def extract_pred_data(model, data_split):
#Putanje do značajki za odgovarajuću podjelu podataka
feat_path = glob.glob(os.path.sep.join([base_feats_dir, data_split, "*tfrecord"]))
#Definiranje dataset-a, broj vremenskih koraka nadopunjuje se do najvećeg broja koraka u cijelom skupu(457)
#kako bi kod izrade grafa imali optimiziranu implementaciju
#Struktura izlaza dataset-a: video_id, num_images, num_labels, image_seq, labels
dataset = build_test_pipeline(feat_path, example2video(feature_dim=FEATURE_DIM, training=False),
batch_size=BATCH_SIZE, padded_batch=True,
padded_shapes=([], [], [], [457, None], [457]),
padding_values=("", 0, 0, 0.,0))
#Izvlačenja duljina nizova bez nadopuna
y_lens = tf.concat([data_point[1] for data_point in dataset.as_numpy_iterator()], axis=0).numpy()
#Vjerojatnost oznake u vremenskom koraku
y_probs = tf.concat([model.predict(data_point[3]) for data_point in dataset.as_numpy_iterator()], axis=0)
#Povjerenje najizglednije oznake
y_conf = tf.reduce_max(y_probs, axis=-1).numpy()
#Predikcija oznake
y_pred = tf.argmax(y_probs, axis=-1, output_type=tf.int32).numpy()
#Stvarna oznaka
y_true = tf.concat([data_point[4] for data_point in dataset.as_numpy_iterator()], axis=0).numpy()
return y_lens, y_conf, y_pred, y_true
import tensorflow as tf
from phd_lib.data_pipeline.pipe_builders import build_test_pipeline
from phd_lib.data_pipeline.tfrecord_helpers import example2video
from phd_lib.models.test_models import SegmentationLoss_v2
from phd_lib.models.ms_tcn_model import build_ms_tcn_model
from tensorflow.keras.utils import to_categorical
import os
import glob
#%%Podatci
#Putanja do opažanja
path = r"D:\Phd_data\Video_feats_tfrecords\feat_extraction\HE\train"
train_tfs = glob.glob(os.path.join(path, "*.tfrecord"))
#%%Dva batch-a sa dva opažanja po batch-u
ds_no_pad = build_test_pipeline(train_tfs, example2video(2048), 2).take(2)
ds_padded = build_test_pipeline(train_tfs,
example2video(2048),
2,
padded_shapes=([457, None], [457])).take(2)
for element in ds_no_pad.take(2):
pred_1 = element[0]
y_1 = element[1]
for element in ds_padded.take(2):
pred_2 = element[0]
y_2 = element[1]
y_1_oh = to_categorical(y_1)
y_2_oh = to_categorical(y_2)
#%%Izgradi 2 modela, sa i bez maskiranja SA JEDNIM IZLAZOM i kompaliraj ih
filenames = glob.glob(os.path.join(ulazni_dir,
kadar,
data_split,
"*.tfrecord"), recursive=True)
return filenames
#Lokacija ulaznih podataka - potrebno je napraviti učenje za kadar HE i Fokus
train_filenames = get_feature_files(FEAT_TYPE, KADAR, "train")
val_filenames = get_feature_files(FEAT_TYPE, KADAR, "val")
#Generiranje train i val dataseta, za konkatenaciju kadrova dimenzija značajki je 4096, za ostale kadrove 2048
feature_dim = {"HE": 2048, "Fokus": 2048, "concat": 4096}
#Oba dataset-a su nadopunjena kako bi mogli koristi batch size veći od 1
train_dataset = build_train_pipeline(train_filenames,
example2video(feature_dim=feature_dim[KADAR], training=True),
batch_size=BATCH_SIZE, padded_batch=True)
val_dataset = build_train_pipeline(val_filenames,
example2video(feature_dim=feature_dim[KADAR], training=True),
batch_size=BATCH_SIZE, padded_batch=True)
#%%Definiranje modela za učenje
print("[INFO] priprema modela")
tf.keras.backend.clear_session()
#Komponente modela
if MODEL_TYPE == "CONV":
model = build_ms_tcn_model(input_shape=(None, feature_dim[KADAR]), num_layers_PG=5, R_stages=1, num_layers_R=5, filters=64, training=True, dropout_rate=0.5, shared_R=False)
else:
model = build_lstm_model(input_shape=(None, feature_dim[KADAR]), mask_value=0., bidirect=False, num_lstm_layers=1,
lstm_units=256, lstm_dropout=0.5, lstm_recurrent_dropout=0.,
"\n",
"MAD vremena trajanja aktivnosti:\n",
f"MAD: {MAD}\n",
"\n"
"Segmentacijski F1 rezultat uz minimalan prag preklapanja od 15, 25 i 50%\n",
f"F1@10: {f1_scores[0]}\n",
f"F1@25: {f1_scores[1]}\n",
f"F1@50: {f1_scores[2]}\n",
"\n",
"Srednja prosječna preciznost uz minimalan prag preklapanja od 15, 25 i 50%\n",
f"mAP@10: {mAP_scores[0]}\n",
f"mAP@25: {mAP_scores[1]}\n",
f"mAP@50: {mAP_scores[2]}\n",
"\n"])
#%%Prikaz rezultata
with open(full_metrics_path, "r") as f:
for line in f:
print(line.rstrip())
print()
print("Rezultati su pohranjeni u datoteku: ", full_metrics_path)
#%%Standardni testovi
test=0
if test:
#Testiranje rezultata izračuna metrike
feat_path = glob.glob(os.path.sep.join([base_feats_dir, args["data_split"], "*tfrecord"]))
dataset = build_test_pipeline(feat_path, example2video(feature_dim=FEATURE_DIM, training=True),
batch_size=BATCH_SIZE, padded_batch=True,
padded_shapes=([457, None], [457]))
print(model.evaluate(dataset))
ulazni_dir = feat_type_dict[feat_type]
filenames = glob.glob(os.path.join(ulazni_dir, kadar, data_split,
"*.tfrecord"),
recursive=True)
return filenames
#Lokacija ulaznih podataka
feat_path = get_feature_files(FEAT_TYPE, KADAR, data_split=args["data_split"])
#Definiranje dataset-a, broj vremenskih koraka nadopunjuje se do najvećeg broja koraka u cijelom skupu(457)
#kako bi kod izrade grafa imali optimiziranu implementaciju
#Struktura izlaza dataset-a: video_id, num_images, num_labels, image_seq, labels
dataset = build_test_pipeline(feat_path,
example2video(feature_dim=FEATURE_DIM,
training=False),
batch_size=BATCH_SIZE,
padded_batch=True,
padded_shapes=([], [], [], [457, None], [457]),
padding_values=("", 0, 0, 0., 0))
#Izvlačenja Video_id-a
y_id = tf.concat([data_point[0] for data_point in dataset.as_numpy_iterator()],
axis=0).numpy()
#Izvlačenja duljina nizova bez nadopuna
y_lens = tf.concat(
[data_point[1] for data_point in dataset.as_numpy_iterator()],
axis=0).numpy()
#Vjerojatnost oznake u vremenskom koraku