Spaces:
Sleeping
Sleeping
working
Browse files- movinet/data.py +79 -0
- movinet/model.py +9 -0
- movinet/scripts/train.py +58 -0
- playgrounds/main.py +0 -3
- playgrounds/verify_metal.py +14 -0
- requirements.txt +35 -0
- site_packages/models +1 -0
movinet/data.py
ADDED
@@ -0,0 +1,79 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from pathlib import Path
|
2 |
+
import random
|
3 |
+
from typing import Literal
|
4 |
+
import cv2
|
5 |
+
import numpy as np
|
6 |
+
import tensorflow as tf
|
7 |
+
|
8 |
+
TRAINING_RATIO = 0.1
|
9 |
+
VALIDATION_RATIO = 0.01
|
10 |
+
|
11 |
+
def format_frames(frame, output_size):
|
12 |
+
frame = tf.image.convert_image_dtype(frame, tf.float32)
|
13 |
+
frame = tf.image.resize_with_pad(frame, *output_size)
|
14 |
+
return frame
|
15 |
+
|
16 |
+
def frames_from_video_file(video_path: str, n_frames: int, output_size=(256, 256), frame_step=15):
|
17 |
+
capture = cv2.VideoCapture(video_path)
|
18 |
+
if not capture.isOpened(): raise ValueError('Video file could not be opened.')
|
19 |
+
total_frames = capture.get(cv2.CAP_PROP_FRAME_COUNT)
|
20 |
+
need_frames = 1 + (n_frames - 1) * frame_step
|
21 |
+
|
22 |
+
if need_frames <= total_frames:
|
23 |
+
start = random.randint(0, total_frames - need_frames + 1)
|
24 |
+
capture.set(cv2.CAP_PROP_POS_FRAMES, start)
|
25 |
+
|
26 |
+
frames = []
|
27 |
+
for _ in range(n_frames - 1):
|
28 |
+
for _ in range(frame_step):
|
29 |
+
ok, frame = capture.read()
|
30 |
+
if ok:
|
31 |
+
frames.append(format_frames(frame, output_size))
|
32 |
+
else:
|
33 |
+
frames.append(np.zeros((output_size[0], output_size[1], 3)))
|
34 |
+
capture.release()
|
35 |
+
|
36 |
+
frames = np.array(frames)
|
37 |
+
frames = frames[..., [2, 1, 0]]
|
38 |
+
return frames
|
39 |
+
|
40 |
+
def Data(data_dir: Path):
|
41 |
+
return {
|
42 |
+
'training':{
|
43 |
+
a.name: (lambda ps: ps[:int(len(ps) * TRAINING_RATIO)])([x for x in a.iterdir()])
|
44 |
+
for a in data_dir.iterdir()
|
45 |
+
},
|
46 |
+
'validation': {
|
47 |
+
a.name: (lambda ps: ps[
|
48 |
+
int(len(ps) * TRAINING_RATIO) :
|
49 |
+
int(len(ps) * (TRAINING_RATIO + VALIDATION_RATIO))
|
50 |
+
])([x for x in a.iterdir()])
|
51 |
+
for a in data_dir.iterdir()
|
52 |
+
},
|
53 |
+
}
|
54 |
+
|
55 |
+
def frame_generator(data_dir: Path, n_frames: int, split: Literal['training', 'validation']):
|
56 |
+
class_names = sorted([x.name for x in data_dir.iterdir()])
|
57 |
+
class_ids_for_name = {
|
58 |
+
name: i
|
59 |
+
for i, name in enumerate(class_names)
|
60 |
+
}
|
61 |
+
data = Data(data_dir)
|
62 |
+
|
63 |
+
def generator():
|
64 |
+
pairs = [
|
65 |
+
(path, name)
|
66 |
+
for name, paths in data[split].items()
|
67 |
+
for path in paths
|
68 |
+
]
|
69 |
+
random.shuffle(pairs)
|
70 |
+
for path, name in pairs:
|
71 |
+
video_frames = frames_from_video_file(str(path), n_frames)
|
72 |
+
label = class_ids_for_name[name]
|
73 |
+
yield video_frames, label
|
74 |
+
return generator
|
75 |
+
|
76 |
+
def total_steps(data_dir: Path):
|
77 |
+
data = Data(data_dir)
|
78 |
+
size = lambda d: sum([len(x) for x in d.values()])
|
79 |
+
return size(data['training']), size(data['validation'])
|
movinet/model.py
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from site_packages.models.official.projects.movinet.modeling import movinet_model
|
2 |
+
|
3 |
+
def build_classifier(batch_size: int, num_frames: int, resolution: int, backbone, num_classes: int):
|
4 |
+
model = movinet_model.MovinetClassifier(
|
5 |
+
backbone=backbone,
|
6 |
+
num_classes=num_classes,
|
7 |
+
)
|
8 |
+
model.build([batch_size, num_frames, resolution, resolution, 3])
|
9 |
+
return model
|
movinet/scripts/train.py
ADDED
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
from pathlib import Path
|
3 |
+
|
4 |
+
import tensorflow as tf
|
5 |
+
import tf_keras as keras
|
6 |
+
|
7 |
+
from site_packages.models.official.projects.movinet.modeling import movinet
|
8 |
+
|
9 |
+
from movinet.data import frame_generator, total_steps
|
10 |
+
from movinet.model import build_classifier
|
11 |
+
|
12 |
+
model_id = 'a0'
|
13 |
+
resolution = 256
|
14 |
+
batch_size = 8
|
15 |
+
num_frames = 8
|
16 |
+
num_classes = 6
|
17 |
+
model_save_path = "out/aero-recognize-classifier.keras"
|
18 |
+
num_epochs = 2
|
19 |
+
|
20 |
+
print('Load data.')
|
21 |
+
data_dir = Path('assets/datasets/Aero')
|
22 |
+
output_signature = (
|
23 |
+
tf.TensorSpec(shape=(None, None, None, 3), dtype=tf.float32),
|
24 |
+
tf.TensorSpec(shape=(), dtype=tf.int16),
|
25 |
+
)
|
26 |
+
training_data = tf.data.Dataset.from_generator(
|
27 |
+
frame_generator(data_dir, num_frames, 'training'),
|
28 |
+
output_signature=output_signature,
|
29 |
+
)
|
30 |
+
training_data = training_data.batch(batch_size)
|
31 |
+
validation_data = tf.data.Dataset.from_generator(
|
32 |
+
frame_generator(data_dir, num_frames, 'validation'),
|
33 |
+
output_signature=output_signature,
|
34 |
+
)
|
35 |
+
validation_data = validation_data.batch(batch_size)
|
36 |
+
|
37 |
+
print('Build model.')
|
38 |
+
backbone = movinet.Movinet(model_id=model_id)
|
39 |
+
backbone.trainable = True
|
40 |
+
model = build_classifier(batch_size, num_frames, resolution, backbone, 6)
|
41 |
+
|
42 |
+
print('Start training.')
|
43 |
+
model_dir = os.path.dirname(model_save_path)
|
44 |
+
save_model = keras.callbacks.ModelCheckpoint(filepath=model_save_path)
|
45 |
+
loss = keras.losses.SparseCategoricalCrossentropy(from_logits=True)
|
46 |
+
# optimizer = keras.optimizers.legacy.Adam(learning_rate=0.001)
|
47 |
+
model.compile(optimizer='adam', loss=loss, metrics=['accuracy'])
|
48 |
+
train_steps, validation_steps = total_steps(data_dir)
|
49 |
+
results = model.fit(
|
50 |
+
training_data,
|
51 |
+
steps_per_epoch=train_steps,
|
52 |
+
validation_data=validation_data,
|
53 |
+
validation_steps=validation_steps,
|
54 |
+
epochs=num_epochs,
|
55 |
+
validation_freq=1,
|
56 |
+
verbose=1,
|
57 |
+
callbacks=[save_model],
|
58 |
+
)
|
playgrounds/main.py
DELETED
@@ -1,3 +0,0 @@
|
|
1 |
-
from playgrounds.yolo import main
|
2 |
-
|
3 |
-
main()
|
|
|
|
|
|
|
|
playgrounds/verify_metal.py
ADDED
@@ -0,0 +1,14 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import tensorflow as tf
|
2 |
+
|
3 |
+
cifar = tf.keras.datasets.cifar100
|
4 |
+
(x_train, y_train), (x_test, y_test) = cifar.load_data()
|
5 |
+
model = tf.keras.applications.ResNet50(
|
6 |
+
include_top=True,
|
7 |
+
weights=None,
|
8 |
+
input_shape=(32, 32, 3),
|
9 |
+
classes=100,
|
10 |
+
)
|
11 |
+
|
12 |
+
loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=False)
|
13 |
+
model.compile(optimizer="adam", loss=loss_fn, metrics=["accuracy"])
|
14 |
+
model.fit(x_train, y_train, epochs=5, batch_size=64)
|
requirements.txt
ADDED
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
gradio
|
2 |
+
tensorflow
|
3 |
+
opencv-python
|
4 |
+
|
5 |
+
# cspell: disable
|
6 |
+
# models/official
|
7 |
+
six
|
8 |
+
google-api-python-client>=1.6.7
|
9 |
+
kaggle>=1.3.9
|
10 |
+
numpy>=1.20
|
11 |
+
oauth2client
|
12 |
+
pandas>=0.22.0
|
13 |
+
psutil>=5.4.3
|
14 |
+
py-cpuinfo>=3.3.0
|
15 |
+
scipy>=0.19.1
|
16 |
+
tensorflow-hub>=0.6.0
|
17 |
+
tensorflow-model-optimization>=0.4.1
|
18 |
+
tensorflow-datasets
|
19 |
+
tf-keras
|
20 |
+
gin-config
|
21 |
+
tf_slim>=1.1.0
|
22 |
+
Cython
|
23 |
+
matplotlib
|
24 |
+
# Loader becomes a required positional argument in 6.0 in yaml.load
|
25 |
+
pyyaml>=6.0.0
|
26 |
+
# CV related dependencies
|
27 |
+
opencv-python-headless
|
28 |
+
Pillow
|
29 |
+
pycocotools
|
30 |
+
# NLP related dependencies
|
31 |
+
seqeval
|
32 |
+
sentencepiece
|
33 |
+
sacrebleu
|
34 |
+
# Projects/vit dependencies
|
35 |
+
immutabledict
|
site_packages/models
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
Subproject commit d14cf43b09cc29d68900bb9f766de19b01acde40
|