def _load_feature_extractor():
global inference_engine
import torch
from sense import engine
from sense import feature_extractors
if inference_engine is None:
feature_extractor = feature_extractors.StridedInflatedEfficientNet()
# Remove internal padding for feature extraction and training
checkpoint = torch.load('resources/backbone/strided_inflated_efficientnet.ckpt')
feature_extractor.load_state_dict(checkpoint)
feature_extractor.eval()
# Create Inference Engine
inference_engine = engine.InferenceEngine(feature_extractor, use_gpu=True)
print(
"Warning: if start training, files in the path_out folder will be overwritten."
)
while True:
overwrite_files = input(
"Enter Y if allow to overwrite files, enter N then the program will stop: "
)
if overwrite_files.lower() == "y":
break
elif overwrite_files.lower() == "n":
sys.exit()
else:
print('Wrong input')
# Load feature extractor
feature_extractor = feature_extractors.StridedInflatedEfficientNet()
if resume:
# load the last classifier
checkpoint_classifier = torch.load(
os.path.join(path_out, 'last_classifier.checkpoint'))
checkpoint = torch.load(
'resources/backbone/strided_inflated_efficientnet.ckpt')
# Update original weights in case some intermediate layers have been finetuned
name_finetuned_layers = set(checkpoint.keys()).intersection(
checkpoint_classifier.keys())
for key in name_finetuned_layers:
checkpoint[key] = checkpoint_classifier.pop(key)
else:
checkpoint = torch.load(
'resources/backbone/strided_inflated_efficientnet.ckpt')
def init_model(transform):
use_gpu = True
inference_engine = None
neural_network = None
postprocessors = None
if transform == 'gesture':
# Load feature extractor
feature_extractor = feature_extractors.StridedInflatedEfficientNet()
feature_extractor.load_weights_from_resources(
'backbone/strided_inflated_efficientnet.ckpt')
feature_extractor.eval()
# Load a logistic regression classifier
gesture_classifier = LogisticRegression(
num_in=feature_extractor.feature_dim, num_out=30)
checkpoint = load_weights_from_resources(
'gesture_detection/efficientnet_logistic_regression.ckpt')
gesture_classifier.load_state_dict(checkpoint)
gesture_classifier.eval()
# Concatenate feature extractor and met converter
neural_network = Pipe(feature_extractor, gesture_classifier)
postprocessors = [
PostprocessClassificationOutput(INT2LAB, smoothing=4)
]
elif transform == 'fitness':
weight = float(60)
height = float(170)
age = float(20)
gender = 'female'
# Load feature extractor
feature_extractor = feature_extractors.StridedInflatedMobileNetV2()
feature_extractor.load_weights_from_resources(
'backbone/strided_inflated_mobilenet.ckpt')
feature_extractor.eval()
# Load fitness activity classifier
gesture_classifier = LogisticRegression(
num_in=feature_extractor.feature_dim, num_out=81)
checkpoint = load_weights_from_resources(
'fitness_activity_recognition/mobilenet_logistic_regression.ckpt')
gesture_classifier.load_state_dict(checkpoint)
gesture_classifier.eval()
# Load MET value converter
met_value_converter = calorie_estimation.METValueMLPConverter()
checkpoint = load_weights_from_resources(
'calorie_estimation/mobilenet_features_met_converter.ckpt')
met_value_converter.load_state_dict(checkpoint)
met_value_converter.eval()
# Concatenate feature extractor with downstream nets
neural_network = Pipe(
feature_extractor,
feature_converter=[gesture_classifier, met_value_converter])
postprocessors = [
PostprocessClassificationOutput(INT2LAB, smoothing=8, indices=[0]),
calorie_estimation.CalorieAccumulator(weight=weight,
height=height,
age=age,
gender=gender,
smoothing=12,
indices=[1])
]
if neural_network is not None:
inference_engine = InferenceEngine(neural_network, use_gpu=use_gpu)
start_inference(inference_engine)
return (inference_engine, postprocessors), None