def train_faster_rcnn_e2e(cfg):
# Input variables denoting features and labeled ground truth rois (as 5-tuples per roi)
image_input = input_variable(shape=(cfg.NUM_CHANNELS, cfg.IMAGE_HEIGHT, cfg.IMAGE_WIDTH),
dynamic_axes=[Axis.default_batch_axis()],
name=cfg["MODEL"].FEATURE_NODE_NAME)
roi_input = input_variable((cfg.INPUT_ROIS_PER_IMAGE, 5), dynamic_axes=[Axis.default_batch_axis()])
dims_input = input_variable((6), dynamic_axes=[Axis.default_batch_axis()])
dims_node = alias(dims_input, name='dims_input')
# Instantiate the Faster R-CNN prediction model and loss function
loss, pred_error = create_faster_rcnn_model(image_input, roi_input, dims_node, cfg)
if cfg["CNTK"].DEBUG_OUTPUT:
print("Storing graphs and models to %s." % cfg.OUTPUT_PATH)
plot(loss, os.path.join(cfg.OUTPUT_PATH, "graph_frcn_train_e2e." + cfg["CNTK"].GRAPH_TYPE))
# Set learning parameters
e2e_lr_factor = cfg["MODEL"].E2E_LR_FACTOR
e2e_lr_per_sample_scaled = [x * e2e_lr_factor for x in cfg["CNTK"].E2E_LR_PER_SAMPLE]
mm_schedule = momentum_schedule(cfg["CNTK"].MOMENTUM_PER_MB)
print("Using base model: {}".format(cfg["MODEL"].BASE_MODEL))
print("lr_per_sample: {}".format(e2e_lr_per_sample_scaled))
train_model(image_input, roi_input, dims_input, loss, pred_error,
e2e_lr_per_sample_scaled, mm_schedule, cfg["CNTK"].L2_REG_WEIGHT, cfg["CNTK"].E2E_MAX_EPOCHS, cfg)
return create_faster_rcnn_eval_model(loss, image_input, dims_input, cfg)
def train_faster_rcnn(cfg):
# Train only if no model exists yet
model_path = cfg['MODEL_PATH']
if os.path.exists(model_path) and cfg["CNTK"].MAKE_MODE:
print("Loading existing model from %s" % model_path)
eval_model = load_model(model_path)
else:
if cfg["CNTK"].TRAIN_E2E:
eval_model = train_faster_rcnn_e2e(cfg)
else:
eval_model = train_faster_rcnn_alternating(cfg)
eval_model.save(model_path)
print("Stored eval model at %s" % model_path)
if cfg["CNTK"].DEBUG_OUTPUT:
plot(
eval_model,
os.path.join(
cfg.OUTPUT_PATH, "graph_frcn_eval_{}_{}.{}".format(
cfg["MODEL"].BASE_MODEL,
"e2e" if cfg["CNTK"].TRAIN_E2E else "4stage",
cfg["CNTK"].GRAPH_TYPE)))
upload_checkpoint_file(os.path.join(
os.environ['AZUREML_NATIVE_SHARE_DIRECTORY'], 'output',
cfg['OUTPUT_MODEL_NAME']),
cfg['OUTPUT_MODEL_NAME'],
add_timestamp=True)
print("Model saved to Azure blob")
return eval_model
def create_fast_rcnn_eval_model(model, image_input, roi_proposals, cfg):
print("creating eval model")
predictor = clone_model(model,
[cfg["MODEL"].FEATURE_NODE_NAME, "roi_proposals"],
["cls_score", "bbox_regr"], CloneMethod.freeze)
pred_net = predictor(image_input, roi_proposals)
cls_score = pred_net.outputs[0]
bbox_regr = pred_net.outputs[1]
if cfg.BBOX_NORMALIZE_TARGETS:
num_boxes = int(bbox_regr.shape[1] / 4)
bbox_normalize_means = np.array(cfg.BBOX_NORMALIZE_MEANS * num_boxes)
bbox_normalize_stds = np.array(cfg.BBOX_NORMALIZE_STDS * num_boxes)
bbox_regr = plus(element_times(bbox_regr, bbox_normalize_stds),
bbox_normalize_means,
name='bbox_regr')
cls_pred = softmax(cls_score, axis=1, name='cls_pred')
eval_model = combine([cls_pred, bbox_regr])
if cfg["CNTK"].DEBUG_OUTPUT:
plot(
eval_model,
os.path.join(cfg.OUTPUT_PATH,
"graph_frcn_eval." + cfg["CNTK"].GRAPH_TYPE))
return eval_model
def train_faster_rcnn_e2e(base_model_file_name, debug_output=False):
# Input variables denoting features and labeled ground truth rois (as 5-tuples per roi)
image_input = input_variable((num_channels, image_height, image_width), dynamic_axes=[Axis.default_batch_axis()], name=feature_node_name)
roi_input = input_variable((cfg["CNTK"].INPUT_ROIS_PER_IMAGE, 5), dynamic_axes=[Axis.default_batch_axis()])
dims_input = input_variable((6), dynamic_axes=[Axis.default_batch_axis()])
dims_node = alias(dims_input, name='dims_input')
# Instantiate the Faster R-CNN prediction model and loss function
loss, pred_error = create_faster_rcnn_predictor(base_model_file_name, image_input, roi_input, dims_node)
if debug_output:
print("Storing graphs and models to %s." % globalvars['output_path'])
plot(loss, os.path.join(globalvars['output_path'], "graph_frcn_train_e2e." + cfg["CNTK"].GRAPH_TYPE))
# Set learning parameters
e2e_lr_factor = globalvars['e2e_lr_factor']
e2e_lr_per_sample_scaled = [x * e2e_lr_factor for x in cfg["CNTK"].E2E_LR_PER_SAMPLE]
mm_schedule = momentum_schedule(cfg["CNTK"].MOMENTUM_PER_MB)
print("Using base model: {}".format(cfg["CNTK"].BASE_MODEL))
print("lr_per_sample: {}".format(e2e_lr_per_sample_scaled))
train_model(image_input, roi_input, dims_input, loss, pred_error,
e2e_lr_per_sample_scaled, mm_schedule, cfg["CNTK"].L2_REG_WEIGHT, globalvars['e2e_epochs'])
return create_eval_model(loss, image_input, dims_input)
def train_faster_rcnn(cfg):
# Train only if no model exists yet
model_path = cfg['MODEL_PATH']
if os.path.exists(model_path) and cfg["CNTK"].MAKE_MODE:
print("Loading existing model from %s" % model_path)
eval_model = load_model(model_path)
else:
if cfg["CNTK"].TRAIN_E2E:
eval_model = train_faster_rcnn_e2e(cfg)
else:
eval_model = train_faster_rcnn_alternating(cfg)
eval_model.save(model_path)
if cfg["CNTK"].DEBUG_OUTPUT:
plot(
eval_model,
os.path.join(
cfg.OUTPUT_PATH, "graph_frcn_eval_{}_{}.{}".format(
cfg["MODEL"].BASE_MODEL,
"e2e" if cfg["CNTK"].TRAIN_E2E else "4stage",
cfg["CNTK"].GRAPH_TYPE)))
print("Stored eval model at %s" % model_path)
if cfg.DISTRIBUTED_FLG:
distributed.Communicator.finalize()
return eval_model
def train_faster_rcnn_e2e(base_model_file_name, debug_output=False):
# Input variables denoting features and labeled ground truth rois (as 5-tuples per roi)
image_input = input_variable((num_channels, image_height, image_width), dynamic_axes=[Axis.default_batch_axis()], name=feature_node_name)
roi_input = input_variable((cfg["CNTK"].INPUT_ROIS_PER_IMAGE, 5), dynamic_axes=[Axis.default_batch_axis()])
dims_input = input_variable((6), dynamic_axes=[Axis.default_batch_axis()])
dims_node = alias(dims_input, name='dims_input')
# Instantiate the Faster R-CNN prediction model and loss function
loss, pred_error = create_faster_rcnn_predictor(base_model_file_name, image_input, roi_input, dims_node)
if debug_output:
print("Storing graphs and models to %s." % globalvars['output_path'])
plot(loss, os.path.join(globalvars['output_path'], "graph_frcn_train_e2e." + cfg["CNTK"].GRAPH_TYPE))
# Set learning parameters
e2e_lr_factor = globalvars['e2e_lr_factor']
e2e_lr_per_sample_scaled = [x * e2e_lr_factor for x in cfg["CNTK"].E2E_LR_PER_SAMPLE]
mm_schedule = momentum_schedule(cfg["CNTK"].MOMENTUM_PER_MB)
print("Using base model: {}".format(cfg["CNTK"].BASE_MODEL))
print("lr_per_sample: {}".format(e2e_lr_per_sample_scaled))
train_model(image_input, roi_input, dims_input, loss, pred_error,
e2e_lr_per_sample_scaled, mm_schedule, cfg["CNTK"].L2_REG_WEIGHT, globalvars['e2e_epochs'])
return create_eval_model(loss, image_input, dims_input)
def train_faster_rcnn_e2e(cfg):
# Input variables denoting features and labeled ground truth rois (as 5-tuples per roi)
image_input = input_variable(shape=(cfg.NUM_CHANNELS, cfg.IMAGE_HEIGHT, cfg.IMAGE_WIDTH),
dynamic_axes=[Axis.default_batch_axis()],
name=cfg["MODEL"].FEATURE_NODE_NAME)
roi_input = input_variable((cfg.INPUT_ROIS_PER_IMAGE, 5), dynamic_axes=[Axis.default_batch_axis()])
dims_input = input_variable((6), dynamic_axes=[Axis.default_batch_axis()])
dims_node = alias(dims_input, name='dims_input')
# Instantiate the Faster R-CNN prediction model and loss function
loss, pred_error = create_faster_rcnn_model(image_input, roi_input, dims_node, cfg)
if cfg["CNTK"].DEBUG_OUTPUT:
print("Storing graphs and models to %s." % cfg.OUTPUT_PATH)
plot(loss, os.path.join(cfg.OUTPUT_PATH, "graph_frcn_train_e2e." + cfg["CNTK"].GRAPH_TYPE))
# Set learning parameters
e2e_lr_factor = cfg["MODEL"].E2E_LR_FACTOR
e2e_lr_per_sample_scaled = [x * e2e_lr_factor for x in cfg["CNTK"].E2E_LR_PER_SAMPLE]
mm_schedule = momentum_schedule(cfg["CNTK"].MOMENTUM_PER_MB)
print("Using base model: {}".format(cfg["MODEL"].BASE_MODEL))
print("lr_per_sample: {}".format(e2e_lr_per_sample_scaled))
train_model(image_input, roi_input, dims_input, loss, pred_error,
e2e_lr_per_sample_scaled, mm_schedule, cfg["CNTK"].L2_REG_WEIGHT, cfg["CNTK"].E2E_MAX_EPOCHS, cfg)
return create_faster_rcnn_eval_model(loss, image_input, dims_input, cfg)
def plot_model(root, output_file="model.svg"):
"""Plots the CNTK model starting from the root node to an output image
Pre-requisites:
Install graphviz executables from graphviz.org
Update your PATH environment to include the path to graphviz
pip install graphviz
pip install pydot_ng
"""
graph.plot(root, output_file)
_logger.info("Model graph plotted to: {}".format(output_file))
def train_fast_rcnn(debug_output=False, model_path=model_file):
if debug_output:
print("Storing graphs and intermediate models to %s." % os.path.join(abs_path, "Output"))
# Create the minibatch source
minibatch_source = create_mb_source(image_height, image_width, num_channels,
num_classes, num_rois, base_path, "train")
# Input variables denoting features, rois and label data
image_input = C.input_variable((num_channels, image_height, image_width))
roi_input = C.input_variable((num_rois, 4))
label_input = C.input_variable((num_rois, num_classes))
# define mapping from reader streams to network inputs
input_map = {
image_input: minibatch_source.streams.features,
roi_input: minibatch_source.streams.rois,
label_input: minibatch_source.streams.roiLabels
}
# Instantiate the Fast R-CNN prediction model and loss function
frcn_output = frcn_predictor(image_input, roi_input, num_classes, model_path)
ce = cross_entropy_with_softmax(frcn_output, label_input, axis=1)
pe = classification_error(frcn_output, label_input, axis=1)
if debug_output:
plot(frcn_output, os.path.join(abs_path, "Output", "graph_frcn.png"))
# Set learning parameters
l2_reg_weight = 0.0005
lr_per_sample = [0.00001] * 10 + [0.000001] * 5 + [0.0000001]
lr_schedule = learning_rate_schedule(lr_per_sample, unit=UnitType.sample)
mm_schedule = momentum_as_time_constant_schedule(momentum_time_constant)
# Instantiate the trainer object
learner = momentum_sgd(frcn_output.parameters, lr_schedule, mm_schedule, l2_regularization_weight=l2_reg_weight)
progress_printer = ProgressPrinter(tag='Training', num_epochs=max_epochs)
trainer = Trainer(frcn_output, (ce, pe), learner, progress_printer)
# Get minibatches of images and perform model training
print("Training Fast R-CNN model for %s epochs." % max_epochs)
log_number_of_parameters(frcn_output)
for epoch in range(max_epochs): # loop over epochs
sample_count = 0
while sample_count < epoch_size: # loop over minibatches in the epoch
data = minibatch_source.next_minibatch(min(mb_size, epoch_size-sample_count), input_map=input_map)
trainer.train_minibatch(data) # update model with it
sample_count += trainer.previous_minibatch_sample_count # count samples processed so far
trainer.summarize_training_progress()
if debug_output:
frcn_output.save(os.path.join(abs_path, "Output", "frcn_py_%s.model" % (epoch+1)))
return frcn_output
def train_fast_rcnn(debug_output=False):
if debug_output:
print("Storing graphs and intermediate models to %s." % os.path.join(abs_path, "Output"))
# Create the minibatch source
minibatch_source = create_mb_source(image_height, image_width, num_channels,
num_classes, num_rois, base_path, "train")
# Input variables denoting features, rois and label data
image_input = input((num_channels, image_height, image_width))
roi_input = input((num_rois, 4))
label_input = input((num_rois, num_classes))
# define mapping from reader streams to network inputs
input_map = {
image_input: minibatch_source.streams.features,
roi_input: minibatch_source.streams.rois,
label_input: minibatch_source.streams.roiLabels
}
# Instantiate the Fast R-CNN prediction model and loss function
frcn_output = frcn_predictor(image_input, roi_input, num_classes)
ce = cross_entropy_with_softmax(frcn_output, label_input, axis=1)
pe = classification_error(frcn_output, label_input, axis=1)
if debug_output:
plot(frcn_output, os.path.join(abs_path, "Output", "graph_frcn.png"))
# Set learning parameters
l2_reg_weight = 0.0005
lr_per_sample = [0.00001] * 10 + [0.000001] * 5 + [0.0000001]
lr_schedule = learning_rate_schedule(lr_per_sample, unit=UnitType.sample)
mm_schedule = momentum_as_time_constant_schedule(momentum_time_constant)
# Instantiate the trainer object
learner = momentum_sgd(frcn_output.parameters, lr_schedule, mm_schedule, l2_regularization_weight=l2_reg_weight)
progress_printer = ProgressPrinter(tag='Training', num_epochs=max_epochs)
trainer = Trainer(frcn_output, (ce, pe), learner, progress_printer)
# Get minibatches of images and perform model training
print("Training Fast R-CNN model for %s epochs." % max_epochs)
log_number_of_parameters(frcn_output)
for epoch in range(max_epochs): # loop over epochs
sample_count = 0
while sample_count < epoch_size: # loop over minibatches in the epoch
data = minibatch_source.next_minibatch(min(mb_size, epoch_size-sample_count), input_map=input_map)
trainer.train_minibatch(data) # update model with it
sample_count += trainer.previous_minibatch_sample_count # count samples processed so far
trainer.summarize_training_progress()
if debug_output:
frcn_output.save(os.path.join(abs_path, "Output", "frcn_py_%s.model" % (epoch+1)))
return frcn_output
def train (self, train_file, output_resources_pickle_file, \
network_type = 'unidirectional', \
num_epochs = 1, batch_size = 50, \
dropout = 0.2, reg_alpha = 0.0, \
num_hidden_units = 150, num_layers = 1):
train_X, train_Y = self.reader.read_and_parse_training_data(train_file, output_resources_pickle_file)
print("Data Shape: ")
print(train_X.shape) # (15380, 613)
print(train_Y.shape) # (15380, 613, 8)
#self.wordvecs.shape (66962, 50)
print("Hyper parameters:")
print("output_resources_pickle_file = {}".format(output_resources_pickle_file))
print("network_type = {}".format(network_type))
print("num_epochs= {}".format(num_epochs ))
print("batch_size = {}".format(batch_size ))
print("dropout = ".format(dropout ))
print("reg_alpha = {}".format(reg_alpha ))
print("num_hidden_units = {}".format(num_hidden_units))
print("num_layers = {}".format(num_layers ))
# Instantiate the model function;
features = C.sequence.input_variable(self.wordvecs.shape[0])
labels = C.input_variable(train_Y.shape[2], dynamic_axes=[C.Axis.default_batch_axis()])
self.model = self.__create_model(features, train_Y.shape[2], num_hidden_units, dropout)
plot_path = "./lstm_model.png"
plot(self.model, plot_path)
# Instantiate the loss and error function
loss = C.cross_entropy_with_softmax(self.model, labels)
error = C.classification_error(self.model, labels)
# LR schedule
learning_rate = 0.02
lr_schedule = C.learning_parameter_schedule(learning_rate)
momentum_schedule = C.momentum_schedule(0.9, minibatch_size=batch_size)
learner = C.fsadagrad(self.model.parameters, lr = lr_schedule, momentum = momentum_schedule, unit_gain = True)
# Setup the progress updater
progress_printer = C.logging.ProgressPrinter(freq=100, first=10, tag='Training', num_epochs=num_epochs)
# Instantiate the trainer. We have all data in memory. https://github.com/Microsoft/CNTK/blob/master/Manual/Manual_How_to_feed_data.ipynb
print('Start training')
train_summary = loss.train((train_X.astype('float32'), train_Y.astype('float32')), parameter_learners=[learner], callbacks=[progress_printer])
def train_faster_rcnn(cfg):
# Train only if no model exists yet
model_path = cfg['MODEL_PATH']
if os.path.exists(model_path) and cfg["CNTK"].MAKE_MODE:
print("Loading existing model from %s" % model_path)
eval_model = load_model(model_path)
else:
if cfg["CNTK"].TRAIN_E2E:
eval_model = train_faster_rcnn_e2e(cfg)
else:
eval_model = train_faster_rcnn_alternating(cfg)
eval_model.save(model_path)
if cfg["CNTK"].DEBUG_OUTPUT:
plot(eval_model, os.path.join(cfg.OUTPUT_PATH, "graph_frcn_eval_{}_{}.{}"
.format(cfg["MODEL"].BASE_MODEL, "e2e" if cfg["CNTK"].TRAIN_E2E else "4stage", cfg["CNTK"].GRAPH_TYPE)))
print("Stored eval model at %s" % model_path)
return eval_model
def create_fast_rcnn_eval_model(model, image_input, roi_proposals, cfg):
print("creating eval model")
predictor = clone_model(model, [cfg["MODEL"].FEATURE_NODE_NAME, "roi_proposals"], ["cls_score", "bbox_regr"], CloneMethod.freeze)
pred_net = predictor(image_input, roi_proposals)
cls_score = pred_net.outputs[0]
bbox_regr = pred_net.outputs[1]
if cfg.BBOX_NORMALIZE_TARGETS:
num_boxes = int(bbox_regr.shape[1] / 4)
bbox_normalize_means = np.array(cfg.BBOX_NORMALIZE_MEANS * num_boxes)
bbox_normalize_stds = np.array(cfg.BBOX_NORMALIZE_STDS * num_boxes)
bbox_regr = plus(element_times(bbox_regr, bbox_normalize_stds), bbox_normalize_means, name='bbox_regr')
cls_pred = softmax(cls_score, axis=1, name='cls_pred')
eval_model = combine([cls_pred, bbox_regr])
if cfg["CNTK"].DEBUG_OUTPUT:
plot(eval_model, os.path.join(cfg.OUTPUT_PATH, "graph_frcn_eval." + cfg["CNTK"].GRAPH_TYPE))
return eval_model
def create_model(base_model_file, input_features, params):
num_classes = params['num_classes']
dropout_rate = params['dropout_rate']
freeze_weights = params['freeze_weights']
# Load the pretrained classification net and find nodes
base_model = load_model(base_model_file)
log = logging.getLogger("neuralnets1.utils.create_model")
log.info('Loaded base model - %s with layers:' % base_model_file)
node_outputs = get_node_outputs(base_model)
[log.info('%s , %s' % (layer.name, layer.shape)) for layer in node_outputs]
graph.plot(base_model,
filename="base_model.pdf") # Write graph visualization
feature_node = find_by_name(base_model, 'features')
beforePooling_node = find_by_name(base_model, "z.x.x.r")
# Clone model until right before the pooling layer, ie. until including z.x.x.r
modelCloned = combine([beforePooling_node.owner]).clone(
CloneMethod.freeze if freeze_weights else CloneMethod.clone,
{feature_node: placeholder(name='features')})
# Center the input around zero and set model input.
# Do this early, to avoid CNTK bug with wrongly estimated layer shapes
feat_norm = input_features - constant(114)
model = modelCloned(feat_norm)
# Add pool layer
avgPool = GlobalAveragePooling(name="poolingLayer")(
model) # assign name to the layer and add to the model
# Add drop out layer
if dropout_rate > 0:
avgPoolDrop = Dropout(dropout_rate)(
avgPool
) # add drop out layer with specified drop out rate and add it to the model
else:
avgPoolDrop = avgPool
# Add new dense layer for class prediction
finalModel = Dense(num_classes, activation=None, name="Dense")(avgPoolDrop)
return finalModel
def plot_model(root, output_file="model.png"):
"""Plots the CNTK model starting from the root node to an output image
Pre-requisites:
Install graphviz executables from graphviz.org
Update your PATH environment variable to include the path to graphviz
pip install graphviz
pip install pydot_ng
"""
text = graph.plot(root, output_file)
print(text)
def train_fast_rcnn(cfg):
# Train only if no model exists yet
model_path = cfg['MODEL_PATH']
if os.path.exists(model_path) and cfg["CNTK"].MAKE_MODE:
print("Loading existing model from %s" % model_path)
return load_model(model_path)
else:
# Input variables denoting features and labeled ground truth rois (as 5-tuples per roi)
image_input = input_variable(shape=(cfg.NUM_CHANNELS, cfg.IMAGE_HEIGHT,
cfg.IMAGE_WIDTH),
dynamic_axes=[Axis.default_batch_axis()],
name=cfg["MODEL"].FEATURE_NODE_NAME)
roi_proposals = input_variable(
(cfg.NUM_ROI_PROPOSALS, 4),
dynamic_axes=[Axis.default_batch_axis()],
name="roi_proposals")
label_targets = input_variable(
(cfg.NUM_ROI_PROPOSALS, cfg["DATA"].NUM_CLASSES),
dynamic_axes=[Axis.default_batch_axis()])
bbox_targets = input_variable(
(cfg.NUM_ROI_PROPOSALS, 4 * cfg["DATA"].NUM_CLASSES),
dynamic_axes=[Axis.default_batch_axis()])
bbox_inside_weights = input_variable(
(cfg.NUM_ROI_PROPOSALS, 4 * cfg["DATA"].NUM_CLASSES),
dynamic_axes=[Axis.default_batch_axis()])
# Instantiate the Fast R-CNN prediction model and loss function
loss, pred_error = create_fast_rcnn_model(image_input, roi_proposals,
label_targets, bbox_targets,
bbox_inside_weights, cfg)
if isinstance(loss, cntk.Variable):
loss = combine([loss])
if cfg["CNTK"].DEBUG_OUTPUT:
print("Storing graphs and models to %s." % cfg.OUTPUT_PATH)
plot(
loss,
os.path.join(cfg.OUTPUT_PATH,
"graph_frcn_train." + cfg["CNTK"].GRAPH_TYPE))
# Set learning parameters
lr_factor = cfg["CNTK"].LR_FACTOR
lr_per_sample_scaled = [
x * lr_factor for x in cfg["CNTK"].LR_PER_SAMPLE
]
mm_schedule = momentum_schedule(cfg["CNTK"].MOMENTUM_PER_MB)
l2_reg_weight = cfg["CNTK"].L2_REG_WEIGHT
epochs_to_train = cfg["CNTK"].MAX_EPOCHS
print("Using base model: {}".format(cfg["MODEL"].BASE_MODEL))
print("lr_per_sample: {}".format(lr_per_sample_scaled))
# --- train ---
# Instantiate the learners and the trainer object
params = loss.parameters
biases = [p for p in params if '.b' in p.name or 'b' == p.name]
others = [p for p in params if not p in biases]
bias_lr_mult = cfg["CNTK"].BIAS_LR_MULT
lr_schedule = learning_parameter_schedule_per_sample(
lr_per_sample_scaled)
learner = momentum_sgd(others,
lr_schedule,
mm_schedule,
l2_regularization_weight=l2_reg_weight,
unit_gain=False,
use_mean_gradient=True)
bias_lr_per_sample = [
v * bias_lr_mult for v in cfg["CNTK"].LR_PER_SAMPLE
]
bias_lr_schedule = learning_parameter_schedule_per_sample(
bias_lr_per_sample)
bias_learner = momentum_sgd(biases,
bias_lr_schedule,
mm_schedule,
l2_regularization_weight=l2_reg_weight,
unit_gain=False,
use_mean_gradient=True)
trainer = Trainer(None, (loss, pred_error), [learner, bias_learner])
# Get minibatches of images and perform model training
print("Training model for %s epochs." % epochs_to_train)
log_number_of_parameters(loss)
# Create the minibatch source
if cfg.USE_PRECOMPUTED_PROPOSALS:
proposal_provider = ProposalProvider.fromfile(
cfg["DATA"].TRAIN_PRECOMPUTED_PROPOSALS_FILE,
cfg.NUM_ROI_PROPOSALS)
else:
proposal_provider = ProposalProvider.fromconfig(cfg)
od_minibatch_source = ObjectDetectionMinibatchSource(
cfg["DATA"].TRAIN_MAP_FILE,
cfg["DATA"].TRAIN_ROI_FILE,
max_annotations_per_image=cfg.INPUT_ROIS_PER_IMAGE,
pad_width=cfg.IMAGE_WIDTH,
pad_height=cfg.IMAGE_HEIGHT,
pad_value=cfg["MODEL"].IMG_PAD_COLOR,
randomize=True,
use_flipping=cfg["TRAIN"].USE_FLIPPED,
max_images=cfg["DATA"].NUM_TRAIN_IMAGES,
num_classes=cfg["DATA"].NUM_CLASSES,
proposal_provider=proposal_provider,
provide_targets=True,
proposal_iou_threshold=cfg.BBOX_THRESH,
normalize_means=None
if not cfg.BBOX_NORMALIZE_TARGETS else cfg.BBOX_NORMALIZE_MEANS,
normalize_stds=None
if not cfg.BBOX_NORMALIZE_TARGETS else cfg.BBOX_NORMALIZE_STDS)
# define mapping from reader streams to network inputs
input_map = {
od_minibatch_source.image_si: image_input,
od_minibatch_source.proposals_si: roi_proposals,
od_minibatch_source.label_targets_si: label_targets,
od_minibatch_source.bbox_targets_si: bbox_targets,
od_minibatch_source.bbiw_si: bbox_inside_weights
}
progress_printer = ProgressPrinter(tag='Training',
num_epochs=epochs_to_train,
gen_heartbeat=True)
for epoch in range(epochs_to_train): # loop over epochs
sample_count = 0
while sample_count < cfg[
"DATA"].NUM_TRAIN_IMAGES: # loop over minibatches in the epoch
data = od_minibatch_source.next_minibatch(min(
cfg.MB_SIZE, cfg["DATA"].NUM_TRAIN_IMAGES - sample_count),
input_map=input_map)
trainer.train_minibatch(data) # update model with it
sample_count += trainer.previous_minibatch_sample_count # count samples processed so far
progress_printer.update_with_trainer(
trainer, with_metric=True) # log progress
if sample_count % 100 == 0:
continue
#print("Processed {} samples".format(sample_count))
progress_printer.epoch_summary(with_metric=True)
eval_model = create_fast_rcnn_eval_model(loss, image_input,
roi_proposals, cfg)
eval_model.save(cfg['MODEL_PATH'])
return eval_model
model_path = os.path.join(globalvars['output_path'], "faster_rcnn_eval_{}_{}.model"
.format(cfg["CNTK"].BASE_MODEL, "e2e" if globalvars['train_e2e'] else "4stage"))
# Train only if no model exists yet
if os.path.exists(model_path) and cfg["CNTK"].MAKE_MODE:
print("Loading existing model from %s" % model_path)
eval_model = load_model(model_path)
else:
if globalvars['train_e2e']:
eval_model = train_faster_rcnn_e2e(base_model_file, debug_output=cfg["CNTK"].DEBUG_OUTPUT)
else:
eval_model = train_faster_rcnn_alternating(base_model_file, debug_output=cfg["CNTK"].DEBUG_OUTPUT)
eval_model.save(model_path)
if cfg["CNTK"].DEBUG_OUTPUT:
plot(eval_model, os.path.join(globalvars['output_path'], "graph_frcn_eval_{}_{}.{}"
.format(cfg["CNTK"].BASE_MODEL, "e2e" if globalvars['train_e2e'] else "4stage", cfg["CNTK"].GRAPH_TYPE)))
print("Stored eval model at %s" % model_path)
# Compute mean average precision on test set
eval_faster_rcnn_mAP(eval_model)
# Plot results on test set
if cfg["CNTK"].VISUALIZE_RESULTS:
from plot_helpers import eval_and_plot_faster_rcnn
num_eval = min(num_test_images, 100)
img_shape = (num_channels, image_height, image_width)
results_folder = os.path.join(globalvars['output_path'], cfg["CNTK"].DATASET)
eval_and_plot_faster_rcnn(eval_model, num_eval, globalvars['test_map_file'], img_shape,
results_folder, feature_node_name, globalvars['classes'],
drawUnregressedRois=cfg["CNTK"].DRAW_UNREGRESSED_ROIS,
def train_faster_rcnn_alternating(cfg):
'''
4-Step Alternating Training scheme from the Faster R-CNN paper:
# Create initial network, only rpn, without detection network
# --> train only the rpn (and conv3_1 and up for VGG16)
# buffer region proposals from rpn
# Create full network, initialize conv layers with imagenet, use buffered proposals
# --> train only detection network (and conv3_1 and up for VGG16)
# Keep conv weights from detection network and fix them
# --> train only rpn
# buffer region proposals from rpn
# Keep conv and rpn weights from step 3 and fix them
# --> train only detection network
'''
# setting pre- and post-nms top N to training values since buffered proposals are used for further training
test_pre = cfg["TEST"].RPN_PRE_NMS_TOP_N
test_post = cfg["TEST"].RPN_POST_NMS_TOP_N
cfg["TEST"].RPN_PRE_NMS_TOP_N = cfg["TRAIN"].RPN_PRE_NMS_TOP_N
cfg["TEST"].RPN_POST_NMS_TOP_N = cfg["TRAIN"].RPN_POST_NMS_TOP_N
# Learning parameters
rpn_lr_factor = cfg["MODEL"].RPN_LR_FACTOR
rpn_lr_per_sample_scaled = [x * rpn_lr_factor for x in cfg["CNTK"].RPN_LR_PER_SAMPLE]
frcn_lr_factor = cfg["MODEL"].FRCN_LR_FACTOR
frcn_lr_per_sample_scaled = [x * frcn_lr_factor for x in cfg["CNTK"].FRCN_LR_PER_SAMPLE]
l2_reg_weight = cfg["CNTK"].L2_REG_WEIGHT
mm_schedule = momentum_schedule(cfg["CNTK"].MOMENTUM_PER_MB)
rpn_epochs = cfg["CNTK"].RPN_EPOCHS
frcn_epochs = cfg["CNTK"].FRCN_EPOCHS
feature_node_name = cfg["MODEL"].FEATURE_NODE_NAME
last_conv_node_name = cfg["MODEL"].LAST_CONV_NODE_NAME
print("Using base model: {}".format(cfg["MODEL"].BASE_MODEL))
print("rpn_lr_per_sample: {}".format(rpn_lr_per_sample_scaled))
print("frcn_lr_per_sample: {}".format(frcn_lr_per_sample_scaled))
debug_output=cfg["CNTK"].DEBUG_OUTPUT
if debug_output:
print("Storing graphs and models to %s." % cfg.OUTPUT_PATH)
# Input variables denoting features, labeled ground truth rois (as 5-tuples per roi) and image dimensions
image_input = input_variable(shape=(cfg.NUM_CHANNELS, cfg.IMAGE_HEIGHT, cfg.IMAGE_WIDTH),
dynamic_axes=[Axis.default_batch_axis()],
name=feature_node_name)
feat_norm = image_input - Constant([[[v]] for v in cfg["MODEL"].IMG_PAD_COLOR])
roi_input = input_variable((cfg.INPUT_ROIS_PER_IMAGE, 5), dynamic_axes=[Axis.default_batch_axis()])
scaled_gt_boxes = alias(roi_input, name='roi_input')
dims_input = input_variable((6), dynamic_axes=[Axis.default_batch_axis()])
dims_node = alias(dims_input, name='dims_input')
rpn_rois_input = input_variable((cfg["TRAIN"].RPN_POST_NMS_TOP_N, 4), dynamic_axes=[Axis.default_batch_axis()])
rpn_rois_buf = alias(rpn_rois_input, name='rpn_rois')
# base image classification model (e.g. VGG16 or AlexNet)
base_model = load_model(cfg['BASE_MODEL_PATH'])
print("stage 1a - rpn")
if True:
# Create initial network, only rpn, without detection network
# initial weights train?
# conv: base_model only conv3_1 and up
# rpn: init new yes
# frcn: - -
# conv layers
conv_layers = clone_conv_layers(base_model, cfg)
conv_out = conv_layers(feat_norm)
# RPN and losses
rpn_rois, rpn_losses = create_rpn(conv_out, scaled_gt_boxes, dims_node, cfg)
stage1_rpn_network = combine([rpn_rois, rpn_losses])
# train
if debug_output: plot(stage1_rpn_network, os.path.join(cfg.OUTPUT_PATH, "graph_frcn_train_stage1a_rpn." + cfg["CNTK"].GRAPH_TYPE))
train_model(image_input, roi_input, dims_input, rpn_losses, rpn_losses,
rpn_lr_per_sample_scaled, mm_schedule, l2_reg_weight, rpn_epochs, cfg)
print("stage 1a - buffering rpn proposals")
buffered_proposals_s1 = compute_rpn_proposals(stage1_rpn_network, image_input, roi_input, dims_input, cfg)
print("stage 1b - frcn")
if True:
# Create full network, initialize conv layers with imagenet, fix rpn weights
# initial weights train?
# conv: base_model only conv3_1 and up
# rpn: stage1a rpn model no --> use buffered proposals
# frcn: base_model + new yes
# conv_layers
conv_layers = clone_conv_layers(base_model, cfg)
conv_out = conv_layers(feat_norm)
# use buffered proposals in target layer
rois, label_targets, bbox_targets, bbox_inside_weights = \
create_proposal_target_layer(rpn_rois_buf, scaled_gt_boxes, cfg)
# Fast RCNN and losses
fc_layers = clone_model(base_model, [cfg["MODEL"].POOL_NODE_NAME], [cfg["MODEL"].LAST_HIDDEN_NODE_NAME], CloneMethod.clone)
cls_score, bbox_pred = create_fast_rcnn_predictor(conv_out, rois, fc_layers, cfg)
detection_losses = create_detection_losses(cls_score, label_targets, bbox_pred, rois, bbox_targets, bbox_inside_weights, cfg)
pred_error = classification_error(cls_score, label_targets, axis=1, name="pred_error")
stage1_frcn_network = combine([rois, cls_score, bbox_pred, detection_losses, pred_error])
# train
if debug_output: plot(stage1_frcn_network, os.path.join(cfg.OUTPUT_PATH, "graph_frcn_train_stage1b_frcn." + cfg["CNTK"].GRAPH_TYPE))
train_model(image_input, roi_input, dims_input, detection_losses, pred_error,
frcn_lr_per_sample_scaled, mm_schedule, l2_reg_weight, frcn_epochs, cfg,
rpn_rois_input=rpn_rois_input, buffered_rpn_proposals=buffered_proposals_s1)
buffered_proposals_s1 = None
print("stage 2a - rpn")
if True:
# Keep conv weights from detection network and fix them
# initial weights train?
# conv: stage1b frcn model no
# rpn: stage1a rpn model yes
# frcn: - -
# conv_layers
conv_layers = clone_model(stage1_frcn_network, [feature_node_name], [last_conv_node_name], CloneMethod.freeze)
conv_out = conv_layers(image_input)
# RPN and losses
rpn = clone_model(stage1_rpn_network, [last_conv_node_name, "roi_input", "dims_input"], ["rpn_rois", "rpn_losses"], CloneMethod.clone)
rpn_net = rpn(conv_out, dims_node, scaled_gt_boxes)
rpn_rois = rpn_net.outputs[0]
rpn_losses = rpn_net.outputs[1]
stage2_rpn_network = combine([rpn_rois, rpn_losses])
# train
if debug_output: plot(stage2_rpn_network, os.path.join(cfg.OUTPUT_PATH, "graph_frcn_train_stage2a_rpn." + cfg["CNTK"].GRAPH_TYPE))
train_model(image_input, roi_input, dims_input, rpn_losses, rpn_losses,
rpn_lr_per_sample_scaled, mm_schedule, l2_reg_weight, rpn_epochs, cfg)
print("stage 2a - buffering rpn proposals")
buffered_proposals_s2 = compute_rpn_proposals(stage2_rpn_network, image_input, roi_input, dims_input, cfg)
print("stage 2b - frcn")
if True:
# Keep conv and rpn weights from step 3 and fix them
# initial weights train?
# conv: stage2a rpn model no
# rpn: stage2a rpn model no --> use buffered proposals
# frcn: stage1b frcn model yes -
# conv_layers
conv_layers = clone_model(stage2_rpn_network, [feature_node_name], [last_conv_node_name], CloneMethod.freeze)
conv_out = conv_layers(image_input)
# Fast RCNN and losses
frcn = clone_model(stage1_frcn_network, [last_conv_node_name, "rpn_rois", "roi_input"],
["cls_score", "bbox_regr", "rpn_target_rois", "detection_losses", "pred_error"], CloneMethod.clone)
stage2_frcn_network = frcn(conv_out, rpn_rois_buf, scaled_gt_boxes)
detection_losses = stage2_frcn_network.outputs[3]
pred_error = stage2_frcn_network.outputs[4]
# train
if debug_output: plot(stage2_frcn_network, os.path.join(cfg.OUTPUT_PATH, "graph_frcn_train_stage2b_frcn." + cfg["CNTK"].GRAPH_TYPE))
train_model(image_input, roi_input, dims_input, detection_losses, pred_error,
frcn_lr_per_sample_scaled, mm_schedule, l2_reg_weight, frcn_epochs, cfg,
rpn_rois_input=rpn_rois_input, buffered_rpn_proposals=buffered_proposals_s2)
buffered_proposals_s2 = None
# resetting config values to original test values
cfg["TEST"].RPN_PRE_NMS_TOP_N = test_pre
cfg["TEST"].RPN_POST_NMS_TOP_N = test_post
return create_faster_rcnn_eval_model(stage2_frcn_network, image_input, dims_input, cfg, rpn_model=stage2_rpn_network)
def train_faster_rcnn_alternating(base_model_file_name, debug_output=False):
'''
4-Step Alternating Training scheme from the Faster R-CNN paper:
# Create initial network, only rpn, without detection network
# --> train only the rpn (and conv3_1 and up for VGG16)
# buffer region proposals from rpn
# Create full network, initialize conv layers with imagenet, use buffered proposals
# --> train only detection network (and conv3_1 and up for VGG16)
# Keep conv weights from detection network and fix them
# --> train only rpn
# buffer region proposals from rpn
# Keep conv and rpn weights from step 3 and fix them
# --> train only detection network
'''
# Learning parameters
rpn_lr_factor = globalvars['rpn_lr_factor']
rpn_lr_per_sample_scaled = [x * rpn_lr_factor for x in cfg["CNTK"].RPN_LR_PER_SAMPLE]
frcn_lr_factor = globalvars['frcn_lr_factor']
frcn_lr_per_sample_scaled = [x * frcn_lr_factor for x in cfg["CNTK"].FRCN_LR_PER_SAMPLE]
l2_reg_weight = cfg["CNTK"].L2_REG_WEIGHT
mm_schedule = momentum_schedule(globalvars['momentum_per_mb'])
rpn_epochs = globalvars['rpn_epochs']
frcn_epochs = globalvars['frcn_epochs']
print("Using base model: {}".format(cfg["CNTK"].BASE_MODEL))
print("rpn_lr_per_sample: {}".format(rpn_lr_per_sample_scaled))
print("frcn_lr_per_sample: {}".format(frcn_lr_per_sample_scaled))
if debug_output:
print("Storing graphs and models to %s." % globalvars['output_path'])
# Input variables denoting features, labeled ground truth rois (as 5-tuples per roi) and image dimensions
image_input = input_variable((num_channels, image_height, image_width), dynamic_axes=[Axis.default_batch_axis()],
name=feature_node_name)
feat_norm = image_input - normalization_const
roi_input = input_variable((cfg["CNTK"].INPUT_ROIS_PER_IMAGE, 5), dynamic_axes=[Axis.default_batch_axis()])
scaled_gt_boxes = alias(roi_input, name='roi_input')
dims_input = input_variable((6), dynamic_axes=[Axis.default_batch_axis()])
dims_node = alias(dims_input, name='dims_input')
rpn_rois_input = input_variable((cfg["TRAIN"].RPN_POST_NMS_TOP_N, 4), dynamic_axes=[Axis.default_batch_axis()])
rpn_rois_buf = alias(rpn_rois_input, name='rpn_rois')
# base image classification model (e.g. VGG16 or AlexNet)
base_model = load_model(base_model_file_name)
print("stage 1a - rpn")
if True:
# Create initial network, only rpn, without detection network
# initial weights train?
# conv: base_model only conv3_1 and up
# rpn: init new yes
# frcn: - -
# conv layers
conv_layers = clone_conv_layers(base_model)
conv_out = conv_layers(feat_norm)
# RPN and losses
rpn_rois, rpn_losses = create_rpn(conv_out, scaled_gt_boxes, dims_node, proposal_layer_param_string=cfg["CNTK"].PROPOSAL_LAYER_PARAMS)
stage1_rpn_network = combine([rpn_rois, rpn_losses])
# train
if debug_output: plot(stage1_rpn_network, os.path.join(globalvars['output_path'], "graph_frcn_train_stage1a_rpn." + cfg["CNTK"].GRAPH_TYPE))
train_model(image_input, roi_input, dims_input, rpn_losses, rpn_losses,
rpn_lr_per_sample_scaled, mm_schedule, l2_reg_weight, epochs_to_train=rpn_epochs)
print("stage 1a - buffering rpn proposals")
buffered_proposals_s1 = compute_rpn_proposals(stage1_rpn_network, image_input, roi_input, dims_input)
print("stage 1b - frcn")
if True:
# Create full network, initialize conv layers with imagenet, fix rpn weights
# initial weights train?
# conv: base_model only conv3_1 and up
# rpn: stage1a rpn model no --> use buffered proposals
# frcn: base_model + new yes
# conv_layers
conv_layers = clone_conv_layers(base_model)
conv_out = conv_layers(feat_norm)
# use buffered proposals in target layer
rois, label_targets, bbox_targets, bbox_inside_weights = \
create_proposal_target_layer(rpn_rois_buf, scaled_gt_boxes, num_classes=globalvars['num_classes'])
# Fast RCNN and losses
fc_layers = clone_model(base_model, [pool_node_name], [last_hidden_node_name], CloneMethod.clone)
cls_score, bbox_pred = create_fast_rcnn_predictor(conv_out, rois, fc_layers)
detection_losses = create_detection_losses(cls_score, label_targets, rois, bbox_pred, bbox_targets, bbox_inside_weights)
pred_error = classification_error(cls_score, label_targets, axis=1, name="pred_error")
stage1_frcn_network = combine([rois, cls_score, bbox_pred, detection_losses, pred_error])
# train
if debug_output: plot(stage1_frcn_network, os.path.join(globalvars['output_path'], "graph_frcn_train_stage1b_frcn." + cfg["CNTK"].GRAPH_TYPE))
train_model(image_input, roi_input, dims_input, detection_losses, pred_error,
frcn_lr_per_sample_scaled, mm_schedule, l2_reg_weight, epochs_to_train=frcn_epochs,
rpn_rois_input=rpn_rois_input, buffered_rpn_proposals=buffered_proposals_s1)
buffered_proposals_s1 = None
print("stage 2a - rpn")
if True:
# Keep conv weights from detection network and fix them
# initial weights train?
# conv: stage1b frcn model no
# rpn: stage1a rpn model yes
# frcn: - -
# conv_layers
conv_layers = clone_model(stage1_frcn_network, [feature_node_name], [last_conv_node_name], CloneMethod.freeze)
conv_out = conv_layers(image_input)
# RPN and losses
rpn = clone_model(stage1_rpn_network, [last_conv_node_name, "roi_input", "dims_input"], ["rpn_rois", "rpn_losses"], CloneMethod.clone)
rpn_net = rpn(conv_out, dims_node, scaled_gt_boxes)
rpn_rois = rpn_net.outputs[0]
rpn_losses = rpn_net.outputs[1]
stage2_rpn_network = combine([rpn_rois, rpn_losses])
# train
if debug_output: plot(stage2_rpn_network, os.path.join(globalvars['output_path'], "graph_frcn_train_stage2a_rpn." + cfg["CNTK"].GRAPH_TYPE))
train_model(image_input, roi_input, dims_input, rpn_losses, rpn_losses,
rpn_lr_per_sample_scaled, mm_schedule, l2_reg_weight, epochs_to_train=rpn_epochs)
print("stage 2a - buffering rpn proposals")
buffered_proposals_s2 = compute_rpn_proposals(stage2_rpn_network, image_input, roi_input, dims_input)
print("stage 2b - frcn")
if True:
# Keep conv and rpn weights from step 3 and fix them
# initial weights train?
# conv: stage2a rpn model no
# rpn: stage2a rpn model no --> use buffered proposals
# frcn: stage1b frcn model yes -
# conv_layers
conv_layers = clone_model(stage2_rpn_network, [feature_node_name], [last_conv_node_name], CloneMethod.freeze)
conv_out = conv_layers(image_input)
# Fast RCNN and losses
frcn = clone_model(stage1_frcn_network, [last_conv_node_name, "rpn_rois", "roi_input"],
["cls_score", "bbox_regr", "rpn_target_rois", "detection_losses", "pred_error"], CloneMethod.clone)
stage2_frcn_network = frcn(conv_out, rpn_rois_buf, scaled_gt_boxes)
detection_losses = stage2_frcn_network.outputs[3]
pred_error = stage2_frcn_network.outputs[4]
# train
if debug_output: plot(stage2_frcn_network, os.path.join(globalvars['output_path'], "graph_frcn_train_stage2b_frcn." + cfg["CNTK"].GRAPH_TYPE))
train_model(image_input, roi_input, dims_input, detection_losses, pred_error,
frcn_lr_per_sample_scaled, mm_schedule, l2_reg_weight, epochs_to_train=frcn_epochs,
rpn_rois_input=rpn_rois_input, buffered_rpn_proposals=buffered_proposals_s2)
buffered_proposals_s2 = None
return create_eval_model(stage2_frcn_network, image_input, dims_input, rpn_model=stage2_rpn_network)
def train_faster_rcnn_alternating(base_model_file_name, debug_output=False):
'''
4-Step Alternating Training scheme from the Faster R-CNN paper:
# Create initial network, only rpn, without detection network
# --> train only the rpn (and conv3_1 and up for VGG16)
# buffer region proposals from rpn
# Create full network, initialize conv layers with imagenet, use buffered proposals
# --> train only detection network (and conv3_1 and up for VGG16)
# Keep conv weights from detection network and fix them
# --> train only rpn
# buffer region proposals from rpn
# Keep conv and rpn weights from step 3 and fix them
# --> train only detection network
'''
# Learning parameters
rpn_lr_factor = globalvars['rpn_lr_factor']
rpn_lr_per_sample_scaled = [
x * rpn_lr_factor for x in cfg["CNTK"].RPN_LR_PER_SAMPLE
]
frcn_lr_factor = globalvars['frcn_lr_factor']
frcn_lr_per_sample_scaled = [
x * frcn_lr_factor for x in cfg["CNTK"].FRCN_LR_PER_SAMPLE
]
l2_reg_weight = cfg["CNTK"].L2_REG_WEIGHT
mm_schedule = momentum_schedule(globalvars['momentum_per_mb'])
rpn_epochs = globalvars['rpn_epochs']
frcn_epochs = globalvars['frcn_epochs']
print("Using base model: {}".format(cfg["CNTK"].BASE_MODEL))
print("rpn_lr_per_sample: {}".format(rpn_lr_per_sample_scaled))
print("frcn_lr_per_sample: {}".format(frcn_lr_per_sample_scaled))
if debug_output:
print("Storing graphs and models to %s." % globalvars['output_path'])
# Input variables denoting features, labeled ground truth rois (as 5-tuples per roi) and image dimensions
image_input = input_variable((num_channels, image_height, image_width),
dynamic_axes=[Axis.default_batch_axis()],
name=feature_node_name)
feat_norm = image_input - normalization_const
roi_input = input_variable((cfg["CNTK"].INPUT_ROIS_PER_IMAGE, 5),
dynamic_axes=[Axis.default_batch_axis()])
scaled_gt_boxes = alias(roi_input, name='roi_input')
dims_input = input_variable((6), dynamic_axes=[Axis.default_batch_axis()])
dims_node = alias(dims_input, name='dims_input')
rpn_rois_input = input_variable((cfg["TRAIN"].RPN_POST_NMS_TOP_N, 4),
dynamic_axes=[Axis.default_batch_axis()])
rpn_rois_buf = alias(rpn_rois_input, name='rpn_rois')
# base image classification model (e.g. VGG16 or AlexNet)
base_model = load_model(base_model_file_name)
print("stage 1a - rpn")
if True:
# Create initial network, only rpn, without detection network
# initial weights train?
# conv: base_model only conv3_1 and up
# rpn: init new yes
# frcn: - -
# conv layers
conv_layers = clone_conv_layers(base_model)
conv_out = conv_layers(feat_norm)
# RPN and losses
rpn_rois, rpn_losses = create_rpn(
conv_out,
scaled_gt_boxes,
dims_node,
proposal_layer_param_string=cfg["CNTK"].PROPOSAL_LAYER_PARAMS)
stage1_rpn_network = combine([rpn_rois, rpn_losses])
# train
if debug_output:
plot(
stage1_rpn_network,
os.path.join(
globalvars['output_path'],
"graph_frcn_train_stage1a_rpn." + cfg["CNTK"].GRAPH_TYPE))
train_model(image_input,
roi_input,
dims_input,
rpn_losses,
rpn_losses,
rpn_lr_per_sample_scaled,
mm_schedule,
l2_reg_weight,
epochs_to_train=rpn_epochs)
print("stage 1a - buffering rpn proposals")
buffered_proposals_s1 = compute_rpn_proposals(stage1_rpn_network,
image_input, roi_input,
dims_input)
print("stage 1b - frcn")
if True:
# Create full network, initialize conv layers with imagenet, fix rpn weights
# initial weights train?
# conv: base_model only conv3_1 and up
# rpn: stage1a rpn model no --> use buffered proposals
# frcn: base_model + new yes
# conv_layers
conv_layers = clone_conv_layers(base_model)
conv_out = conv_layers(feat_norm)
# use buffered proposals in target layer
rois, label_targets, bbox_targets, bbox_inside_weights = \
create_proposal_target_layer(rpn_rois_buf, scaled_gt_boxes, num_classes=globalvars['num_classes'])
# Fast RCNN and losses
fc_layers = clone_model(base_model, [pool_node_name],
[last_hidden_node_name], CloneMethod.clone)
cls_score, bbox_pred = create_fast_rcnn_predictor(
conv_out, rois, fc_layers)
detection_losses = create_detection_losses(cls_score, label_targets,
rois, bbox_pred,
bbox_targets,
bbox_inside_weights)
pred_error = classification_error(cls_score,
label_targets,
axis=1,
name="pred_error")
stage1_frcn_network = combine(
[rois, cls_score, bbox_pred, detection_losses, pred_error])
# train
if debug_output:
plot(
stage1_frcn_network,
os.path.join(
globalvars['output_path'],
"graph_frcn_train_stage1b_frcn." + cfg["CNTK"].GRAPH_TYPE))
train_model(image_input,
roi_input,
dims_input,
detection_losses,
pred_error,
frcn_lr_per_sample_scaled,
mm_schedule,
l2_reg_weight,
epochs_to_train=frcn_epochs,
rpn_rois_input=rpn_rois_input,
buffered_rpn_proposals=buffered_proposals_s1)
buffered_proposals_s1 = None
print("stage 2a - rpn")
if True:
# Keep conv weights from detection network and fix them
# initial weights train?
# conv: stage1b frcn model no
# rpn: stage1a rpn model yes
# frcn: - -
# conv_layers
conv_layers = clone_model(stage1_frcn_network, [feature_node_name],
[last_conv_node_name], CloneMethod.freeze)
conv_out = conv_layers(image_input)
# RPN and losses
rpn = clone_model(stage1_rpn_network,
[last_conv_node_name, "roi_input", "dims_input"],
["rpn_rois", "rpn_losses"], CloneMethod.clone)
rpn_net = rpn(conv_out, dims_node, scaled_gt_boxes)
rpn_rois = rpn_net.outputs[0]
rpn_losses = rpn_net.outputs[1]
stage2_rpn_network = combine([rpn_rois, rpn_losses])
# train
if debug_output:
plot(
stage2_rpn_network,
os.path.join(
globalvars['output_path'],
"graph_frcn_train_stage2a_rpn." + cfg["CNTK"].GRAPH_TYPE))
train_model(image_input,
roi_input,
dims_input,
rpn_losses,
rpn_losses,
rpn_lr_per_sample_scaled,
mm_schedule,
l2_reg_weight,
epochs_to_train=rpn_epochs)
print("stage 2a - buffering rpn proposals")
buffered_proposals_s2 = compute_rpn_proposals(stage2_rpn_network,
image_input, roi_input,
dims_input)
print("stage 2b - frcn")
if True:
# Keep conv and rpn weights from step 3 and fix them
# initial weights train?
# conv: stage2a rpn model no
# rpn: stage2a rpn model no --> use buffered proposals
# frcn: stage1b frcn model yes -
# conv_layers
conv_layers = clone_model(stage2_rpn_network, [feature_node_name],
[last_conv_node_name], CloneMethod.freeze)
conv_out = conv_layers(image_input)
# Fast RCNN and losses
frcn = clone_model(stage1_frcn_network,
[last_conv_node_name, "rpn_rois", "roi_input"], [
"cls_score", "bbox_regr", "rpn_target_rois",
"detection_losses", "pred_error"
], CloneMethod.clone)
stage2_frcn_network = frcn(conv_out, rpn_rois_buf, scaled_gt_boxes)
detection_losses = stage2_frcn_network.outputs[3]
pred_error = stage2_frcn_network.outputs[4]
# train
if debug_output:
plot(
stage2_frcn_network,
os.path.join(
globalvars['output_path'],
"graph_frcn_train_stage2b_frcn." + cfg["CNTK"].GRAPH_TYPE))
train_model(image_input,
roi_input,
dims_input,
detection_losses,
pred_error,
frcn_lr_per_sample_scaled,
mm_schedule,
l2_reg_weight,
epochs_to_train=frcn_epochs,
rpn_rois_input=rpn_rois_input,
buffered_rpn_proposals=buffered_proposals_s2)
buffered_proposals_s2 = None
return create_eval_model(stage2_frcn_network,
image_input,
dims_input,
rpn_model=stage2_rpn_network)
print("Loading existing model from %s" % model_path)
eval_model = load_model(model_path)
else:
if globalvars['train_e2e']:
eval_model = train_faster_rcnn_e2e(
base_model_file, debug_output=cfg["CNTK"].DEBUG_OUTPUT)
else:
eval_model = train_faster_rcnn_alternating(
base_model_file, debug_output=cfg["CNTK"].DEBUG_OUTPUT)
eval_model.save(model_path)
if cfg["CNTK"].DEBUG_OUTPUT:
plot(
eval_model,
os.path.join(
globalvars['output_path'],
"graph_frcn_eval_{}_{}.{}".format(
cfg["CNTK"].BASE_MODEL,
"e2e" if globalvars['train_e2e'] else "4stage",
cfg["CNTK"].GRAPH_TYPE)))
print("Stored eval model at %s" % model_path)
# Compute mean average precision on test set
eval_faster_rcnn_mAP(eval_model)
# Plot results on test set
if cfg["CNTK"].VISUALIZE_RESULTS:
from plot_helpers import eval_and_plot_faster_rcnn
num_eval = min(num_test_images, 100)
img_shape = (num_channels, image_height, image_width)
results_folder = os.path.join(globalvars['output_path'],
def train_fast_rcnn(debug_output=False, model_path=model_file):
if debug_output:
print("Storing graphs and intermediate models to %s." % os.path.join(abs_path, "Output"))
# Create the minibatch source
minibatch_source = create_mb_source(image_height, image_width, num_channels,
num_classes, num_rois, base_path, "train")
# Input variables denoting features, rois and label data
image_input = C.input_variable((num_channels, image_height, image_width))
roi_input = C.input_variable((num_rois, 4))
label_input = C.input_variable((num_rois, num_classes))
# define mapping from reader streams to network inputs
input_map = {
image_input: minibatch_source.streams.features,
roi_input: minibatch_source.streams.rois,
label_input: minibatch_source.streams.roiLabels
}
# Instantiate the Fast R-CNN prediction model and loss function
frcn_output = frcn_predictor(image_input, roi_input, num_classes, model_path)
ce = cross_entropy_with_softmax(frcn_output, label_input, axis=1)
pe = classification_error(frcn_output, label_input, axis=1)
if debug_output:
plot(frcn_output, os.path.join(abs_path, "Output", "graph_frcn.png"))
# Set learning parameters
l2_reg_weight = 0.0005
lr_per_sample = [0.00001] * 10 + [0.000001] * 5 + [0.0000001]
lr_schedule = learning_parameter_schedule_per_sample(lr_per_sample)
mm_schedule = momentum_schedule_per_sample(momentum_per_sample)
# Instantiate the trainer object as default
learner = momentum_sgd(frcn_output.parameters, lr_schedule, mm_schedule, l2_regularization_weight=l2_reg_weight)
# Preparation for distributed learning, which is compatible for normal learner
learner = distributed.data_parallel_distributed_learner(
learner = learner,
num_quantization_bits = num_quantization_bits, # non-quantized gradient accumulation
distributed_after = warm_up) # no warm start as default
progress_printer = ProgressPrinter(tag='Training', num_epochs=max_epochs, rank=distributed.Communicator.rank())
trainer = Trainer(frcn_output, (ce, pe), learner, progress_printer)
# Get minibatches of images and perform model training
print("Training Fast R-CNN model for %s epochs." % max_epochs)
log_number_of_parameters(frcn_output)
for epoch in range(max_epochs): # loop over epochs
sample_count = 0
while sample_count < epoch_size: # loop over minibatches in the epoch
data = minibatch_source.next_minibatch(min(mb_size * C.Communicator.num_workers(), epoch_size-sample_count),
input_map=input_map,
num_data_partitions=C.Communicator.num_workers(),
partition_index=C.Communicator.rank())
trainer.train_minibatch(data) # update model with it
sample_count += trainer.previous_minibatch_sample_count # count samples processed so far
trainer.summarize_training_progress()
if debug_output:
frcn_output.save(os.path.join(abs_path, "Output", "frcn_py_%s.model" % (epoch+1)))
if distributed_flg:
distributed.Communicator.finalize()
return frcn_output
def train_fast_rcnn(cfg):
# Train only if no model exists yet
model_path = cfg['MODEL_PATH']
if os.path.exists(model_path) and cfg["CNTK"].MAKE_MODE:
print("Loading existing model from %s" % model_path)
return load_model(model_path)
else:
# Input variables denoting features and labeled ground truth rois (as 5-tuples per roi)
image_input = input_variable(shape=(cfg.NUM_CHANNELS, cfg.IMAGE_HEIGHT, cfg.IMAGE_WIDTH),
dynamic_axes=[Axis.default_batch_axis()],
name=cfg["MODEL"].FEATURE_NODE_NAME)
roi_proposals = input_variable((cfg.NUM_ROI_PROPOSALS, 4), dynamic_axes=[Axis.default_batch_axis()], name = "roi_proposals")
label_targets = input_variable((cfg.NUM_ROI_PROPOSALS, cfg["DATA"].NUM_CLASSES), dynamic_axes=[Axis.default_batch_axis()])
bbox_targets = input_variable((cfg.NUM_ROI_PROPOSALS, 4*cfg["DATA"].NUM_CLASSES), dynamic_axes=[Axis.default_batch_axis()])
bbox_inside_weights = input_variable((cfg.NUM_ROI_PROPOSALS, 4*cfg["DATA"].NUM_CLASSES), dynamic_axes=[Axis.default_batch_axis()])
# Instantiate the Fast R-CNN prediction model and loss function
loss, pred_error = create_fast_rcnn_model(image_input, roi_proposals, label_targets, bbox_targets, bbox_inside_weights, cfg)
if isinstance(loss, cntk.Variable):
loss = combine([loss])
if cfg["CNTK"].DEBUG_OUTPUT:
print("Storing graphs and models to %s." % cfg.OUTPUT_PATH)
plot(loss, os.path.join(cfg.OUTPUT_PATH, "graph_frcn_train." + cfg["CNTK"].GRAPH_TYPE))
# Set learning parameters
lr_factor = cfg["CNTK"].LR_FACTOR
lr_per_sample_scaled = [x * lr_factor for x in cfg["CNTK"].LR_PER_SAMPLE]
mm_schedule = momentum_schedule(cfg["CNTK"].MOMENTUM_PER_MB)
l2_reg_weight = cfg["CNTK"].L2_REG_WEIGHT
epochs_to_train = cfg["CNTK"].MAX_EPOCHS
print("Using base model: {}".format(cfg["MODEL"].BASE_MODEL))
print("lr_per_sample: {}".format(lr_per_sample_scaled))
# --- train ---
# Instantiate the learners and the trainer object
params = loss.parameters
biases = [p for p in params if '.b' in p.name or 'b' == p.name]
others = [p for p in params if not p in biases]
bias_lr_mult = cfg["CNTK"].BIAS_LR_MULT
lr_schedule = learning_rate_schedule(lr_per_sample_scaled, unit=UnitType.sample)
learner = momentum_sgd(others, lr_schedule, mm_schedule, l2_regularization_weight=l2_reg_weight, unit_gain=False, use_mean_gradient=True)
bias_lr_per_sample = [v * bias_lr_mult for v in cfg["CNTK"].LR_PER_SAMPLE]
bias_lr_schedule = learning_rate_schedule(bias_lr_per_sample, unit=UnitType.sample)
bias_learner = momentum_sgd(biases, bias_lr_schedule, mm_schedule, l2_regularization_weight=l2_reg_weight, unit_gain=False, use_mean_gradient=True)
trainer = Trainer(None, (loss, pred_error), [learner, bias_learner])
# Get minibatches of images and perform model training
print("Training model for %s epochs." % epochs_to_train)
log_number_of_parameters(loss)
# Create the minibatch source
if cfg.USE_PRECOMPUTED_PROPOSALS:
proposal_provider = ProposalProvider.fromfile(cfg["DATA"].TRAIN_PRECOMPUTED_PROPOSALS_FILE, cfg.NUM_ROI_PROPOSALS)
else:
proposal_provider = ProposalProvider.fromconfig(cfg)
od_minibatch_source = ObjectDetectionMinibatchSource(
cfg["DATA"].TRAIN_MAP_FILE, cfg["DATA"].TRAIN_ROI_FILE,
max_annotations_per_image=cfg.INPUT_ROIS_PER_IMAGE,
pad_width=cfg.IMAGE_WIDTH,
pad_height=cfg.IMAGE_HEIGHT,
pad_value=cfg["MODEL"].IMG_PAD_COLOR,
randomize=True,
use_flipping=cfg["TRAIN"].USE_FLIPPED,
max_images=cfg["DATA"].NUM_TRAIN_IMAGES,
num_classes=cfg["DATA"].NUM_CLASSES,
proposal_provider=proposal_provider,
provide_targets=True,
proposal_iou_threshold = cfg.BBOX_THRESH,
normalize_means = None if not cfg.BBOX_NORMALIZE_TARGETS else cfg.BBOX_NORMALIZE_MEANS,
normalize_stds = None if not cfg.BBOX_NORMALIZE_TARGETS else cfg.BBOX_NORMALIZE_STDS)
# define mapping from reader streams to network inputs
input_map = {
od_minibatch_source.image_si: image_input,
od_minibatch_source.proposals_si: roi_proposals,
od_minibatch_source.label_targets_si: label_targets,
od_minibatch_source.bbox_targets_si: bbox_targets,
od_minibatch_source.bbiw_si: bbox_inside_weights
}
progress_printer = ProgressPrinter(tag='Training', num_epochs=epochs_to_train, gen_heartbeat=True)
for epoch in range(epochs_to_train): # loop over epochs
sample_count = 0
while sample_count < cfg["DATA"].NUM_TRAIN_IMAGES: # loop over minibatches in the epoch
data = od_minibatch_source.next_minibatch(min(cfg.MB_SIZE, cfg["DATA"].NUM_TRAIN_IMAGES - sample_count), input_map=input_map)
trainer.train_minibatch(data) # update model with it
sample_count += trainer.previous_minibatch_sample_count # count samples processed so far
progress_printer.update_with_trainer(trainer, with_metric=True) # log progress
if sample_count % 100 == 0:
print("Processed {} samples".format(sample_count))
progress_printer.epoch_summary(with_metric=True)
eval_model = create_fast_rcnn_eval_model(loss, image_input, roi_proposals, cfg)
eval_model.save(cfg['MODEL_PATH'])
return eval_model
def train_fast_rcnn(debug_output=False, model_path=model_file):
if debug_output:
print("Storing graphs and intermediate models to %s." %
os.path.join(abs_path, "Output"))
# Create the minibatch source
minibatch_source = create_mb_source(image_height, image_width,
num_channels, num_classes, num_rois,
base_path, "train")
# Input variables denoting features, rois and label data
image_input = C.input_variable((num_channels, image_height, image_width))
roi_input = C.input_variable((num_rois, 4))
label_input = C.input_variable((num_rois, num_classes))
# define mapping from reader streams to network inputs
input_map = {
image_input: minibatch_source.streams.features,
roi_input: minibatch_source.streams.rois,
label_input: minibatch_source.streams.roiLabels
}
# Instantiate the Fast R-CNN prediction model and loss function
frcn_output = frcn_predictor(image_input, roi_input, num_classes,
model_path)
ce = cross_entropy_with_softmax(frcn_output, label_input, axis=1)
pe = classification_error(frcn_output, label_input, axis=1)
if debug_output:
plot(frcn_output, os.path.join(abs_path, "Output", "graph_frcn.png"))
# Set learning parameters
l2_reg_weight = 0.0005
lr_per_sample = [0.00001] * 10 + [0.000001] * 5 + [0.0000001]
lr_schedule = learning_parameter_schedule_per_sample(lr_per_sample)
mm_schedule = momentum_schedule_per_sample(momentum_per_sample)
# Instantiate the trainer object as default
learner = momentum_sgd(frcn_output.parameters,
lr_schedule,
mm_schedule,
l2_regularization_weight=l2_reg_weight)
# Preparation for distributed learning, which is compatible for normal learner
learner = distributed.data_parallel_distributed_learner(
learner=learner,
num_quantization_bits=
num_quantization_bits, # non-quantized gradient accumulation
distributed_after=warm_up) # no warm start as default
progress_printer = ProgressPrinter(tag='Training',
num_epochs=max_epochs,
rank=distributed.Communicator.rank())
trainer = Trainer(frcn_output, (ce, pe), learner, progress_printer)
# Get minibatches of images and perform model training
print("Training Fast R-CNN model for %s epochs." % max_epochs)
log_number_of_parameters(frcn_output)
for epoch in range(max_epochs): # loop over epochs
sample_count = 0
while sample_count < epoch_size: # loop over minibatches in the epoch
data = minibatch_source.next_minibatch(
min(mb_size * C.Communicator.num_workers(),
epoch_size - sample_count),
input_map=input_map,
num_data_partitions=C.Communicator.num_workers(),
partition_index=C.Communicator.rank())
trainer.train_minibatch(data) # update model with it
sample_count += trainer.previous_minibatch_sample_count # count samples processed so far
trainer.summarize_training_progress()
if debug_output:
frcn_output.save(
os.path.join(abs_path, "Output",
"frcn_py_%s.model" % (epoch + 1)))
if distributed_flg:
distributed.Communicator.finalize()
return frcn_output
def init_train_fast_rcnn(image_height,
image_width,
num_classes,
num_rois,
mb_size,
max_epochs,
cntk_lr_per_image,
l2_reg_weight,
momentum_time_constant,
base_path,
boSkipTraining=False,
debug_output=False,
tensorboardLogDir=None):
#make sure we use GPU for training
if use_default_device().type() == 0:
print("WARNING: using CPU for training.")
else:
print("Using GPU for training.")
# Instantiate the Fast R-CNN prediction model
image_input = input_variable((3, image_height, image_width))
roi_input = input_variable((num_rois, 4))
label_input = input_variable((num_rois, num_classes))
frcn_output, frcn_penultimateLayer = frcn_predictor(
image_input, roi_input, num_classes, base_path)
if boSkipTraining:
print("Using pre-trained DNN without refinement")
return frcn_penultimateLayer
# Create the minibatch source and define mapping from reader streams to network inputs
minibatch_source, epoch_size = create_mb_source("train",
image_height,
image_width,
num_classes,
num_rois,
base_path,
randomize=True)
input_map = {
image_input: minibatch_source.streams.features,
roi_input: minibatch_source.streams.rois,
label_input: minibatch_source.streams.roiLabels
}
# set loss / error functions
ce = cross_entropy_with_softmax(frcn_output, label_input, axis=1)
pe = classification_error(frcn_output, label_input, axis=1)
if debug_output:
plot(frcn_output, "graph_frcn.png")
# set the progress printer(s)
progress_writers = [ProgressPrinter(tag='Training', num_epochs=max_epochs)]
if tensorboardLogDir != None:
tensorboard_writer = TensorBoardProgressWriter(
freq=10, log_dir=tensorboardLogDir, model=frcn_output)
progress_writers.append(tensorboard_writer)
# Set learning parameters and instantiate the trainer object
lr_per_sample = [f / float(num_rois) for f in cntk_lr_per_image]
lr_schedule = learning_rate_schedule(lr_per_sample, unit=UnitType.sample)
mm_schedule = momentum_as_time_constant_schedule(momentum_time_constant)
learner = momentum_sgd(frcn_output.parameters,
lr_schedule,
mm_schedule,
l2_regularization_weight=l2_reg_weight)
trainer = Trainer(frcn_output, (ce, pe), learner, progress_writers)
# Get minibatches of images and perform model training
print("Training Fast R-CNN model for %s epochs." % max_epochs)
log_number_of_parameters(frcn_output)
for epoch in range(max_epochs):
sample_count = 0
# loop over minibatches in the epoch
while sample_count < epoch_size:
data = minibatch_source.next_minibatch(min(
mb_size, epoch_size - sample_count),
input_map=input_map)
if sample_count % 100 == 1:
print(
"Training in progress: epoch {} of {}, sample count {} of {}"
.format(epoch, max_epochs, sample_count, epoch_size))
trainer.train_minibatch(data)
sample_count += trainer.previous_minibatch_sample_count # count samples processed so far
trainer.summarize_training_progress()
# Log mean of each parameter tensor, so that we can confirm that the parameters change indeed.
if tensorboardLogDir != None:
for parameter in frcn_output.parameters:
tensorboard_writer.write_value(parameter.uid + "/mean",
np.mean(parameter.value), epoch)
tensorboard_writer.write_value(parameter.uid + "/std",
np.std(parameter.value), epoch)
tensorboard_writer.write_value(parameter.uid + "/absSum",
np.sum(np.abs(parameter.value)),
epoch)
if debug_output:
frcn_output.save_model("frcn_py_%s.model" % (epoch + 1))
return frcn_output
def train_faster_rcnn_alternating(cfg):
'''
4-Step Alternating Training scheme from the Faster R-CNN paper:
# Create initial network, only rpn, without detection network
# --> train only the rpn (and conv3_1 and up for VGG16)
# buffer region proposals from rpn
# Create full network, initialize conv layers with imagenet, use buffered proposals
# --> train only detection network (and conv3_1 and up for VGG16)
# Keep conv weights from detection network and fix them
# --> train only rpn
# buffer region proposals from rpn
# Keep conv and rpn weights from step 3 and fix them
# --> train only detection network
'''
# setting pre- and post-nms top N to training values since buffered proposals are used for further training
test_pre = cfg["TEST"].RPN_PRE_NMS_TOP_N
test_post = cfg["TEST"].RPN_POST_NMS_TOP_N
cfg["TEST"].RPN_PRE_NMS_TOP_N = cfg["TRAIN"].RPN_PRE_NMS_TOP_N
cfg["TEST"].RPN_POST_NMS_TOP_N = cfg["TRAIN"].RPN_POST_NMS_TOP_N
# Learning parameters
rpn_lr_factor = cfg["MODEL"].RPN_LR_FACTOR
rpn_lr_per_sample_scaled = [x * rpn_lr_factor for x in cfg["CNTK"].RPN_LR_PER_SAMPLE]
frcn_lr_factor = cfg["MODEL"].FRCN_LR_FACTOR
frcn_lr_per_sample_scaled = [x * frcn_lr_factor for x in cfg["CNTK"].FRCN_LR_PER_SAMPLE]
l2_reg_weight = cfg["CNTK"].L2_REG_WEIGHT
mm_schedule = momentum_schedule(cfg["CNTK"].MOMENTUM_PER_MB)
rpn_epochs = cfg["CNTK"].RPN_EPOCHS
frcn_epochs = cfg["CNTK"].FRCN_EPOCHS
feature_node_name = cfg["MODEL"].FEATURE_NODE_NAME
last_conv_node_name = cfg["MODEL"].LAST_CONV_NODE_NAME
print("Using base model: {}".format(cfg["MODEL"].BASE_MODEL))
print("rpn_lr_per_sample: {}".format(rpn_lr_per_sample_scaled))
print("frcn_lr_per_sample: {}".format(frcn_lr_per_sample_scaled))
debug_output=cfg["CNTK"].DEBUG_OUTPUT
if debug_output:
print("Storing graphs and models to %s." % cfg.OUTPUT_PATH)
# Input variables denoting features, labeled ground truth rois (as 5-tuples per roi) and image dimensions
image_input = input_variable(shape=(cfg.NUM_CHANNELS, cfg.IMAGE_HEIGHT, cfg.IMAGE_WIDTH),
dynamic_axes=[Axis.default_batch_axis()],
name=feature_node_name)
feat_norm = image_input - Constant([[[v]] for v in cfg["MODEL"].IMG_PAD_COLOR])
roi_input = input_variable((cfg.INPUT_ROIS_PER_IMAGE, 5), dynamic_axes=[Axis.default_batch_axis()])
scaled_gt_boxes = alias(roi_input, name='roi_input')
dims_input = input_variable((6), dynamic_axes=[Axis.default_batch_axis()])
dims_node = alias(dims_input, name='dims_input')
rpn_rois_input = input_variable((cfg["TRAIN"].RPN_POST_NMS_TOP_N, 4), dynamic_axes=[Axis.default_batch_axis()])
rpn_rois_buf = alias(rpn_rois_input, name='rpn_rois')
# base image classification model (e.g. VGG16 or AlexNet)
base_model = load_model(cfg['BASE_MODEL_PATH'])
print("stage 1a - rpn")
if True:
# Create initial network, only rpn, without detection network
# initial weights train?
# conv: base_model only conv3_1 and up
# rpn: init new yes
# frcn: - -
# conv layers
conv_layers = clone_conv_layers(base_model, cfg)
conv_out = conv_layers(feat_norm)
# RPN and losses
rpn_rois, rpn_losses = create_rpn(conv_out, scaled_gt_boxes, dims_node, cfg)
stage1_rpn_network = combine([rpn_rois, rpn_losses])
# train
if debug_output: plot(stage1_rpn_network, os.path.join(cfg.OUTPUT_PATH, "graph_frcn_train_stage1a_rpn." + cfg["CNTK"].GRAPH_TYPE))
train_model(image_input, roi_input, dims_input, rpn_losses, rpn_losses,
rpn_lr_per_sample_scaled, mm_schedule, l2_reg_weight, rpn_epochs, cfg)
print("stage 1a - buffering rpn proposals")
buffered_proposals_s1 = compute_rpn_proposals(stage1_rpn_network, image_input, roi_input, dims_input, cfg)
print("stage 1b - frcn")
if True:
# Create full network, initialize conv layers with imagenet, fix rpn weights
# initial weights train?
# conv: base_model only conv3_1 and up
# rpn: stage1a rpn model no --> use buffered proposals
# frcn: base_model + new yes
# conv_layers
conv_layers = clone_conv_layers(base_model, cfg)
conv_out = conv_layers(feat_norm)
# use buffered proposals in target layer
rois, label_targets, bbox_targets, bbox_inside_weights = \
create_proposal_target_layer(rpn_rois_buf, scaled_gt_boxes, cfg)
# Fast RCNN and losses
fc_layers = clone_model(base_model, [cfg["MODEL"].POOL_NODE_NAME], [cfg["MODEL"].LAST_HIDDEN_NODE_NAME], CloneMethod.clone)
cls_score, bbox_pred = create_fast_rcnn_predictor(conv_out, rois, fc_layers, cfg)
detection_losses = create_detection_losses(cls_score, label_targets, bbox_pred, rois, bbox_targets, bbox_inside_weights, cfg)
pred_error = classification_error(cls_score, label_targets, axis=1, name="pred_error")
stage1_frcn_network = combine([rois, cls_score, bbox_pred, detection_losses, pred_error])
# train
if debug_output: plot(stage1_frcn_network, os.path.join(cfg.OUTPUT_PATH, "graph_frcn_train_stage1b_frcn." + cfg["CNTK"].GRAPH_TYPE))
train_model(image_input, roi_input, dims_input, detection_losses, pred_error,
frcn_lr_per_sample_scaled, mm_schedule, l2_reg_weight, frcn_epochs, cfg,
rpn_rois_input=rpn_rois_input, buffered_rpn_proposals=buffered_proposals_s1)
buffered_proposals_s1 = None
print("stage 2a - rpn")
if True:
# Keep conv weights from detection network and fix them
# initial weights train?
# conv: stage1b frcn model no
# rpn: stage1a rpn model yes
# frcn: - -
# conv_layers
conv_layers = clone_model(stage1_frcn_network, [feature_node_name], [last_conv_node_name], CloneMethod.freeze)
conv_out = conv_layers(image_input)
# RPN and losses
rpn = clone_model(stage1_rpn_network, [last_conv_node_name, "roi_input", "dims_input"], ["rpn_rois", "rpn_losses"], CloneMethod.clone)
rpn_net = rpn(conv_out, dims_node, scaled_gt_boxes)
rpn_rois = rpn_net.outputs[0]
rpn_losses = rpn_net.outputs[1]
stage2_rpn_network = combine([rpn_rois, rpn_losses])
# train
if debug_output: plot(stage2_rpn_network, os.path.join(cfg.OUTPUT_PATH, "graph_frcn_train_stage2a_rpn." + cfg["CNTK"].GRAPH_TYPE))
train_model(image_input, roi_input, dims_input, rpn_losses, rpn_losses,
rpn_lr_per_sample_scaled, mm_schedule, l2_reg_weight, rpn_epochs, cfg)
print("stage 2a - buffering rpn proposals")
buffered_proposals_s2 = compute_rpn_proposals(stage2_rpn_network, image_input, roi_input, dims_input, cfg)
print("stage 2b - frcn")
if True:
# Keep conv and rpn weights from step 3 and fix them
# initial weights train?
# conv: stage2a rpn model no
# rpn: stage2a rpn model no --> use buffered proposals
# frcn: stage1b frcn model yes -
# conv_layers
conv_layers = clone_model(stage2_rpn_network, [feature_node_name], [last_conv_node_name], CloneMethod.freeze)
conv_out = conv_layers(image_input)
# Fast RCNN and losses
frcn = clone_model(stage1_frcn_network, [last_conv_node_name, "rpn_rois", "roi_input"],
["cls_score", "bbox_regr", "rpn_target_rois", "detection_losses", "pred_error"], CloneMethod.clone)
stage2_frcn_network = frcn(conv_out, rpn_rois_buf, scaled_gt_boxes)
detection_losses = stage2_frcn_network.outputs[3]
pred_error = stage2_frcn_network.outputs[4]
# train
if debug_output: plot(stage2_frcn_network, os.path.join(cfg.OUTPUT_PATH, "graph_frcn_train_stage2b_frcn." + cfg["CNTK"].GRAPH_TYPE))
train_model(image_input, roi_input, dims_input, detection_losses, pred_error,
frcn_lr_per_sample_scaled, mm_schedule, l2_reg_weight, frcn_epochs, cfg,
rpn_rois_input=rpn_rois_input, buffered_rpn_proposals=buffered_proposals_s2)
buffered_proposals_s2 = None
# resetting config values to original test values
cfg["TEST"].RPN_PRE_NMS_TOP_N = test_pre
cfg["TEST"].RPN_POST_NMS_TOP_N = test_post
return create_faster_rcnn_eval_model(stage2_frcn_network, image_input, dims_input, cfg, rpn_model=stage2_rpn_network)
