# Display image and additional stats
print("image_id ", image_id)
log("image", image)
log("mask", mask)
log("class_ids", class_ids)
log("bbox", bbox)
# Display image and instances
visualize.display_instances(image, bbox, mask, class_ids, dataset.class_names)
# Load random image and mask.
image_id = np.random.choice(dataset.image_ids, 1)[0]
image = dataset.load_image(image_id)
mask, class_ids = dataset.load_mask(image_id)
original_shape = image.shape
# Resize
image, window, scale, padding, _ = utils.resize_image(
image,
min_dim=config.IMAGE_MIN_DIM,
max_dim=config.IMAGE_MAX_DIM,
mode=config.IMAGE_RESIZE_MODE)
mask = utils.resize_mask(mask, scale, padding)
# Compute Bounding box
bbox = utils.extract_bboxes(mask)
# Display image and additional stats
print("image_id: ", image_id)
print("Original shape: ", original_shape)
log("image", image)
log("mask", mask)
log("class_ids", class_ids)
log("bbox", bbox)
# Display image and instances
visualize.display_instances(image, bbox, mask, class_ids, dataset.class_names)
image_id = np.random.choice(dataset.image_ids, 1)[0]
image, image_meta, class_ids, bbox, mask = modellib.load_image_gt(
dataset, config, image_id, use_mini_mask=False)
log("image", image)
log("image_meta", image_meta)
log("class_ids", class_ids)
log("bbox", bbox)
log("mask", mask)
display_images([image]+[mask[:,:,i] for i in range(min(mask.shape[-1], 7))])
visualize.display_instances(image, bbox, mask, class_ids, dataset.class_names)
# Generate Anchors
backbone_shapes = modellib.compute_backbone_shapes(config, config.IMAGE_SHAPE)
anchors = utils.generate_pyramid_anchors(config.RPN_ANCHOR_SCALES,
config.RPN_ANCHOR_RATIOS,
backbone_shapes,
config.BACKBONE_STRIDES,
config.RPN_ANCHOR_STRIDE)
# Print summary of anchors
num_levels = len(backbone_shapes)
anchors_per_cell = len(config.RPN_ANCHOR_RATIOS)
print("Count: ", anchors.shape[0])
print("Scales: ", config.RPN_ANCHOR_SCALES)
print("ratios: ", config.RPN_ANCHOR_RATIOS)
print("Anchors per Cell: ", anchors_per_cell)
print("Levels: ", num_levels)
anchors_per_level = []
for l in range(num_levels):
num_cells = backbone_shapes[l][0] * backbone_shapes[l][1]
anchors_per_level.append(anchors_per_cell * num_cells // config.RPN_ANCHOR_STRIDE**2)
print("Anchors in Level {}: {}".format(l, anchors_per_level[l]))
## Visualize anchors of one cell at the center of the feature map of a specific level
# Load and draw random image
image_id = np.random.choice(dataset.image_ids, 1)[0]
image, image_meta, _, _, _ = modellib.load_image_gt(dataset, config, image_id)
fig, ax = plt.subplots(1, figsize=(10, 10))
ax.imshow(image)
levels = len(backbone_shapes)
for level in range(levels):
colors = visualize.random_colors(levels)
# Compute the index of the anchors at the center of the image
level_start = sum(anchors_per_level[:level]) # sum of anchors of previous levels
level_anchors = anchors[level_start:level_start+anchors_per_level[level]]
print("Level {}. Anchors: {:6} Feature map Shape: {}".format(level, level_anchors.shape[0],
backbone_shapes[level]))
center_cell = backbone_shapes[level] // 2
center_cell_index = (center_cell[0] * backbone_shapes[level][1] + center_cell[1])
level_center = center_cell_index * anchors_per_cell
center_anchor = anchors_per_cell * (
(center_cell[0] * backbone_shapes[level][1] / config.RPN_ANCHOR_STRIDE**2) \
+ center_cell[1] / config.RPN_ANCHOR_STRIDE)
level_center = int(center_anchor)
# Draw anchors. Brightness show the order in the array, dark to bright.
for i, rect in enumerate(level_anchors[level_center:level_center+anchors_per_cell]):
y1, x1, y2, x2 = rect
p = patches.Rectangle((x1, y1), x2-x1, y2-y1, linewidth=2, facecolor='none',
edgecolor=(i+1)*np.array(colors[level]) / anchors_per_cell)
ax.add_patch(p)
# Create data generator
random_rois = 4000
g = modellib.data_generator(
dataset, config, shuffle=True, random_rois=random_rois,
batch_size=4,
detection_targets=True)
# Get Next Image
if random_rois:
[normalized_images, image_meta, rpn_match, rpn_bbox, gt_class_ids, gt_boxes, gt_masks, rpn_rois, rois], \
[mrcnn_class_ids, mrcnn_bbox, mrcnn_mask] = next(g)
log("rois", rois)
log("mrcnn_class_ids", mrcnn_class_ids)
log("mrcnn_bbox", mrcnn_bbox)
log("mrcnn_mask", mrcnn_mask)
else:
[normalized_images, image_meta, rpn_match, rpn_bbox, gt_boxes, gt_masks], _ = next(g)
log("gt_class_ids", gt_class_ids)
log("gt_boxes", gt_boxes)
log("gt_masks", gt_masks)
log("rpn_match", rpn_match, )
log("rpn_bbox", rpn_bbox)
image_id = modellib.parse_image_meta(image_meta)["image_id"][0]
print("image_id: ", image_id, dataset.image_reference(image_id))
# Remove the last dim in mrcnn_class_ids. It's only added
# to satisfy Keras restriction on target shape.
mrcnn_class_ids = mrcnn_class_ids[:, :, 0]
b = 0
# Restore original image (reverse normalization)
sample_image = modellib.unmold_image(normalized_images[b], config)
# Compute anchor shifts.
indices = np.where(rpn_match[b] == 1)[0]
refined_anchors = utils.apply_box_deltas(anchors[indices], rpn_bbox[b, :len(indices)] * config.RPN_BBOX_STD_DEV)
log("anchors", anchors)
log("refined_anchors", refined_anchors)
# Get list of positive anchors
positive_anchor_ids = np.where(rpn_match[b] == 1)[0]
print("Positive anchors: {}".format(len(positive_anchor_ids)))
negative_anchor_ids = np.where(rpn_match[b] == -1)[0]
print("Negative anchors: {}".format(len(negative_anchor_ids)))
neutral_anchor_ids = np.where(rpn_match[b] == 0)[0]
print("Neutral anchors: {}".format(len(neutral_anchor_ids)))
# ROI breakdown by class
for c, n in zip(dataset.class_names, np.bincount(mrcnn_class_ids[b].flatten())):
if n:
print("{:23}: {}".format(c[:20], n))
# Show positive anchors
visualize.draw_boxes(sample_image, boxes=anchors[positive_anchor_ids],
refined_boxes=refined_anchors)
# Show negative anchors
visualize.draw_boxes(sample_image, boxes=anchors[negative_anchor_ids])
# Show neutral anchors. They don't contribute to training.
visualize.draw_boxes(sample_image, boxes=anchors[np.random.choice(neutral_anchor_ids, 100)])
if random_rois:
# Class aware bboxes
bbox_specific = mrcnn_bbox[b, np.arange(mrcnn_bbox.shape[1]), mrcnn_class_ids[b], :]
# Refined ROIs
refined_rois = utils.apply_box_deltas(rois[b].astype(np.float32), bbox_specific[:, :4] * config.BBOX_STD_DEV)
# Class aware masks
mask_specific = mrcnn_mask[b, np.arange(mrcnn_mask.shape[1]), :, :, mrcnn_class_ids[b]]
visualize.draw_rois(sample_image, rois[b], refined_rois, mask_specific, mrcnn_class_ids[b], dataset.class_names)
# Any repeated ROIs?
rows = np.ascontiguousarray(rois[b]).view(np.dtype((np.void, rois.dtype.itemsize * rois.shape[-1])))
_, idx = np.unique(rows, return_index=True)
print("Unique ROIs: {} out of {}".format(len(idx), rois.shape[1]))
if random_rois:
# Dispalay ROIs and corresponding masks and bounding boxes
ids = random.sample(range(rois.shape[1]), 8)
images = []
titles = []
for i in ids:
image = visualize.draw_box(sample_image.copy(), rois[b,i,:4].astype(np.int32), [255, 0, 0])
image = visualize.draw_box(image, refined_rois[i].astype(np.int64), [0, 255, 0])
images.append(image)
titles.append("ROI {}".format(i))
images.append(mask_specific[i] * 255)
titles.append(dataset.class_names[mrcnn_class_ids[b,i]][:20])
display_images(images, titles, cols=4, cmap="Blues", interpolation="none")
# Check ratio of positive ROIs in a set of images.
if random_rois:
limit = 10
temp_g = modellib.data_generator(
dataset, config, shuffle=True, random_rois=10000,
batch_size=1, detection_targets=True)
total = 0
for i in range(limit):
_, [ids, _, _] = next(temp_g)
positive_rois = np.sum(ids[0] > 0)
total += positive_rois
print("{:5} {:5.2f}".format(positive_rois, positive_rois/ids.shape[1]))
print("Average percent: {:.2f}".format(total/(limit*ids.shape[1])))
exit()
def all_steps(dataset, datacfg, dnncfg):
'''
## Single entry point for all the steps for inspecting dataset
'''
## Uncomment for debugging
# inspectdata.load_and_display_dataset(dataset, datacfg)
# In[7]:
log.info("[7]. ---------------")
log.info("Load and display random images and masks---------------")
log.info("Bounding Boxes---------------")
load_and_display_random_sample(dataset, datacfg)
# In[9]:
log.info("[9]. ---------------")
log.info("Resize Images---------------")
load_and_resize_images(dataset, datacfg, dnncfg)
# In[10]:
log.info("[10]. ---------------")
log.info("Mini Masks---------------")
image_id = load_mini_masks(dataset, datacfg, dnncfg)
log.info("image_id: {}".format(image_id))
# In[11]:
log.info("[11]. ---------------")
log.info("Add augmentation and mask resizing---------------")
add_augmentation(dataset, datacfg, dnncfg, image_id)
info = dataset.image_info[image_id]
log.debug("info: {}".format(info))
# In[12]:
log.info("[12]. ---------------")
log.info("Anchors---------------")
backbone_shapes, anchors, anchors_per_level, anchors_per_cell = generate_anchors(dnncfg)
# In[13]:
log.info("[13]. ---------------")
log.info("Visualize anchors of one cell at the center of the feature map of a specific level---------------")
visualize_anchors_at_center(dataset, datacfg, dnncfg, backbone_shapes, anchors, anchors_per_level, anchors_per_cell)
# In[14]:
log.info("[14]. ---------------")
log.info("info---------------")
image_ids = dataset.image_ids
log.info(image_ids)
image_index = -1
image_index = (image_index + 1) % len(image_ids)
log.info("image_index:{}".format(image_index))
# In[15]:
log.info("[15]. ---------------")
log.info("data_generator---------------")
## Data Generator
# Create data generator
random_rois = 2000
g = modellib.data_generator(
dataset, datacfg, dnncfg, shuffle=True, random_rois=random_rois,
batch_size=4,
detection_targets=True)
# Uncomment to run the generator through a lot of images
# to catch rare errors
# for i in range(1000):
# log.debug(i)
# _, _ = next(g)
# Get Next Image
if random_rois:
[normalized_images, image_meta, rpn_match, rpn_bbox, gt_class_ids, gt_boxes, gt_masks, rpn_rois, rois], [mrcnn_class_ids, mrcnn_bbox, mrcnn_mask] = next(g)
customlog("rois", rois)
customlog("mrcnn_class_ids", mrcnn_class_ids)
customlog("mrcnn_bbox", mrcnn_bbox)
customlog("mrcnn_mask", mrcnn_mask)
else:
[normalized_images, image_meta, rpn_match, rpn_bbox, gt_boxes, gt_masks], _ = next(g)
customlog("gt_class_ids", gt_class_ids)
customlog("gt_boxes", gt_boxes)
customlog("gt_masks", gt_masks)
customlog("rpn_match", rpn_match, )
customlog("rpn_bbox", rpn_bbox)
image_id = modellib.parse_image_meta(image_meta)["image_id"][0]
# Remove the last dim in mrcnn_class_ids. It's only added
# to satisfy Keras restriction on target shape.
mrcnn_class_ids = mrcnn_class_ids[:,:,0]
# In[16]:
log.info("[16]. ---------------")
b = 0
# Restore original image (reverse normalization)
sample_image = modellib.unmold_image(normalized_images[b], dnncfg)
# Compute anchor shifts.
indices = np.where(rpn_match[b] == 1)[0]
refined_anchors = utils.apply_box_deltas(anchors[indices], rpn_bbox[b, :len(indices)] * dnncfg.RPN_BBOX_STD_DEV)
customlog("anchors", anchors)
customlog("refined_anchors", refined_anchors)
# Get list of positive anchors
positive_anchor_ids = np.where(rpn_match[b] == 1)[0]
log.info("Positive anchors: {}".format(len(positive_anchor_ids)))
negative_anchor_ids = np.where(rpn_match[b] == -1)[0]
log.info("Negative anchors: {}".format(len(negative_anchor_ids)))
neutral_anchor_ids = np.where(rpn_match[b] == 0)[0]
log.info("Neutral anchors: {}".format(len(neutral_anchor_ids)))
log.info("ROI breakdown by class---------------")
# ROI breakdown by class
for c, n in zip(dataset.class_names, np.bincount(mrcnn_class_ids[b].flatten())):
if n:
log.info("{:23}: {}".format(c[:20], n))
log.info("Show positive anchors---------------")
# Show positive anchors
fig, ax = plt.subplots(1, figsize=(16, 16))
visualize.draw_boxes(sample_image, boxes=anchors[positive_anchor_ids],
refined_boxes=refined_anchors, ax=ax)
# In[17]:
log.info("[17]. ---------------")
log.info("Show negative anchors---------------")
# Show negative anchors
visualize.draw_boxes(sample_image, boxes=anchors[negative_anchor_ids])
# In[18]:
log.info("[18]. ---------------")
log.info("Show neutral anchors. They don't contribute to training---------------")
# Show neutral anchors. They don't contribute to training.
visualize.draw_boxes(sample_image, boxes=anchors[np.random.choice(neutral_anchor_ids, 100)])
# In[19]:
log.info("[19]. ---------------")
log.info("ROIs---------------")
## ROIs
if random_rois:
# Class aware bboxes
bbox_specific = mrcnn_bbox[b, np.arange(mrcnn_bbox.shape[1]), mrcnn_class_ids[b], :]
# Refined ROIs
refined_rois = utils.apply_box_deltas(rois[b].astype(np.float32), bbox_specific[:,:4] * dnncfg.BBOX_STD_DEV)
# Class aware masks
mask_specific = mrcnn_mask[b, np.arange(mrcnn_mask.shape[1]), :, :, mrcnn_class_ids[b]]
visualize.draw_rois(sample_image, rois[b], refined_rois, mask_specific, mrcnn_class_ids[b], dataset.class_names)
# Any repeated ROIs?
rows = np.ascontiguousarray(rois[b]).view(np.dtype((np.void, rois.dtype.itemsize * rois.shape[-1])))
_, idx = np.unique(rows, return_index=True)
log.info("Unique ROIs: {} out of {}".format(len(idx), rois.shape[1]))
# In[20]:
log.info("[20]. ---------------")
log.info("Dispalay ROIs and corresponding masks and bounding boxes---------------")
if random_rois:
# Dispalay ROIs and corresponding masks and bounding boxes
ids = random.sample(range(rois.shape[1]), 8)
images = []
titles = []
for i in ids:
image = visualize.draw_box(sample_image.copy(), rois[b,i,:4].astype(np.int32), [255, 0, 0])
image = visualize.draw_box(image, refined_rois[i].astype(np.int64), [0, 255, 0])
images.append(image)
titles.append("ROI {}".format(i))
images.append(mask_specific[i] * 255)
titles.append(dataset.class_names[mrcnn_class_ids[b,i]][:20])
visualize.display_images(images, titles, cols=4, cmap="Blues", interpolation="none")
# In[21]:
log.info("[21]. ---------------")
log.info("Check ratio of positive ROIs in a set of images.---------------")
# Check ratio of positive ROIs in a set of images.
if random_rois:
limit = 10
temp_g = modellib.data_generator(
dataset, datacfg, dnncfg, shuffle=True, random_rois=10000,
batch_size=1, detection_targets=True)
total = 0
for i in range(limit):
_, [ids, _, _] = next(temp_g)
positive_rois = np.sum(ids[0] > 0)
total += positive_rois
log.info("{:5} {:5.2f}".format(positive_rois, positive_rois/ids.shape[1]))
log.info("Average percent: {:.2f}".format(total/(limit*ids.shape[1])))
def train(self,
train_dataset,
val_dataset,
learning_rate,
epochs,
layers,
augmentation=None,
custom_callbacks=None,
no_augmentation_sources=None):
assert self.mode == "training", "Create model in training mode."
# Pre-defined layer regular expressions
layer_regex = {
# all layers but the backbone
"heads": r"(mrcnn\_.*)|(rpn\_.*)|(fpn\_.*)",
# From a specific Resnet stage and up
"3+":
r"(res3.*)|(bn3.*)|(res4.*)|(bn4.*)|(res5.*)|(bn5.*)|(mrcnn\_.*)|(rpn\_.*)|(fpn\_.*)",
"4+":
r"(res4.*)|(bn4.*)|(res5.*)|(bn5.*)|(mrcnn\_.*)|(rpn\_.*)|(fpn\_.*)",
"5+": r"(res5.*)|(bn5.*)|(mrcnn\_.*)|(rpn\_.*)|(fpn\_.*)",
# All layers
"all": ".*",
}
if layers in layer_regex.keys():
layers = layer_regex[layers]
# Data generators
train_generator = modellib.data_generator(
train_dataset,
self.config,
shuffle=True,
augmentation=augmentation,
batch_size=self.config.BATCH_SIZE,
no_augmentation_sources=no_augmentation_sources)
val_generator = modellib.data_generator(
val_dataset,
self.config,
shuffle=True,
batch_size=self.config.BATCH_SIZE)
# Create log_dir if it does not exist
if not os.path.exists(self.log_dir):
os.makedirs(self.log_dir)
# Callbacks
callbacks = [
keras.callbacks.TensorBoard(log_dir=self.log_dir,
histogram_freq=0,
write_graph=True,
write_images=False),
keras.callbacks.ModelCheckpoint(self.checkpoint_path,
verbose=0,
save_weights_only=True),
]
# Add custom callbacks to the list
if custom_callbacks:
callbacks += custom_callbacks
# Train
modellib.log("\nStarting at epoch {}. LR={}\n".format(
self.epoch, learning_rate))
modellib.log("Checkpoint Path: {}".format(self.checkpoint_path))
self.set_trainable(layers)
self.compile(learning_rate, self.config.LEARNING_MOMENTUM)
# Work-around for Windows: Keras fails on Windows when using
# multiprocessing workers. See discussion here:
# https://github.com/matterport/Mask_RCNN/issues/13#issuecomment-353124009
if os.name is 'nt':
workers = 0
else:
workers = modellib.multiprocessing.cpu_count()
self.history = self.keras_model.fit_generator(
train_generator,
initial_epoch=self.epoch,
epochs=epochs,
steps_per_epoch=self.config.STEPS_PER_EPOCH,
callbacks=callbacks,
validation_data=val_generator,
validation_steps=self.config.VALIDATION_STEPS,
max_queue_size=100,
workers=workers,
use_multiprocessing=True,
)
self.epoch = max(self.epoch, epochs)
def inspect_data(dataset, config):
print("Image Count: {}".format(len(dataset.image_ids)))
print("Class Count: {}".format(dataset.num_classes))
for i, info in enumerate(dataset.class_info):
print("{:3}. {:50}".format(i, info['name']))
# ## Display Samples
#
# Load and display images and masks.
# In[4]:
# Load and display random samples
image_ids = np.random.choice(dataset.image_ids, 4)
for image_id in image_ids:
image = dataset.load_image(image_id)
mask, class_ids = dataset.load_mask(image_id)
visualize.display_top_masks(image, mask, class_ids,
dataset.class_names)
# ## Bounding Boxes
#
# Rather than using bounding box coordinates provided by the source datasets, we compute the bounding boxes from masks instead. This allows us to handle bounding boxes consistently regardless of the source dataset, and it also makes it easier to resize, rotate, or crop images because we simply generate the bounding boxes from the updates masks rather than computing bounding box transformation for each type of image transformation.
# In[5]:
# Load random image and mask.
image_id = random.choice(dataset.image_ids)
image = dataset.load_image(image_id)
mask, class_ids = dataset.load_mask(image_id)
# Compute Bounding box
bbox = utils.extract_bboxes(mask)
# Display image and additional stats
print("image_id ", image_id, dataset.image_reference(image_id))
log("image", image)
log("mask", mask)
log("class_ids", class_ids)
log("bbox", bbox)
# Display image and instances
visualize.display_instances(image, bbox, mask, class_ids,
dataset.class_names)
# ## Resize Images
#
# To support multiple images per batch, images are resized to one size (1024x1024). Aspect ratio is preserved, though. If an image is not square, then zero padding is added at the top/bottom or right/left.
# In[6]:
# Load random image and mask.
image_id = np.random.choice(dataset.image_ids, 1)[0]
image = dataset.load_image(image_id)
mask, class_ids = dataset.load_mask(image_id)
original_shape = image.shape
# Resize
image, window, scale, padding, _ = utils.resize_image(
image,
min_dim=config.IMAGE_MIN_DIM,
max_dim=config.IMAGE_MAX_DIM,
mode=config.IMAGE_RESIZE_MODE)
mask = utils.resize_mask(mask, scale, padding)
# Compute Bounding box
bbox = utils.extract_bboxes(mask)
# Display image and additional stats
print("image_id: ", image_id, dataset.image_reference(image_id))
print("Original shape: ", original_shape)
log("image", image)
log("mask", mask)
log("class_ids", class_ids)
log("bbox", bbox)
# Display image and instances
visualize.display_instances(image, bbox, mask, class_ids,
dataset.class_names)
# ## Mini Masks
#
# Instance binary masks can get large when training with high resolution images. For example, if training with 1024x1024 image then the mask of a single instance requires 1MB of memory (Numpy uses bytes for boolean values). If an image has 100 instances then that's 100MB for the masks alone.
#
# To improve training speed, we optimize masks by:
# * We store mask pixels that are inside the object bounding box, rather than a mask of the full image. Most objects are small compared to the image size, so we save space by not storing a lot of zeros around the object.
# * We resize the mask to a smaller size (e.g. 56x56). For objects that are larger than the selected size we lose a bit of accuracy. But most object annotations are not very accuracy to begin with, so this loss is negligable for most practical purposes. Thie size of the mini_mask can be set in the config class.
#
# To visualize the effect of mask resizing, and to verify the code correctness, we visualize some examples.
# In[7]:
image_id = np.random.choice(dataset.image_ids, 1)[0]
image, image_meta, class_ids, bbox, mask = modellib.load_image_gt(
dataset, config, image_id, use_mini_mask=False)
log("image", image)
log("image_meta", image_meta)
log("class_ids", class_ids)
log("bbox", bbox)
log("mask", mask)
display_images([image] +
[mask[:, :, i] for i in range(min(mask.shape[-1], 7))])
# In[8]:
visualize.display_instances(image, bbox, mask, class_ids,
dataset.class_names)
# In[9]:
# Add augmentation and mask resizing.
image, image_meta, class_ids, bbox, mask = modellib.load_image_gt(
dataset, config, image_id, augment=True, use_mini_mask=True)
log("mask", mask)
display_images([image] +
[mask[:, :, i] for i in range(min(mask.shape[-1], 7))])
# In[10]:
mask = utils.expand_mask(bbox, mask, image.shape)
visualize.display_instances(image, bbox, mask, class_ids,
dataset.class_names)
# ## Anchors
#
# The order of anchors is important. Use the same order in training and prediction phases. And it must match the order of the convolution execution.
#
# For an FPN network, the anchors must be ordered in a way that makes it easy to match anchors to the output of the convolution layers that predict anchor scores and shifts.
# * Sort by pyramid level first. All anchors of the first level, then all of the second and so on. This makes it easier to separate anchors by level.
# * Within each level, sort anchors by feature map processing sequence. Typically, a convolution layer processes a feature map starting from top-left and moving right row by row.
# * For each feature map cell, pick any sorting order for the anchors of different ratios. Here we match the order of ratios passed to the function.
#
# **Anchor Stride:**
# In the FPN architecture, feature maps at the first few layers are high resolution. For example, if the input image is 1024x1024 then the feature meap of the first layer is 256x256, which generates about 200K anchors (256*256*3). These anchors are 32x32 pixels and their stride relative to image pixels is 4 pixels, so there is a lot of overlap. We can reduce the load significantly if we generate anchors for every other cell in the feature map. A stride of 2 will cut the number of anchors by 4, for example.
#
# In this implementation we use an anchor stride of 2, which is different from the paper.
# In[11]:
# Generate Anchors
backbone_shapes = modellib.compute_backbone_shapes(config,
config.IMAGE_SHAPE)
anchors = utils.generate_pyramid_anchors(config.RPN_ANCHOR_SCALES,
config.RPN_ANCHOR_RATIOS,
backbone_shapes,
config.BACKBONE_STRIDES,
config.RPN_ANCHOR_STRIDE)
# Print summary of anchors
num_levels = len(backbone_shapes)
anchors_per_cell = len(config.RPN_ANCHOR_RATIOS)
print("Count: ", anchors.shape[0])
print("Scales: ", config.RPN_ANCHOR_SCALES)
print("ratios: ", config.RPN_ANCHOR_RATIOS)
print("Anchors per Cell: ", anchors_per_cell)
print("Levels: ", num_levels)
anchors_per_level = []
for l in range(num_levels):
num_cells = backbone_shapes[l][0] * backbone_shapes[l][1]
anchors_per_level.append(anchors_per_cell * num_cells //
config.RPN_ANCHOR_STRIDE**2)
print("Anchors in Level {}: {}".format(l, anchors_per_level[l]))
# Visualize anchors of one cell at the center of the feature map of a specific level.
# In[12]:
## Visualize anchors of one cell at the center of the feature map of a specific level
# Load and draw random image
image_id = np.random.choice(dataset.image_ids, 1)[0]
image, image_meta, _, _, _ = modellib.load_image_gt(
dataset, config, image_id)
fig, ax = plt.subplots(1, figsize=(10, 10))
ax.imshow(image)
levels = len(backbone_shapes)
for level in range(levels):
colors = visualize.random_colors(levels)
# Compute the index of the anchors at the center of the image
level_start = sum(
anchors_per_level[:level]) # sum of anchors of previous levels
level_anchors = anchors[level_start:level_start +
anchors_per_level[level]]
print("Level {}. Anchors: {:6} Feature map Shape: {}".format(
level, level_anchors.shape[0], backbone_shapes[level]))
center_cell = backbone_shapes[level] // 2
center_cell_index = (center_cell[0] * backbone_shapes[level][1] +
center_cell[1])
level_center = center_cell_index * anchors_per_cell
center_anchor = anchors_per_cell * (
(center_cell[0] * backbone_shapes[level][1] / config.RPN_ANCHOR_STRIDE**2) \
+ center_cell[1] / config.RPN_ANCHOR_STRIDE)
level_center = int(center_anchor)
# Draw anchors. Brightness show the order in the array, dark to bright.
for i, rect in enumerate(level_anchors[level_center:level_center +
anchors_per_cell]):
y1, x1, y2, x2 = rect
p = patches.Rectangle(
(x1, y1),
x2 - x1,
y2 - y1,
linewidth=2,
facecolor='none',
edgecolor=(i + 1) * np.array(colors[level]) / anchors_per_cell)
ax.add_patch(p)
# ## Data Generator
#
# In[13]:
# Create data generator
random_rois = 2000
g = modellib.data_generator(dataset,
config,
shuffle=True,
random_rois=random_rois,
batch_size=4,
detection_targets=True)
# In[14]:
# Uncomment to run the generator through a lot of images
# to catch rare errors
# for i in range(1000):
# print(i)
# _, _ = next(g)
# In[15]:
# Get Next Image
if random_rois:
[
normalized_images, image_meta, rpn_match, rpn_bbox, gt_class_ids,
gt_boxes, gt_masks, rpn_rois, rois
], [mrcnn_class_ids, mrcnn_bbox, mrcnn_mask] = next(g)
log("rois", rois)
log("mrcnn_class_ids", mrcnn_class_ids)
log("mrcnn_bbox", mrcnn_bbox)
log("mrcnn_mask", mrcnn_mask)
else:
[
normalized_images, image_meta, rpn_match, rpn_bbox, gt_boxes,
gt_masks
], _ = next(g)
log("gt_class_ids", gt_class_ids)
log("gt_boxes", gt_boxes)
log("gt_masks", gt_masks)
log(
"rpn_match",
rpn_match,
)
log("rpn_bbox", rpn_bbox)
image_id = modellib.parse_image_meta(image_meta)["image_id"][0]
print("image_id: ", image_id, dataset.image_reference(image_id))
# Remove the last dim in mrcnn_class_ids. It's only added
# to satisfy Keras restriction on target shape.
mrcnn_class_ids = mrcnn_class_ids[:, :, 0]
# In[16]:
b = 0
# Restore original image (reverse normalization)
sample_image = modellib.unmold_image(normalized_images[b], config)
# Compute anchor shifts.
indices = np.where(rpn_match[b] == 1)[0]
refined_anchors = utils.apply_box_deltas(
anchors[indices], rpn_bbox[b, :len(indices)] * config.RPN_BBOX_STD_DEV)
log("anchors", anchors)
log("refined_anchors", refined_anchors)
# Get list of positive anchors
positive_anchor_ids = np.where(rpn_match[b] == 1)[0]
print("Positive anchors: {}".format(len(positive_anchor_ids)))
negative_anchor_ids = np.where(rpn_match[b] == -1)[0]
print("Negative anchors: {}".format(len(negative_anchor_ids)))
neutral_anchor_ids = np.where(rpn_match[b] == 0)[0]
print("Neutral anchors: {}".format(len(neutral_anchor_ids)))
# ROI breakdown by class
for c, n in zip(dataset.class_names,
np.bincount(mrcnn_class_ids[b].flatten())):
if n:
print("{:23}: {}".format(c[:20], n))
# Show positive anchors
fig, ax = plt.subplots(1, figsize=(16, 16))
visualize.draw_boxes(sample_image,
boxes=anchors[positive_anchor_ids],
refined_boxes=refined_anchors,
ax=ax)
# In[17]:
# Show negative anchors
visualize.draw_boxes(sample_image, boxes=anchors[negative_anchor_ids])
# In[18]:
# Show neutral anchors. They don't contribute to training.
visualize.draw_boxes(sample_image,
boxes=anchors[np.random.choice(
neutral_anchor_ids, 100)])
# ## ROIs
# In[19]:
if random_rois:
# Class aware bboxes
bbox_specific = mrcnn_bbox[b,
np.arange(mrcnn_bbox.shape[1]),
mrcnn_class_ids[b], :]
# Refined ROIs
refined_rois = utils.apply_box_deltas(
rois[b].astype(np.float32),
bbox_specific[:, :4] * config.BBOX_STD_DEV)
# Class aware masks
mask_specific = mrcnn_mask[b,
np.arange(mrcnn_mask.shape[1]), :, :,
mrcnn_class_ids[b]]
visualize.draw_rois(sample_image, rois[b], refined_rois, mask_specific,
mrcnn_class_ids[b], dataset.class_names)
# Any repeated ROIs?
rows = np.ascontiguousarray(rois[b]).view(
np.dtype((np.void, rois.dtype.itemsize * rois.shape[-1])))
_, idx = np.unique(rows, return_index=True)
print("Unique ROIs: {} out of {}".format(len(idx), rois.shape[1]))
# In[20]:
if random_rois:
# Dispalay ROIs and corresponding masks and bounding boxes
ids = random.sample(range(rois.shape[1]), 8)
images = []
titles = []
for i in ids:
image = visualize.draw_box(sample_image.copy(),
rois[b, i, :4].astype(np.int32),
[255, 0, 0])
image = visualize.draw_box(image, refined_rois[i].astype(np.int64),
[0, 255, 0])
images.append(image)
titles.append("ROI {}".format(i))
images.append(mask_specific[i] * 255)
titles.append(dataset.class_names[mrcnn_class_ids[b, i]][:20])
display_images(images,
titles,
cols=4,
cmap="Blues",
interpolation="none")
# In[21]:
# Check ratio of positive ROIs in a set of images.
if random_rois:
limit = 10
temp_g = modellib.data_generator(dataset,
config,
shuffle=True,
random_rois=10000,
batch_size=1,
detection_targets=True)
total = 0
for i in range(limit):
_, [ids, _, _] = next(temp_g)
positive_rois = np.sum(ids[0] > 0)
total += positive_rois
print("{:5} {:5.2f}".format(positive_rois,
positive_rois / ids.shape[1]))
print("Average percent: {:.2f}".format(total / (limit * ids.shape[1])))
def train(self, train_dataset, val_dataset, learning_rate, epochs, layers,
augmentation=None, custom_callbacks=None, no_augmentation_sources=None):
"""Train the model.
train_dataset, val_dataset: Training and validation Dataset objects.
learning_rate: The learning rate to train with
epochs: Number of training epochs. Note that previous training epochs
are considered to be done alreay, so this actually determines
the epochs to train in total rather than in this particaular
call.
layers: Allows selecting wich layers to train. It can be:
- A regular expression to match layer names to train
- One of these predefined values:
heads: The RPN, classifier and mask heads of the network
all: All the layers
3+: Train Resnet stage 3 and up
4+: Train Resnet stage 4 and up
5+: Train Resnet stage 5 and up
augmentation: Optional. An imgaug (https://github.com/aleju/imgaug)
augmentation. For example, passing imgaug.augmenters.Fliplr(0.5)
flips images right/left 50% of the time. You can pass complex
augmentations as well. This augmentation applies 50% of the
time, and when it does it flips images right/left half the time
and adds a Gaussian blur with a random sigma in range 0 to 5.
augmentation = imgaug.augmenters.Sometimes(0.5, [
imgaug.augmenters.Fliplr(0.5),
imgaug.augmenters.GaussianBlur(sigma=(0.0, 5.0))
])
custom_callbacks: Optional. Add custom callbacks to be called
with the keras fit_generator method. Must be list of type keras.callbacks.
no_augmentation_sources: Optional. List of sources to exclude for
augmentation. A source is string that identifies a dataset and is
defined in the Dataset class.
"""
assert self.mode == "training", "Create model in training mode."
# Pre-defined layer regular expressions
layer_regex = {
# all layers but the backbone
"heads": r"(mrcnn\_.*)|(rpn\_.*)|(fpn\_.*)",
# From a specific Resnet stage and up
"3+": r"(res3.*)|(bn3.*)|(res4.*)|(bn4.*)|(res5.*)|(bn5.*)|(mrcnn\_.*)|(rpn\_.*)|(fpn\_.*)",
"4+": r"(res4.*)|(bn4.*)|(res5.*)|(bn5.*)|(mrcnn\_.*)|(rpn\_.*)|(fpn\_.*)",
"5+": r"(res5.*)|(bn5.*)|(mrcnn\_.*)|(rpn\_.*)|(fpn\_.*)",
# All layers
"all": ".*",
}
if layers in layer_regex.keys():
layers = layer_regex[layers]
# Data generators
train_generator = data_generator(train_dataset, self.config, shuffle=True,
augmentation=augmentation,
batch_size=self.config.BATCH_SIZE,
no_augmentation_sources=no_augmentation_sources)
val_generator = data_generator(val_dataset, self.config, shuffle=True,
batch_size=self.config.BATCH_SIZE)
# Create log_dir if it does not exist
if not os.path.exists(self.log_dir):
os.makedirs(self.log_dir)
# Callbacks
callbacks = []
# Add custom callbacks to the list
if custom_callbacks:
callbacks += custom_callbacks
# Train
log("\nStarting at epoch {}. LR={}\n".format(self.epoch, learning_rate))
log("Checkpoint Path: {}".format(self.checkpoint_path))
self.set_trainable(layers)
self.compile(learning_rate, self.config.LEARNING_MOMENTUM)
# Work-around for Windows: Keras fails on Windows when using
# multiprocessing workers. See discussion here:
# https://github.com/matterport/Mask_RCNN/issues/13#issuecomment-353124009
if os.name == 'nt':
workers = 0
else:
workers = multiprocessing.cpu_count()
self.keras_model.fit_generator(
train_generator,
initial_epoch=self.epoch,
epochs=epochs,
steps_per_epoch=self.config.STEPS_PER_EPOCH,
callbacks=callbacks,
validation_data=val_generator,
validation_steps=self.config.VALIDATION_STEPS,
max_queue_size=100,
workers=workers,
use_multiprocessing=True,
)
self.epoch = max(self.epoch, epochs)
