def get_image_ids_scope(session):
if LoggedIn() != True:
return defaultRedirect()
project = get_current_project(session)
version = get_current_version(session)
data = request.get_json(force=True)
search_term = data.get("search_term", None)
print(search_term, file=sys.stderr)
if search_term is None:
Images = session.query(Image).filter_by(version_id=version.id).order_by(Image.original_filename.desc()).limit(128)
else:
search_term = "%" + search_term + "%"
Images = session.query(Image).filter_by(version_id=version.id).filter(Image.original_filename.like(search_term)).order_by(Image.original_filename.desc()).limit(128)
gcs = storage.Client()
gcs = get_gcs_service_account(gcs)
bucket = gcs.get_bucket(settings.CLOUD_STORAGE_BUCKET)
Pre_condition_checked_images = []
for i in Images:
if i.soft_delete != True:
if i.url_signed_expiry is None or i.url_signed_expiry <= time.time():
rebuild_secure_urls(session, project, version, i)
Pre_condition_checked_images.append(i)
out = jsonify(images=[i.serialize() for i in Pre_condition_checked_images])
return out, 200, {'ContentType':'application/json'}
def runNewModel(session):
if LoggedIn() != True:
return defaultRedirect()
project = get_current_project(session)
JOB_NAME = "a_" + str(project.id)
# Creating model
requestDict = {'name': JOB_NAME,
'description': 'Built by runNewModel()'}
request = ml.projects().models().create(parent=projectID, body=requestDict)
try:
response = request.execute()
print(response, file=sys.stderr)
operationID = response['name']
out = 'success'
return out, 200, {'ContentType':'application/json'}
except errors.HttpError as EOFError:
print('There was an error. Check the details:', file=sys.stderr)
print(EOFError._get_reason(), file=sys.stderr)
out = 'failed'
return out, 200, {'ContentType':'application/json'}
return "Success", 200
def test_out_scope(session):
if LoggedIn() != True:
return defaultRedirect()
project = get_current_project(session)
version = get_current_version(session)
Images = session.query(Image).filter_by(version_id=version.id, is_test_image=True).order_by(Image.id.desc())
gcs = storage.Client()
gcs = get_gcs_service_account(gcs)
bucket = gcs.get_bucket(settings.CLOUD_STORAGE_BUCKET)
Public_urls = []
expiration_time = int(time.time() + 300)
file = str(project.id)+"/"+str(version.id)+"/"+ "test_inference_out/"+ "0_.jpg"
blob = bucket.blob(file)
public_url = blob.generate_signed_url(expiration=expiration_time)
Public_urls.append(public_url)
out = {}
out['image_ids'] = [i.id for i in Images]
out['width'] = [i.width for i in Images]
out['height'] = [i.height for i in Images]
out['public_url'] = [i for i in Public_urls]
return json.dumps(out), 200, {'ContentType':'application/json'}
def remove_duplicate_filenames_scope(session):
if LoggedIn() != True:
return defaultRedirect()
# May want to remove specific duplicates only
data = request.get_json(force=True)
images = data['images']
version = get_current_version(session)
Existing_images = session.query(Image).filter_by(version_id=version.id)
seen_file_once = []
looked_at = 0
soft_delete_marked = 0
for i in Existing_images:
looked_at += 1
for j in images:
if i.id == j['id']:
if i.original_filename not in seen_file_once:
seen_file_once.append(i.original_filename)
else:
i.soft_delete = True
session.add(i)
soft_delete_marked +=1
if looked_at % 100 == 0:
print("Looked at", looked_at, "Removed", soft_delete_marked)
print("Removed", soft_delete_marked, "duplicates")
out = 'success'
return json.dumps(out), 200, {'ContentType':'application/json'}
def toggle_test_image_all_scope(session):
if LoggedIn() != True:
return defaultRedirect()
data = request.get_json(force=True) # Force = true if not set as application/json'
images = data['images']
print("len(images)", len(images), file=sys.stderr)
version = get_current_version(session)
for i in images:
image_db = session.query(Image).filter_by(version_id=version.id, id=i['id']).first()
if image_db.is_test_image == True:
version.train_length += 1
version.test_length -= 1
else:
version.train_length -= 1
version.test_length += 1
image_db.is_test_image = not image_db.is_test_image
session.add(image_db)
session.add(version)
out = 'success'
return json.dumps(out), 200, {'ContentType':'application/json'}
def image_delete_scope(session):
if LoggedIn() != True:
return defaultRedirect()
data = request.get_json(force=True) # Force = true if not set as application/json'
image = data['image']
version = get_current_version(session)
Existing_images = session.query(Image).filter_by(version_id=version.id)
# Could delete multiple if get list of images to delete....
# Same issue as above otherwise, could do in query by checking id ie
# session.query(Employer).filter_by(id=employer_id).one()
for i in Existing_images:
if i.id == image['id']:
i.soft_delete = True
if i.is_test_image is True:
version.test_length -= 1
else:
version.train_length -= 1
# TODO Handle updating test / train length cache without db hit
session.add(i)
out = 'success'
return json.dumps(out), 200, {'ContentType':'application/json'}
def get_boxes_scope(session):
if LoggedIn() != True:
return defaultRedirect()
data = request.get_json(force=True)
if data['image_id'] is not None:
print("current_image.id", data['image_id'], file=sys.stderr)
image_id = data['image_id']
boxes = session.query(Box).filter_by(image_id=image_id).order_by(Box.id.desc())
else:
# This could be more sophisticated ie store last image we were working with
version = get_current_version(session)
image = session.query(Image).filter_by(version_id=version.id).order_by(Image.id.desc()).first()
boxes = session.query(Box).filter_by(image_id=image.id).order_by(Box.id.desc())
labels = []
for b in boxes:
label = session.query(Label).filter_by(id=b.label_id)
labels.append(label[0].serialize())
out = jsonify(boxes=[i.serialize() for i in boxes], labels=labels)
return out, 200, {'ContentType':'application/json'}
def new_box_scope(session):
if LoggedIn() != True:
return defaultRedirect()
data = request.get_json(force=True) # Force = true if not set as application/json'
boxes = data['boxes']
existing_boxes = session.query(Box).filter_by(image_id=data['image_id']).all()
# TODO better way to do this
for box_old in existing_boxes:
for box_new in boxes:
if box_old.id != box_new['id']:
session.delete(box_old)
if len(boxes) == 0: # Handle case of all boxes being deleted
session.delete(box_old)
for box in boxes:
if box['width'] > 5 and box['height'] > 5:
new_box = Box(x_min=box['x_min'], y_min=box['y_min'], x_max=box['x_max'], y_max=box['y_max'],
width=box['width'], height=box['height'],
image_id = box['image_id'],
label_id = box['label']['id'])
session.add(new_box)
session.commit()
print(box['id'], file=sys.stderr)
return json.dumps({'success':True}), 200, {'ContentType':'application/json'}
def remove_duplicate_filenames_all_scope(session):
if LoggedIn() != True:
return defaultRedirect()
version = get_current_version(session)
Existing_images = session.query(Image).filter_by(version_id=version.id)
seen_file_once = []
looked_at = 0
soft_delete_marked = 0
# basic greedy approach
# If we have seen the filename mark future objects with same file name as soft delete, else add to list
for i in Existing_images:
if i.original_filename in seen_file_once:
i.soft_delete = True
session.add(i)
soft_delete_marked +=1
else:
seen_file_once.append(i.original_filename)
looked_at += 1
if looked_at % 100 == 0:
print("Looked at", looked_at, "Removed", soft_delete_marked)
print("Removed", soft_delete_marked, "duplicates")
out = 'success'
return json.dumps(out), 200, {'ContentType':'application/json'}
def runTraining(session):
if LoggedIn() != True:
return defaultRedirect()
have_error, params = training_pre_conditions(session)
if have_error:
print("have error", params, file=sys.stderr)
return json.dumps(params), 200, {'ContentType':'application/json'}
# TODO Thinking on reasonable way to "copy" a version and track changes
project = get_current_project(session)
version = get_current_version(session)
machine_learning_settings = get_ml_settings(session=session, version=version)
JOB_NAME = "train_" + machine_learning_settings.JOB_NAME
print(JOB_NAME, file=sys.stderr)
REGION="us-central1"
RUNTIME_VERSION="1.2"
root_dir = "gs://" + settings.CLOUD_STORAGE_BUCKET + "/" + str(project.id) + "/" + str(version.id) + "/ml/" + str(machine_learning_settings.ml_compute_engine_id) + "/"
JOB_DIR = root_dir + "train"
pipeline_config_path = root_dir + "faster_rcnn_resnet.config"
MAIN_TRAINER_MODULE='object_detection.train'
training_inputs = {'scaleTier': 'CUSTOM',
'masterType': 'standard_gpu',
'workerType': 'standard_gpu',
'parameterServerType': 'standard_gpu',
'workerCount': 2,
'parameterServerCount': 1,
'packageUris': ['gs://' + settings.CLOUD_STORAGE_BUCKET + '/' + settings.LIB_OBJECT_DETECTION_PYTHON,
'gs://' + settings.CLOUD_STORAGE_BUCKET + '/' + settings.LIB_SLIM_PYTHON ],
'pythonModule': MAIN_TRAINER_MODULE,
'args': ['--train_dir', JOB_DIR,
'--pipeline_config_path', pipeline_config_path],
'region': REGION,
'jobDir': JOB_DIR,
'runtimeVersion': RUNTIME_VERSION }
job_spec = {'jobId': JOB_NAME, 'trainingInput': training_inputs}
request = ml.projects().jobs().create(body=job_spec, parent=projectID)
try:
response = request.execute()
print(response, file=sys.stderr)
out = 'success'
return out, 200, {'ContentType':'application/json'}
except errors.HttpError as EOFError:
print('There was an error. Check the details:', file=sys.stderr)
print(EOFError._get_reason(), file=sys.stderr)
out = 'failed'
return out, 500, {'ContentType':'application/json'}
return "success", 200
def version_view_scope(session):
if LoggedIn() != True:
return defaultRedirect()
version = get_current_version(session)
out = jsonify(version=version.serialize())
return out, 200, {'ContentType':'application/json'}
def user_view():
if LoggedIn() != True:
return defaultRedirect()
with sessionMaker.session_scope() as session:
user_id = getUserID()
user = session.query(User).filter_by(id=getUserID()).first()
out = jsonify(user=user.serialize())
return out, 200, {'ContentType': 'application/json'}
def trainingFrozenRun(session):
if LoggedIn() != True:
return defaultRedirect()
project = get_current_project(session)
version = get_current_version(session)
machine_learning_settings = get_ml_settings(session=session, version=version)
#now=strftime("%Y_%m_%d_%H_%M_%S", gmtime())
JOB_NAME = "frozen_user_" + machine_learning_settings.JOB_NAME
print(JOB_NAME, file=sys.stderr)
root_dir = "gs://" + settings.CLOUD_STORAGE_BUCKET + "/" + str(project.id) + "/" + str(version.id) + "/ml/" + str(machine_learning_settings.ml_compute_engine_id) + "/"
JOB_DIR = root_dir + str(machine_learning_settings.re_train_id) + "/frozen"
REGION ="us-central1"
RUNTIME_VERSION ="1.2"
# Should be updated during training and store in db?
trained_checkpoint_prefix = configNew.check_actual_model_path_name(session=session)
pipeline_config_path = root_dir + "faster_rcnn_resnet.config"
MAIN_TRAINER_MODULE ="object_detection.export_inference_graph"
training_inputs = {'scaleTier': 'CUSTOM',
'masterType': 'large_model',
'workerCount': 0,
'packageUris': ['gs://' + settings.CLOUD_STORAGE_BUCKET + '/' + settings.LIB_OBJECT_DETECTION_PYTHON,
'gs://' + settings.CLOUD_STORAGE_BUCKET + '/' + settings.LIB_SLIM_PYTHON ],
'pythonModule': MAIN_TRAINER_MODULE,
'args': ['--trained_checkpoint_prefix', trained_checkpoint_prefix,
'--pipeline_config_path', pipeline_config_path,
'--input_type', 'encoded_image_string_tensor',
'--output_directory', JOB_DIR],
'region': REGION,
'jobDir': JOB_DIR,
'runtimeVersion': RUNTIME_VERSION }
job_spec = {'jobId': JOB_NAME, 'trainingInput': training_inputs}
request = ml.projects().jobs().create(body=job_spec, parent=projectID)
try:
response = request.execute()
print(response, file=sys.stderr)
out = 'success'
return out, 200, {'ContentType':'application/json'}
except errors.HttpError as EOFError:
print('There was an error. Check the details:', file=sys.stderr)
print(EOFError._get_reason(), file=sys.stderr)
out = 'failed'
return out, 200, {'ContentType':'application/json'}
return "Success", 200
def machine_learning_settings_edit(session, next_id=False):
if LoggedIn() != True:
return defaultRedirect()
project = get_current_project(session)
version = get_current_version(session)
machine_learning_settings = get_ml_settings(session, version)
if next_id is True:
machine_learning_settings.ml_compute_engine_id += 1
session.add(machine_learning_settings)
session.commit()
return "Success", 200
def toggle_done_labeling_scope(session):
if LoggedIn() != True:
return defaultRedirect()
data = request.get_json(force=True) # Force = true if not set as application/json'
image = data['image']
version = get_current_version(session)
Existing_images = session.query(Image).filter_by(version_id=version.id)
for i in Existing_images:
if i.id == image['id']:
i.done_labeling = not i.done_labeling
session.add(i)
out = 'success'
return json.dumps(out), 200, {'ContentType':'application/json'}
def runNewVersion(session):
if LoggedIn() != True:
return defaultRedirect()
project = get_current_project(session)
version = get_current_version(session)
machine_learning_settings = get_ml_settings(session=session, version=version)
project_root = "a_" + str(project.id)
# Creating version
modelID= '{}/models/{}'.format(projectID, project_root)
versionName = "a_" + str(version.id) + "_" + str(machine_learning_settings.ml_compute_engine_id)
versionName += "_" + str(machine_learning_settings.re_train_id)
# Maybe could include more info like date time?
versionDescription = 'created by runNewVersion()'
root_dir = "gs://" + settings.CLOUD_STORAGE_BUCKET + "/" + str(project.id) + "/" + str(version.id)
root_dir += "/ml/" + str(machine_learning_settings.ml_compute_engine_id) + "/" + str(machine_learning_settings.re_train_id) + "/"
JOB_DIR = root_dir + "frozen/saved_model"
requestDict = {'name': versionName,
'description': versionDescription,
'deploymentUri': JOB_DIR,
'runtimeVersion': '1.2'}
request = ml.projects().models().versions().create(
parent=modelID, body=requestDict)
try:
response = request.execute()
print(response, file=sys.stderr)
operationID = response['name']
out = 'success'
return out, 200, {'ContentType':'application/json'}
except errors.HttpError as EOFError:
# Something went wrong, print out some information.
print('There was an error. Check the details:', file=sys.stderr)
print(EOFError._get_reason(), file=sys.stderr)
out = 'failed'
return out, 200, {'ContentType':'application/json'}
return out, 200, {'ContentType':'application/json'}
def label_refresh_scope(session):
if LoggedIn() != True:
return defaultRedirect()
project = get_current_project(session)
Labels_db = session.query(Label).filter_by(
project_id=project.id).order_by(Label.id.desc())
# TODO can do soft_delete != "True" check in here???
Labels = []
for i in Labels_db:
if i.soft_delete != True:
Labels.append(i)
out = {}
out['ids'] = [i.id for i in Labels]
out['names'] = [i.name for i in Labels]
#Colour?
return json.dumps(out), 200, {'ContentType': 'application/json'}
def label_delete_scope(session):
if LoggedIn() != True:
return defaultRedirect()
data = request.get_json(
force=True) # Force = true if not set as application/json'
label = data['label']
project = get_current_project(session)
existing_Labels = session.query(Label).filter_by(
project_id=project.id).order_by(Label.id.desc())
for i in existing_Labels:
if i.id == label['id']:
i.soft_delete = True
session.add(i)
out = 'success'
session.commit()
return json.dumps(out), 200, {'ContentType': 'application/json'}
def machine_learning_settings_new(session):
if LoggedIn() != True:
return defaultRedirect()
project = get_current_project(session)
version = get_current_version(session)
iterations = 2000 # Could get from user
previous_goal_iterations = iterations
ml_compute_engine_id = 0
new_ml_settings = Machine_learning_settings(iterations=iterations,
previous_goal_iterations=previous_goal_iterations,
ml_compute_engine_id=ml_compute_engine_id)
session.add(new_ml_settings)
session.commit()
version.machine_learning_settings_id = new_ml_settings.id
session.add(version)
session.commit()
return "Success", 200
def label_new_scope(session):
if LoggedIn() != True:
return defaultRedirect()
data = request.get_json(
force=True) # Force = true if not set as application/json'
label = data['label']
print(label, data.keys(), file=sys.stderr)
have_error = False
params = {}
#existing_label = session.query(Label).filter_by(id=label['id']).first()
existing_label = None # Maybe do more with this later
project = get_current_project(session)
if label is None:
params['error'] = "No Label"
have_error = True
if existing_label is not None:
params['error'] = "Existing label"
have_error = True
if have_error:
return json.dumps(params), 200, {'ContentType': 'application/json'}
else:
label['colour'] = "blue" # since JS is being strange
new_label = Label(name=label['name'],
colour=label['colour'],
project_id=project.id)
session.add(new_label)
return json.dumps({'success': True}), 200, {
'ContentType': 'application/json'
}
def runTrainingPipeline(re_train=0):
"""
"""
# 1 == retrain
# Can you pass a bool here? prefer that but just doing this for now
if LoggedIn() != True:
return defaultRedirect()
@copy_current_request_context
def task_manager():
def task_manager_scope(session):
print("[Training task manager] Started. Retrain_flag:", re_train, file=sys.stderr)
session = sessionMaker.scoppedSession() # Threadsafe
# Maybe better to have this somewhere else
version = get_current_version(session=session)
if version.machine_learning_settings_id is None:
ml_settings.machine_learning_settings_new(session=session)
# Advance one for training if not retraining
if re_train == 0:
ml_settings.machine_learning_settings_edit(session=session, next_id=True)
project = get_current_project(session=session)
machine_learning_settings = get_ml_settings(session=session, version=version)
JOB_NAME = "__projectID_" + str(project.id) + "__versionID_" + str(version.id) + "__ml_compute_id_" + str(machine_learning_settings.ml_compute_engine_id)
if re_train == 1:
machine_learning_settings.re_train_id += 1
JOB_NAME += "__retrainID_" + str(machine_learning_settings.re_train_id)
machine_learning_settings.JOB_NAME = JOB_NAME
session.add(machine_learning_settings)
session.commit()
# Do YAML for retraining
# TODO way to detect if this is needed or not...
yamlNew(hold_thread=True)
labelMapNew()
fasterRcnnResnetNew(re_train=re_train) # Config file
tfrecordsNew(hold_thread=True)
### TRAINING
runTraining(session)
config = {}
config['PUBSSUB_TOPIC'] = settings.PUB_SUB_TOPIC
config['PROJECT'] = settings.GOOGLE_PROJECT_NAME
publisher = pubsub_v1.PublisherClient()
topic_path = publisher.topic_path(config['PROJECT'], config['PUBSSUB_TOPIC'])
JOB_NAME = "train_" + machine_learning_settings.JOB_NAME
JOB_NAME_FORMATTED = projectID + "/jobs/" + JOB_NAME
training_flag = True
while training_flag is True:
request = ml.projects().jobs().get(name=JOB_NAME_FORMATTED)
# TODO error handling
response = request.execute()
data = json.dumps(response)
print(data, file=sys.stderr)
data = data.encode()
publisher.publish(topic_path, data=data)
a = response['state']
if a == "SUCCEEDED" or a == "FAILED" or a =="CANCELLED":
training_flag = False
else:
time.sleep(30)
#### END TRAINING
# Now need to run new model on re training
if re_train == 0:
runNewModel(session)
##### FROZEN
trainingFrozenRun(session)
JOB_NAME = "frozen_user_" + machine_learning_settings.JOB_NAME
JOB_NAME_FORMATTED = projectID + "/jobs/" + JOB_NAME
frozen_flag = True
while frozen_flag is True:
request = ml.projects().jobs().get(name=JOB_NAME_FORMATTED)
# TODO error handling
response = request.execute()
data = json.dumps(response)
print(data, file=sys.stderr)
data = data.encode()
publisher.publish(topic_path, data=data)
a = response['state']
if a == "SUCCEEDED" or a == "FAILED" or a =="CANCELLED":
frozen_flag = False
else:
time.sleep(30)
#####
runNewVersion(session)
time.sleep(60*8) # Sleep while long running operation
runInferenceSingle()
print("[Training task manager] SUCCESS", file=sys.stderr)
t.cancel()
with sessionMaker.session_scope() as session:
task_manager_scope(session)
t = threading.Timer(0, task_manager)
t.daemon = True
t.start()
out = 'success'
return out, 200, {'ContentType':'application/json'}
def test():
if LoggedIn() != True:
return defaultRedirect()
return render_template('/annotation/test.html')
def uploadPOST():
if LoggedIn() != True:
return defaultRedirect()
@copy_current_request_context
def task_manager(name, extension): # Function is defined here to so as to use request context decorator with scopped session.
def task_manager_scope(session):
counter = 0
project = get_current_project(session)
version = get_current_version(session)
project_id = project.id
version_id = version.id
out = ""
with open(name, "rb") as file:
if extension == ".zip":
try:
zip_ref = zipfile.ZipFile(BytesIO(file.read()), 'r')
temp_dir = tempfile.mkdtemp()
zip_ref.extractall(temp_dir)
zip_ref.close()
filenames = sorted(os.listdir(temp_dir))
len_filenames = len(filenames) # Variable used in loop below so storing here
print("[ZIP processor] Found", len_filenames, file=sys.stderr)
Thread_session = sessionMaker.scoppedSession() # Threadsafe
for filename in filenames:
t_2 = threading.Timer(0, multi_thread_task_manager,
args=(temp_dir, filename, Thread_session,
project_id, version_id))
t_2.start()
# Slow down new threads if too many open
len_threads = len(threading.enumerate())
if len_threads > settings.MAX_UPLOAD_THREADS:
time.sleep(settings.MAX_UPLOAD_THREADS * 25)
if len_threads > settings.TARGET_UPLOAD_THREADS:
time.sleep(settings.TARGET_UPLOAD_THREADS * 5)
counter += 1
if counter % 10 == 0:
print("[ZIP processor]", (counter / len(filenames) ) * 100, "% done." , file=sys.stderr)
Thread_session.remove()
except zipfile.BadZipFile:
out = {"files": [{"name": "Error bad zip file"}]}
else:
content_type = "image/" + str(extension)
file_name = os.path.split(file.name)[1]
out = process_one_image_file(file=file, name=file_name,
content_type=content_type, extension=extension,
session=session, project_id=project_id, version_id=version_id)
out = {"files": [{"name": "Processed files"}]}
print(out, file=sys.stderr)
t.cancel()
with sessionMaker.session_scope() as session:
task_manager_scope(session)
# Update train counts
with sessionMaker.session_scope() as session:
time.sleep(1)
version = get_current_version(session)
image_count = session.query(Image).filter_by(version_id=version.id).filter(Image.is_test_image == False, Image.soft_delete == False).count()
version.train_length = image_count
session.add(version)
file = request.files.get('files[]')
if not file:
return "No file", 400
extension = os.path.splitext(file.filename)[1].lower()
if extension in allowed_file_names:
file.filename = secure_filename(file.filename) # http://flask.pocoo.org/docs/0.12/patterns/fileuploads/
temp_dir = tempfile.mkdtemp()
name = temp_dir + "/" + file.filename
file.save(name)
t = threading.Timer(0, task_manager, args=(name, extension)) # https://stackoverflow.com/questions/29330982/python-timer-nonetype-object-is-not-callable-error
t.daemon = True
t.start()
out = {"files": [{"name": "Processing files. Please wait a few seconds per file uploaded"}]}
else:
out = {"files": [{"name": "Invalid file extension"}]}
return jsonify(out)
def yamlNew(hold_thread=False):
if LoggedIn() != True:
return defaultRedirect()
@copy_current_request_context
def task_manager():
def task_manager_scope(session):
project = get_current_project(session)
version = get_current_version(session)
machine_learning_settings = get_ml_settings(session=session,
version=version)
Images = session.query(Image).filter_by(
version_id=version.id).order_by(Image.id.desc())
annotations_list = []
len_images = Images.count()
counter = 0
for image in Images:
# TODO maybe better to do in database?
if image.soft_delete != True and image.is_test_image != True and image.done_labeling == True:
boxes = session.query(Box).filter_by(
image_id=image.id).order_by(Box.id.desc()).limit(100)
box_dict_list = []
for box in boxes:
label = session.query(Label).filter_by(
id=box.label_id).one()
if label is None:
print("Label is none", file=sys.stderr)
box_dict_list.append({
'label_id': label.id,
'label_name': label.name,
'x_min': box.x_min,
'x_max': box.x_max,
'y_min': box.y_min,
'y_max': box.y_max
})
image_dict = {
'image': {
'image_id': image.id,
'image_width': image.width,
'image_height': image.height,
'original_filename': image.original_filename
}
}
boxes_dict = {'boxes': box_dict_list}
annotations_list.append(
{'annotations': [image_dict, boxes_dict]})
if counter % 10 == 0:
print("Percent done", (counter / len_images) * 100,
file=sys.stderr)
counter += 1
print("annotations_list len",
len(annotations_list),
file=sys.stderr)
yaml_data = yaml.dump(annotations_list, default_flow_style=False)
gcs = storage.Client()
gcs = get_gcs_service_account(gcs)
bucket = gcs.get_bucket(settings.CLOUD_STORAGE_BUCKET)
project_str = str(project.id) + "/" + str(
version.id) + "/ml/" + str(
machine_learning_settings.ml_compute_engine_id
) + "/annotations.yaml"
blob = bucket.blob(project_str)
blob.upload_from_string(yaml_data, content_type='text/yaml')
print("Built YAML, link below", file=sys.stderr)
link = get_secure_link(blob)
print(link, file=sys.stderr)
t.cancel()
with sessionMaker.session_scope() as session:
task_manager_scope(session)
t = threading.Timer(0, task_manager)
t.daemon = True
t.start()
print("[YAML processor] Started", file=sys.stderr)
# Default to False for HTTP
# Use
if hold_thread is True:
t.join()
out = "Started"
return out, 200
def tfrecordsNew(hold_thread=False):
if LoggedIn() != True:
return defaultRedirect()
@copy_current_request_context
def task_manager():
def task_manager_scope(session):
project = get_current_project(session)
version = get_current_version(session)
ml_settings = get_ml_settings(session=session, version=version)
project_str = str(project.id) + "/" + str(version.id) + "/"
gcs = storage.Client()
gcs = get_gcs_service_account(gcs)
bucket = gcs.get_bucket(settings.CLOUD_STORAGE_BUCKET)
blob = bucket.blob(project_str + "ml/" +
str(ml_settings.ml_compute_engine_id) +
"/tfrecords_0.record")
INPUT_YAML = project_str + "ml/" + str(
ml_settings.ml_compute_engine_id) + "/annotations.yaml"
yaml_blob = bucket.blob(INPUT_YAML)
yaml_bytes = yaml_blob.download_as_string()
examples = yaml.load(yaml_bytes)
len_examples = len(examples)
print("Loaded ", len(examples), "examples", file=sys.stderr)
images_dir = project_str + "images/"
for i in range(len(examples)):
examples[i]['annotations'][0]['image'][
'image_id'] = images_dir + str(
examples[i]['annotations'][0]['image']['image_id'])
counter = 0
all_examples = []
# Reassign db ids to be 1 2 3 etc for tensorflow
# TODO this is terrible surely better way to do this
Labels = []
labels = session.query(Label).filter_by(project_id=project.id)
for i in labels:
if i.soft_delete != True:
Labels.append(i)
Labels_unique = set(Labels)
Labels.sort(key=lambda x: x.id)
label_dict = {}
start_at_1_label = 1
lowest_label = 0
for label in Labels:
if label.id > lowest_label:
label_dict[label.id] = start_at_1_label
start_at_1_label += 1
lowest_label = label.id
print("label_dict length", len(label_dict), file=sys.stderr)
temp = tempfile.NamedTemporaryFile()
writer = tf.python_io.TFRecordWriter(str(temp.name))
for example in examples:
tf_example = create_tf_example(example['annotations'],
label_dict)
writer.write(tf_example.SerializeToString())
if counter % 2 == 0:
print("Percent done", (counter / len_examples) * 100)
counter += 1
writer.close()
blob.upload_from_file(temp, content_type='text/record')
temp.close()
link = get_secure_link(blob)
print(blob.name, file=sys.stderr)
print("Built TF records", file=sys.stderr)
t.cancel()
with sessionMaker.session_scope() as session:
task_manager_scope(session)
t = threading.Timer(0, task_manager)
t.daemon = True
t.start()
print("[TF records processor] Started", file=sys.stderr)
if hold_thread is True:
t.join()
return "Started tf_records", 200
def project_get():
if LoggedIn():
return render_template('/workspace/newWorkspace.html')
else:
return defaultRedirect()
def faster_rcnn_resnet_new_scope(session):
if LoggedIn() != True:
return defaultRedirect()
project = get_current_project(session)
version = get_current_version(session)
machine_learning_settings = get_ml_settings(session=session,
version=version)
project_str = str(project.id) + "/" + str(version.id) + "/ml/" + str(
machine_learning_settings.ml_compute_engine_id)
project_str += "/faster_rcnn_resnet.config"
# Faster R-CNN with Resnet-101 (v1)
root_dir = "gs://" + settings.CLOUD_STORAGE_BUCKET + "/" + str(
project.id) + "/" + str(version.id) + "/ml/" + str(
machine_learning_settings.ml_compute_engine_id) + "/"
num_classes_var = version.labels_number # TODO get this automatically
print("version.labels_number", version.labels_number, file=sys.stderr)
min_dimension_var = 720
max_dimension_var = 1280 # TODO get this automaticaly within limit
first_stage_max_proposals_var = 100
label_map_path_var = root_dir + "label_map.pbtxt"
# Testing for multiple records?
input_path_var = root_dir + "tfrecords_*.record"
num_steps_var = machine_learning_settings.iterations
# This is the shared generic starting point
fine_tune_checkpoint_var = "gs://" + settings.CLOUD_STORAGE_BUCKET + "/" + settings.RESNET_PRE_TRAINED_MODEL
if re_train == 1:
machine_learning_settings.previous_goal_iterations = machine_learning_settings.iterations
num_steps_var = machine_learning_settings.iterations + 1500
fine_tune_checkpoint_var = check_actual_model_path_name(
session=session)
model = "model {"
faster_rcnn = "\nfaster_rcnn {"
num_classes = "\nnum_classes: " + str(num_classes_var)
image_resizer = "\nimage_resizer { \nkeep_aspect_ratio_resizer {"
min_dimension = "\nmin_dimension: " + str(min_dimension_var)
max_dimension = "\nmax_dimension: " + str(
max_dimension_var) + "\n} \n}"
feature_extractor = "\nfeature_extractor { \n type: 'faster_rcnn_resnet101' "
first_stage_features_stride = "\nfirst_stage_features_stride: 16 \n } "
first_stage_anchor_generator = """first_stage_anchor_generator \n{ \ngrid_anchor_generator
{ \nscales: [0.25, 0.5, 1.0, 2.0] \naspect_ratios: [0.5, 1.0, 2.0] \nheight_stride: 16 \n
width_stride: 16 \n } \n } \n"""
first_stage_box_predictor_conv_hyperparams = """
first_stage_box_predictor_conv_hyperparams {
op: CONV
regularizer {
l2_regularizer {
weight: 0.0
}
}
initializer {
truncated_normal_initializer {
stddev: 0.01
}
}
}
first_stage_nms_score_threshold: 0.0
first_stage_nms_iou_threshold: 0.7
first_stage_localization_loss_weight: 2.0
first_stage_objectness_loss_weight: 1.0
initial_crop_size: 14
maxpool_kernel_size: 2
maxpool_stride: 2
"""
first_stage_max_proposals = "\nfirst_stage_max_proposals:" + str(
first_stage_max_proposals_var)
second_stage_box_predictor = """
second_stage_box_predictor {
mask_rcnn_box_predictor {
use_dropout: false
dropout_keep_probability: 1.0
fc_hyperparams {
op: FC
regularizer {
l2_regularizer {
weight: 0.0
}
}
initializer {
variance_scaling_initializer {
factor: 1.0
uniform: true
mode: FAN_AVG
}
}
}
}
}
"""
second_stage_post_processing = """
second_stage_post_processing {
batch_non_max_suppression {
score_threshold: 0.0
iou_threshold: 0.6
max_detections_per_class: 100
"""
max_total_detections = "max_total_detections:" + str(
first_stage_max_proposals_var) + "\n}"
score_converter = """
score_converter: SOFTMAX
}
second_stage_localization_loss_weight: 2.0
second_stage_classification_loss_weight: 1.0
"""
second_stage_batch_size = "\nsecond_stage_batch_size: " + str(
first_stage_max_proposals_var) + "\n }\n }\n"
train_config = """
train_config: {
batch_size: 1
optimizer {
momentum_optimizer: {
learning_rate: {
manual_step_learning_rate {
initial_learning_rate: 0.0003
schedule {
step: 0
learning_rate: .0003
}
schedule {
step: 900000
learning_rate: .00003
}
schedule {
step: 1200000
learning_rate: .000003
}
}
}
momentum_optimizer_value: 0.9
}
use_moving_average: false
}
gradient_clipping_by_norm: 10.0
"""
fine_tune_checkpoint = "\nfine_tune_checkpoint: '" + str(
fine_tune_checkpoint_var) + "'"
from_detection_checkpoint = "\nfrom_detection_checkpoint: true"
num_steps = "\nnum_steps: " + str(num_steps_var)
data_augmentation_options = """
data_augmentation_options {
random_horizontal_flip {
}
}
}
"""
train_input_reader = """
train_input_reader: {
tf_record_input_reader {
"""
input_path = "\ninput_path: '" + str(input_path_var) + "' \n}"
label_map_path = "\nlabel_map_path: '" + str(
label_map_path_var) + "'\n}"
config_file_a = model + faster_rcnn + num_classes + image_resizer + min_dimension + max_dimension + feature_extractor + first_stage_features_stride + first_stage_anchor_generator
config_file_b = first_stage_box_predictor_conv_hyperparams + first_stage_max_proposals + second_stage_box_predictor + second_stage_post_processing + max_total_detections
config_file_c = score_converter + second_stage_batch_size + train_config + fine_tune_checkpoint + from_detection_checkpoint + num_steps + data_augmentation_options + train_input_reader + input_path + label_map_path
config_file = config_file_a + config_file_b + config_file_c
gcs = storage.Client()
bucket = gcs.get_bucket(settings.CLOUD_STORAGE_BUCKET)
blob = bucket.blob(project_str)
blob.upload_from_string(config_file, content_type='text/config')
print("Built Config", file=sys.stderr)
out = 'success'
return out, 200, {'ContentType': 'application/json'}
def label_map_new_scope(session):
if LoggedIn() != True:
return defaultRedirect()
project = get_current_project(session)
version = get_current_version(session)
ml_settings = get_ml_settings(session=session, version=version)
Images = session.query(Image).filter_by(version_id=version.id)
Labels = []
# TO DO Refactor ie maintain a cache all label ids used in a version
# Would need to store that cache per version
# And update / delete it as labels are changed OR Collect at YAML stage
labels = session.query(Label).filter_by(project_id=project.id)
for i in labels:
if i.soft_delete != True:
Labels.append(i)
# Map db ids to id s staring with 123
Labels.sort(key=lambda x: x.id)
label_dict = {}
start_at_1_label = 1
lowest_label = 0
for label in Labels:
if label.id > lowest_label:
label_dict[label.id] = start_at_1_label
start_at_1_label += 1
lowest_label = label.id
print("label_dict length", len(label_dict), file=sys.stderr)
project_str = str(project.id) + "/" + str(version.id) + "/ml/" + str(
ml_settings.ml_compute_engine_id)
project_str += "/label_map.pbtext"
file = ""
Labels_unique = set(Labels)
len_labels = len(Labels_unique)
version.labels_number = len_labels
session.add(version)
session.commit()
for i, c in enumerate(Labels_unique):
new = "\nitem {"
id = "\nid: " + str(label_dict[c.id])
name = "\nname: " + str(c.name) + "\n }\n"
file += new + id + name
gcs = storage.Client()
gcs = get_gcs_service_account(gcs)
bucket = gcs.get_bucket(settings.CLOUD_STORAGE_BUCKET)
blob = bucket.blob(project_str)
blob.upload_from_string(file, content_type='text/pbtext')
print("Built label_map", file=sys.stderr)
out = get_secure_link(blob)
return out, 200, {'ContentType': 'application/json'}
def upload():
if LoggedIn() != True:
return defaultRedirect()
return render_template('/storage/uploadView.html')
def runInferenceSingle():
if LoggedIn() != True:
return defaultRedirect()
@copy_current_request_context
def task_manager():
def task_manager_scope(session):
credentials = GoogleCredentials.get_application_default()
ml = discovery.build('ml', 'v1', credentials=credentials)
projectID = 'projects/{}'.format(settings.GOOGLE_PROJECT_NAME)
project = get_current_project(session=session)
version = get_current_version(session=session)
machine_learning_settings = get_ml_settings(session=session,
version=version)
Images_db = session.query(Image).filter_by(version_id=version.id,
is_test_image=True)
REGION = "us-central1"
RUNTIME_VERSION = "1.2"
modelName = "a_" + str(project.id)
versionName = "a_" + str(version.id) + "_" + str(
machine_learning_settings.ml_compute_engine_id)
versionName += "_" + str(machine_learning_settings.re_train_id)
modelVersionName = '{}/models/{}/versions/{}'.format(
projectID, modelName, versionName)
gcs = storage.Client()
bucket = gcs.get_bucket(settings.CLOUD_STORAGE_BUCKET)
filenames = []
root_dir = str(project.id) + "/" + str(version.id) + "/"
for image in Images_db:
#print(image.is_test_image, file=sys.stderr)
if image.soft_delete != True:
filenames.append(root_dir + "images/" + str(image.id))
break
Rows = []
Images = []
print("len(filenames):", len(filenames), file=sys.stderr)
for file in filenames:
blob = bucket.blob(file)
image = blob.download_as_string()
# Resize
image = scipy.misc.imread(BytesIO(image))
if image is None:
raise IOError("Could not open")
# TODO BETTER WAY
#image = scipy.misc.imresize(image, (640, 960))
temp = tempfile.mkdtemp()
new_temp_filename = temp + "/resized.jpg"
scipy.misc.imsave(new_temp_filename, image)
# Otherwise have strange byte issues
blob = bucket.blob(file + "_test_resized")
blob.upload_from_filename(new_temp_filename,
content_type="image/jpg")
image = blob.download_as_string()
encoded_contents = base64.b64encode(image).decode('UTF-8')
row = {'b64': encoded_contents}
Rows.append(row)
Images.append(image)
output = {'instances': Rows}
ml_request = ml.projects().predict(name=modelVersionName,
body=output)
PATH_TO_LABELS = root_dir + "ml/" + str(
machine_learning_settings.ml_compute_engine_id
) + "/label_map.pbtext"
label_map_blob = bucket.blob(PATH_TO_LABELS)
label_map_data = label_map_blob.download_as_string()
category_index = categoryMap(session=session)
try:
time0 = time.time()
response = ml_request.execute()
time1 = time.time()
print("Time in seconds", (time1 - time0), file=sys.stderr)
print(response, file=sys.stderr)
for i in range(len(Images)):
response = response['predictions'][i] # First one
boxes = response['detection_boxes']
scores = response['detection_scores']
classes = response['detection_classes']
boxes = np.array(boxes)
scores = np.array(scores)
classes = np.array(classes, dtype=int)
print(classes, file=sys.stderr)
image_np = scipy.misc.imread(BytesIO(Images[i]))
# Handle gray scale
if len(image_np.shape) == 2:
image_np = np.stack((image_np, ) * 3, axis=2)
print(image_np.shape)
visualization_utils.visualize_boxes_and_labels_on_image_array(
image_np,
boxes,
classes,
scores,
category_index,
use_normalized_coordinates=True,
min_score_thresh=.3,
line_thickness=2)
blob = bucket.blob(root_dir + "test_inference_out/" +
str(i) + "_.jpg")
temp = tempfile.mkdtemp()
new_temp_filename = temp + "/inference_" + str(i) + "_.jpg"
scipy.misc.imsave(new_temp_filename, image_np)
blob.upload_from_filename(new_temp_filename,
content_type="image/jpg")
min_score_thresh = .05
for i in range(len(boxes)):
if scores[i] > min_score_thresh:
class_name = category_index[classes[i]]['name']
print(class_name, scores[i], file=sys.stderr)
# TODO add pub sub messaging
out = 'success'
except errors.HttpError as EOFError:
print('There was an error. Check the details:',
file=sys.stderr)
print(EOFError._get_reason(), file=sys.stderr)
out = 'failed'
t.cancel()
with sessionMaker.session_scope() as session:
task_manager_scope(session)
t = threading.Timer(0, task_manager)
t.daemon = True
t.start()
out = 'success'
return out, 200,
