def clear():
Model.destroy_model()
Laboratory.destroy_laboratory()
Model.create_model('test-model')
Laboratory.create_laboratory('test-lab')
destroy_mapping()
create_mapping('test-map-model-to-lab')
def __init__(self):
#load the model
self.model = Model()
self.model.model_load("./inference_models/ecgnet_0_fold_0.6078759902401878.pt")
#load preprocessng pipeline
#load thresholds
self.postptocessing = PostProcessing(0)
def test_no_dublicated_mapping():
clear()
m = Model.add_host('alice')
n = Laboratory.add_device('node1')
bind(m, n)
bind(m, n)
bind(n, m)
assert len(get_mapping().bindings()) == 1
def test_binding():
clear()
m = Model.add_host('alice')
n = Laboratory.add_device('node1')
bind(m, n)
assert(m.bound() == n)
assert(n.bound() == m)
assert len(get_mapping().bindings()) == 1
def load_12ECG_model(model_input):
"""Load Physionet2017 Model
model_input: This is an argument from running driver.py on command line. I think we just ignore it and hard-code
out model path.
"""
models_list = [
'ecgnet_0_fold_0.631593191670484', 'ecgnet_1_fold_0.6370736239012214',
'ecgnet_2_fold_0.6444454717434089', 'ecgnet_3_fold_0.6195938932528102',
'ecgnet_4_fold_0.6149398148500164', 'ecgnet_5_fold_0.6409127451470004'
]
os.makedirs(model_input + '/pretrained/', exist_ok=True)
# load the model
models = []
for i in models_list:
model_stack = Model(input_size=19000,
n_channels=15,
hparams=hparams,
gpu=[],
inference=True)
model_stack.model_load("./inference_models/" + i + ".pt")
model_stack.model_save(model_input + '/pretrained/' + i + ".pt")
models.append(model_stack)
return models
def __init__(self) -> None:
RefreshProgramLogs()
super().__init__()
self.video = cv2.VideoCapture(self.source)
self.model = Model(utilsdir=c.UTILSDIR,
modeldir=c.MODELDIR,
weights=c.WEIGHTS,
cfg=c.CFG,
labelsdir=c.LABELSDIR,
coco=c.COCONAMES,
cuda=False)
self.active_thread_count: int = None
self.p_time: float = 0
self.frame_counter: int = 0
def test_mapping_by_name():
clear()
m = Model.add_host('alice')
n = Laboratory.add_device('node1')
mi = m.add_interface('eth')
ni = n.add_interface('eth1')
bind('alice', 'node1')
bind('alice.eth', 'node1.eth1')
assert(m.bound() == n)
assert(n.bound() == m)
assert(mi.bound() == ni)
assert(ni.bound() == mi)
class Predict():
def __init__(self):
#load the model
self.model = Model()
self.model.model_load("./inference_models/ecgnet_0_fold_0.6078759902401878.pt")
#load preprocessng pipeline
#load thresholds
self.postptocessing = PostProcessing(0)
# threshold = 0
# for fold in range(6):
# threshold += float(open(f"threshold_{fold}.txt", "r").read())/6
#
# self.postptocessing.threshold = threshold
def predict(self,signal,meta):
############## Preprocessing ##############
#downsampling
X_resampled = np.zeros((signal.shape[0] // 2, 12))
for i in range(12):
X_resampled[:, i] = self.resampling.downsample(signal[:, 0], order=2)
#apply preprocessing
signal = self.apply_amplitude_scaling(X=X_resampled,y=meta)
# padding
sig_length = 19000
if X_resampled.shape[0] < sig_length:
padding = np.zeros((sig_length - X_resampled.shape[0], X_resampled.shape[1]))
X = np.concatenate([X_resampled, padding], axis=0)
if X_resampled.shape[0] > sig_length:
X_resampled = X_resampled[:sig_length, :]
############## Predictions ##############
predict = self.model.predict(X_resampled)
############## Postprocessing ##############
predict = self.postptocessing.run(predict)
predict = list(predict)
if predict[4] > 0 or predict[18] > 0:
predict[4] = 1
predict[18] = 1
if predict[23] > 0 or predict[12] > 0:
predict[23] = 1
predict[12] = 1
if predict[26] > 0 or predict[13] > 0:
predict[26] = 1
predict[13] = 1
return predict
@staticmethod
def apply_amplitude_scaling(X, y):
"""Get rpeaks for each channel and scale waveform amplitude by median rpeak amplitude of lead I."""
if y['rpeaks']:
for channel_rpeaks in y['rpeaks']:
if channel_rpeaks:
return X / np.median(X[y['rpeaks'][0], 0])
return (X - X.mean()) / X.std()
def train(self, positive_dir, negative_dir, hnm_dir):
model = Model.svc()
batch_size = 6000
increment = batch_size // 100
dir_walk_mgr = RecursiveDirectoryWalkerManager()
# Get positive samples
print('Loading positive samples...')
i = batch_size
samples = None
p_len = 0
while i:
if i % increment == 0:
sys.stdout.write('.')
sys.stdout.flush()
f = dir_walk_mgr.get_a_file(directory=positive_dir,
filters=['.jpg'])
if f is None:
print('Not enough positive samples T_T')
break
img = cv2.imread(os.path.normpath(f.path),
1) # Load as RGB for compatibility
if img is None:
continue
gray = ImageUtilities.preprocess(img,
convert_gray=cv2.COLOR_RGB2YCrCb)
gray = imresize(gray, HyperParam.window_size)
for j in range(3):
intensity = 0.3 * j
img = ImageUtilities.transform(gray, intensity=intensity)
if p_len < 100:
cv2.imwrite('./preview/' + str(p_len).zfill(2) + '.jpg',
img)
hist = self.compute_hog(img)
if samples is None:
samples = np.zeros((batch_size, ) + hist.shape,
dtype=np.float32)
samples[p_len, :] = hist
p_len += 1
i -= 1
if p_len >= len(samples): break
print(samples.shape)
positive_samples = np.copy(samples[0:p_len, :])
print('Positive samples loaded:', positive_samples.shape)
# Get negative samples
print('Loading negative samples...')
samples = None
n_len = p_len * 10
i = n_len
pt = 0
while i:
if i % increment == 0:
sys.stdout.write('.')
sys.stdout.flush()
f = dir_walk_mgr.get_a_file(directory=negative_dir,
filters=['.jpg'])
if f is None:
print('Not enough negative samples T_T')
break
img = cv2.imread(os.path.normpath(f.path),
1) # Load as RGB for compatibility
if img is None:
continue
gray = ImageUtilities.preprocess(img,
convert_gray=cv2.COLOR_RGB2YCrCb)
gray = imresize(gray, HyperParam.window_size)
hist = self.compute_hog(gray)
if samples is None:
samples = np.zeros((n_len, ) + hist.shape, dtype=np.float32)
try:
samples[pt, :] = hist.ravel()
except:
pass
pt += 1
i -= 1
print('Negative samples loaded:', samples.shape)
samples = np.concatenate([positive_samples, samples])
# Get hard-negative-mining samples
for di in range(10):
directory = os.path.normpath(
os.path.join(hnm_dir,
str(di + 1).zfill(4)))
if not os.path.isdir(directory):
break
print('Loading hard-negative-mining samples...', directory)
hnm_samples = None
t_len = batch_size * 10
i = t_len
print('target sample size', i)
pt = 0
while i:
if i % increment == 0:
sys.stdout.write('.')
sys.stdout.flush()
f = dir_walk_mgr.get_a_file(directory=directory,
filters=['.jpg'])
if f is None:
print('Not enough hard-negative-mining samples T_T')
break
img = cv2.imread(os.path.normpath(f.path),
1) # Load as RGB for compatibility
if img is None:
continue
gray = ImageUtilities.preprocess(
img, convert_gray=cv2.COLOR_RGB2YCrCb)
gray = imresize(gray, HyperParam.window_size)
hist = self.compute_hog(gray)
if hnm_samples is None:
hnm_samples = np.zeros((t_len, ) + hist.shape,
dtype=np.float32)
try:
hnm_samples[pt, :] = hist.ravel()
except:
pass
pt += 1
i -= 1
hnm_samples = np.copy(hnm_samples[0:pt, :])
print('HNM samples loaded:', hnm_samples.shape)
samples = np.concatenate([samples, hnm_samples])
print('Total samples:', samples.shape)
# Convert to numpy array of float32 and create labels
labels = np.zeros((samples.shape[0], ), dtype=np.int32)
labels[0:p_len] = 1
# Shuffle Samples
rand = np.random.RandomState(321)
shuffle = rand.permutation(len(samples))
samples = samples[shuffle]
labels = labels[shuffle]
print(samples.shape)
print(labels.shape)
print('Training...')
# Create SVM classifier
model.fit(samples, labels)
print(model.best_score_)
with open('svm.dat', 'wb') as f:
pickle.dump(model, f)
'''td = cv2.TrainData.create(InputArray samples, int layout, InputArray responses, InputArray varIdx=noArray(), InputArray sampleIdx=noArray(), InputArray sampleWeights=noArray(), InputArray varType=noArray())
def draw_feature(path='train_log/models/best-accuracy',
scale=1,
data_set='fashion_mnist'):
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print(device)
checkpoint = torch.load(path, map_location=device)
data_loader_train, data_loader_test, data_train, data_test = load_dataset(
data_set=data_set)
net = Model(n_feature)
net.load_state_dict(checkpoint['model_state_dict'])
exact_list = ['feature']
feature_extractor = FeatureExtractor(net, exact_list)
feature_extractor.to(device)
# get weight
weight = checkpoint['model_state_dict']['pred.weight'].to('cpu').data
weight_norm = weight / torch.norm(weight, dim=1, keepdim=True)
print("weight_norm: ", torch.norm(weight, dim=1))
# get feature
features = []
labels = []
for data in data_loader_train:
X_train, y_train = data
X_train = Variable(X_train).to(device)
outputs = feature_extractor(X_train)['feature'].data
features.append(outputs)
labels.append(y_train)
features = torch.cat(features, dim=0).to('cpu').data
features_norm = features / torch.norm(features, dim=1, keepdim=True)
features = features.numpy()
features_norm = features_norm.numpy()
labels = torch.cat(labels, dim=0).to('cpu').data.numpy()
# draw features
label_list = get_label_list(data_set)
plt.figure(1, figsize=(20, 20))
plt.subplot(221)
for i in range(10):
plt.plot([0, scale * weight[i, 0]], [0, scale * weight[i, 1]],
color=color_list[i])
feature = features[labels == i]
plt.scatter(feature[:, 0],
feature[:, 1],
c=color_list[i],
marker='.',
label=label_list[i],
s=1)
plt.legend()
plt.subplot(223)
for i in range(10):
plt.plot([0, weight_norm[i, 0]], [0, weight_norm[i, 1]],
color=color_list[i])
feature = features_norm[labels == i]
plt.scatter(feature[:, 0],
feature[:, 1],
c=color_list[i],
marker='.',
label=label_list[i],
s=1)
plt.legend()
# get feature
features = []
labels = []
for data in data_loader_test:
X_test, y_test = data
X_test = Variable(X_test).to(device)
outputs = feature_extractor(X_test)['feature'].data
features.append(outputs)
labels.append(y_test)
features = torch.cat(features, dim=0).to('cpu').data
features_norm = features / torch.norm(features, dim=1, keepdim=True)
features = features.numpy()
features_norm = features_norm.numpy()
labels = torch.cat(labels, dim=0).to('cpu').data.numpy()
plt.subplot(222)
for i in range(10):
plt.plot([0, scale * weight[i, 0]], [0, scale * weight[i, 1]],
color=color_list[i])
feature = features[labels == i]
plt.scatter(feature[:, 0],
feature[:, 1],
c=color_list[i],
marker='.',
label=label_list[i],
s=1)
plt.legend()
plt.subplot(224)
for i in range(10):
plt.plot([0, weight_norm[i, 0]], [0, weight_norm[i, 1]],
color=color_list[i])
feature = features_norm[labels == i]
plt.scatter(feature[:, 0],
feature[:, 1],
c=color_list[i],
marker='.',
label=label_list[i],
s=1)
plt.legend()
title = os.path.basename(os.getcwd()) + '-' + os.path.basename(path)
plt.suptitle(title)
fname = 'train_log/feature-{}'.format(os.path.basename(path))
figname = 'train_log/{}.png'.format(fname)
os.remove(figname) if os.path.exists(figname) else None
plt.savefig(fname)
plt.close('all')