Мобильная версияЧтобы получить больше информации для обучения, изображения обучающего набора генерируются заново каждую эпоху с новым случайным вращением. .
Однако кажется, что модель не обучается, поскольку она предсказывает один и тот же класс для каждого изображения. Я пробовал менять размер пакета, оптимизатор, скорость обучения, более сложные модели, но ничего не помогло решить проблему.
В этом примере я использую
500 обучающих образцов, 50 проверочных образцов и 10 тестовых образцов. Я пробовал использовать до 2000 обучающих выборок, однако возникла та же проблема.
Это мой результат:
Код: Выделить всё
Using TensorFlow backend.
WARNING:tensorflow:From /home/lisa/.local/lib/python3.7/site-packages/keras/backend/tensorflow_backend.py:4070: The name tf.nn.max_pool is deprecated. Please use tf.nn.max_pool2d instead.
Model: "model_1"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) (None, 224, 224, 3) 0
_________________________________________________________________
conv2d_1 (Conv2D) (None, 222, 222, 32) 896
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 111, 111, 32) 0
_________________________________________________________________
conv2d_2 (Conv2D) (None, 109, 109, 64) 18496
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 54, 54, 64) 0
_________________________________________________________________
conv2d_3 (Conv2D) (None, 52, 52, 128) 73856
_________________________________________________________________
max_pooling2d_3 (MaxPooling2 (None, 26, 26, 128) 0
_________________________________________________________________
conv2d_4 (Conv2D) (None, 24, 24, 128) 147584
_________________________________________________________________
max_pooling2d_4 (MaxPooling2 (None, 12, 12, 128) 0
_________________________________________________________________
flatten_1 (Flatten) (None, 18432) 0
_________________________________________________________________
dense_1 (Dense) (None, 512) 9437696
_________________________________________________________________
dense_2 (Dense) (None, 360) 184680
=================================================================
Total params: 9,863,208
Trainable params: 9,863,208
Non-trainable params: 0
_________________________________________________________________
2019-11-06 11:08:47.885295: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2019-11-06 11:08:47.901431: I tensorflow/core/platform/profile_utils/cpu_utils.cc:94] CPU Frequency: 3408000000 Hz
2019-11-06 11:08:47.902091: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x55f4487aac50 executing computations on platform Host. Devices:
2019-11-06 11:08:47.902139: I tensorflow/compiler/xla/service/service.cc:175] StreamExecutor device (0): ,
2019-11-06 11:08:47.903354: I tensorflow/stream_executor/platform/default/dso_loader.cc:42] Successfully opened dynamic library libcuda.so.1
2019-11-06 11:08:47.921001: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2019-11-06 11:08:47.921953: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1640] Found device 0 with properties:
name: GeForce GTX 970 major: 5 minor: 2 memoryClockRate(GHz): 1.1775
pciBusID: 0000:01:00.0
2019-11-06 11:08:47.922112: I tensorflow/stream_executor/platform/default/dso_loader.cc:42] Successfully opened dynamic library libcudart.so.10.0
2019-11-06 11:08:47.922988: I tensorflow/stream_executor/platform/default/dso_loader.cc:42] Successfully opened dynamic library libcublas.so.10.0
2019-11-06 11:08:47.923739: I tensorflow/stream_executor/platform/default/dso_loader.cc:42] Successfully opened dynamic library libcufft.so.10.0
2019-11-06 11:08:47.923921: I tensorflow/stream_executor/platform/default/dso_loader.cc:42] Successfully opened dynamic library libcurand.so.10.0
2019-11-06 11:08:47.924921: I tensorflow/stream_executor/platform/default/dso_loader.cc:42] Successfully opened dynamic library libcusolver.so.10.0
2019-11-06 11:08:47.925684: I tensorflow/stream_executor/platform/default/dso_loader.cc:42] Successfully opened dynamic library libcusparse.so.10.0
2019-11-06 11:08:47.928111: I tensorflow/stream_executor/platform/default/dso_loader.cc:42] Successfully opened dynamic library libcudnn.so.7
2019-11-06 11:08:47.928199: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2019-11-06 11:08:47.929103: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2019-11-06 11:08:47.929818: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1763] Adding visible gpu devices: 0
2019-11-06 11:08:47.929844: I tensorflow/stream_executor/platform/default/dso_loader.cc:42] Successfully opened dynamic library libcudart.so.10.0
2019-11-06 11:08:47.976192: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1181] Device interconnect StreamExecutor with strength 1 edge matrix:
2019-11-06 11:08:47.976213: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1187] 0
2019-11-06 11:08:47.976219: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1200] 0: N
2019-11-06 11:08:47.976372: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2019-11-06 11:08:47.977217: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2019-11-06 11:08:47.978039: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:1005] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2019-11-06 11:08:47.978851: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1326] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 3466 MB memory) -> physical GPU (device: 0, name: GeForce GTX 970, pci bus id: 0000:01:00.0, compute capability: 5.2)
2019-11-06 11:08:47.980313: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x55f449158000 executing computations on platform CUDA. Devices:
2019-11-06 11:08:47.980326: I tensorflow/compiler/xla/service/service.cc:175] StreamExecutor device (0): GeForce GTX 970, Compute Capability 5.2
WARNING:tensorflow:From /home/lisa/.local/lib/python3.7/site-packages/keras/backend/tensorflow_backend.py:422: The name tf.global_variables is deprecated. Please use tf.compat.v1.global_variables instead.
Epoch 1/50
2019-11-06 11:08:48.922378: I tensorflow/stream_executor/platform/default/dso_loader.cc:42] Successfully opened dynamic library libcublas.so.10.0
2019-11-06 11:08:49.080712: I tensorflow/stream_executor/platform/default/dso_loader.cc:42] Successfully opened dynamic library libcudnn.so.7
16/16 [==============================] - 3s 199ms/step - loss: 10271548.3852 - mse_angle: 88.4758 - val_loss: 6.0310 - val_mse_angle: 83.5972
Epoch 2/50
16/16 [==============================] - 1s 84ms/step - loss: 6.0294 - mse_angle: 87.3988 - val_loss: 6.2498 - val_mse_angle: 90.8889
Epoch 3/50
16/16 [==============================] - 1s 82ms/step - loss: 6.9000 - mse_angle: 90.9215 - val_loss: 6.2606 - val_mse_angle: 96.1042
Epoch 4/50
16/16 [==============================] - 1s 82ms/step - loss: 6.0261 - mse_angle: 90.2238 - val_loss: 6.1281 - val_mse_angle: 89.1111
Epoch 5/50
16/16 [==============================] - 1s 82ms/step - loss: 6.0339 - mse_angle: 90.6246 - val_loss: 6.1609 - val_mse_angle: 84.5764
Epoch 6/50
16/16 [==============================] - 1s 82ms/step - loss: 5.9953 - mse_angle: 90.6105 - val_loss: 6.0373 - val_mse_angle: 97.3819
Epoch 7/50
16/16 [==============================] - 1s 82ms/step - loss: 5.9419 - mse_angle: 90.0617 - val_loss: 6.0082 - val_mse_angle: 99.2257
Epoch 8/50
16/16 [==============================] - 1s 82ms/step - loss: 5.9563 - mse_angle: 89.2258 - val_loss: 6.0243 - val_mse_angle: 99.2257
Epoch 9/50
16/16 [==============================] - 1s 83ms/step - loss: 5.9515 - mse_angle: 92.9902 - val_loss: 6.0726 - val_mse_angle: 87.7812
Epoch 10/50
16/16 [==============================] - 1s 82ms/step - loss: 5.9554 - mse_angle: 89.0434 - val_loss: 6.0980 - val_mse_angle: 81.9757
Epoch 11/50
16/16 [==============================] - 1s 82ms/step - loss: 5.9761 - mse_angle: 90.9699 - val_loss: 6.1573 - val_mse_angle: 99.1910
Epoch 12/50
16/16 [==============================] - 1s 83ms/step - loss: 5.9674 - mse_angle: 87.5254 - val_loss: 6.1502 - val_mse_angle: 91.5312
Epoch 13/50
16/16 [==============================] - 1s 82ms/step - loss: 5.9457 - mse_angle: 90.9098 - val_loss: 6.1447 - val_mse_angle: 89.7708
Epoch 14/50
16/16 [==============================] - 1s 82ms/step - loss: 5.9803 - mse_angle: 92.3281 - val_loss: 6.1520 - val_mse_angle: 97.5417
Epoch 15/50
16/16 [==============================] - 1s 82ms/step - loss: 5.9663 - mse_angle: 91.3766 - val_loss: 6.1332 - val_mse_angle: 81.1562
Epoch 16/50
16/16 [==============================] - 1s 82ms/step - loss: 5.9707 - mse_angle: 89.2891 - val_loss: 6.0442 - val_mse_angle: 88.7361
Epoch 17/50
16/16 [==============================] - 1s 82ms/step - loss: 5.9691 - mse_angle: 87.9980 - val_loss: 5.8971 - val_mse_angle: 81.1562
Epoch 18/50
16/16 [==============================] - 1s 82ms/step - loss: 5.9675 - mse_angle: 87.8605 - val_loss: 5.9070 - val_mse_angle: 81.1562
Epoch 19/50
16/16 [==============================] - 1s 81ms/step - loss: 5.9816 - mse_angle: 88.3820 - val_loss: 6.0384 - val_mse_angle: 90.0694
Epoch 20/50
16/16 [==============================] - 1s 82ms/step - loss: 6.0144 - mse_angle: 91.3855 - val_loss: 6.1066 - val_mse_angle: 90.0694
Epoch 21/50
16/16 [==============================] - 1s 82ms/step - loss: 5.9556 - mse_angle: 92.5727 - val_loss: 6.2307 - val_mse_angle: 86.2465
Epoch 22/50
16/16 [==============================] - 1s 83ms/step - loss: 5.9522 - mse_angle: 90.1418 - val_loss: 6.1750 - val_mse_angle: 81.9062
Epoch 23/50
16/16 [==============================] - 1s 81ms/step - loss: 5.9603 - mse_angle: 88.3703 - val_loss: 6.0286 - val_mse_angle: 81.9062
Epoch 24/50
16/16 [==============================] - 1s 82ms/step - loss: 5.9608 - mse_angle: 90.1531 - val_loss: 5.9816 - val_mse_angle: 97.9549
Epoch 25/50
16/16 [==============================] - 1s 82ms/step - loss: 5.9764 - mse_angle: 88.8660 - val_loss: 6.0606 - val_mse_angle: 89.0174
Epoch 26/50
16/16 [==============================] - 1s 82ms/step - loss: 5.9771 - mse_angle: 90.2336 - val_loss: 6.0759 - val_mse_angle: 83.8507
Epoch 27/50
16/16 [==============================] - 1s 82ms/step - loss: 6.0073 - mse_angle: 90.3863 - val_loss: 6.0298 - val_mse_angle: 83.8507
Epoch 28/50
16/16 [==============================] - 1s 82ms/step - loss: 5.9664 - mse_angle: 89.0832 - val_loss: 5.9718 - val_mse_angle: 83.5972
Epoch 29/50
16/16 [==============================] - 1s 82ms/step - loss: 5.9445 - mse_angle: 88.3340 - val_loss: 5.9844 - val_mse_angle: 82.4306
Epoch 30/50
16/16 [==============================] - 1s 81ms/step - loss: 5.9596 - mse_angle: 90.2934 - val_loss: 5.8805 - val_mse_angle: 83.0521
Epoch 31/50
16/16 [==============================] - 1s 83ms/step - loss: 5.9729 - mse_angle: 91.9238 - val_loss: 5.9500 - val_mse_angle: 84.4444
Epoch 32/50
16/16 [==============================] - 1s 83ms/step - loss: 5.9743 - mse_angle: 90.0250 - val_loss: 6.0221 - val_mse_angle: 97.5556
Epoch 33/50
16/16 [==============================] - 1s 81ms/step - loss: 5.9469 - mse_angle: 86.5922 - val_loss: 6.0201 - val_mse_angle: 87.6076
Epoch 34/50
16/16 [==============================] - 1s 82ms/step - loss: 5.9822 - mse_angle: 93.8836 - val_loss: 5.9119 - val_mse_angle: 81.3472
Epoch 35/50
16/16 [==============================] - 1s 81ms/step - loss: 5.9751 - mse_angle: 88.9707 - val_loss: 5.9052 - val_mse_angle: 99.3993
Epoch 36/50
16/16 [==============================] - 1s 83ms/step - loss: 5.9564 - mse_angle: 89.6219 - val_loss: 5.9162 - val_mse_angle: 92.5278
Epoch 37/50
16/16 [==============================] - 1s 82ms/step - loss: 5.9864 - mse_angle: 94.1816 - val_loss: 5.9559 - val_mse_angle: 90.5278
Epoch 38/50
16/16 [==============================] - 1s 82ms/step - loss: 5.9566 - mse_angle: 88.3102 - val_loss: 6.0087 - val_mse_angle: 99.3993
Epoch 39/50
16/16 [==============================] - 1s 83ms/step - loss: 5.9639 - mse_angle: 91.0492 - val_loss: 5.9907 - val_mse_angle: 94.2361
Epoch 40/50
16/16 [==============================] - 1s 83ms/step - loss: 5.9792 - mse_angle: 88.0059 - val_loss: 5.8827 - val_mse_angle: 94.3056
Epoch 41/50
16/16 [==============================] - 1s 82ms/step - loss: 5.9297 - mse_angle: 92.0566 - val_loss: 5.8013 - val_mse_angle: 94.6319
Epoch 42/50
16/16 [==============================] - 1s 84ms/step - loss: 5.9666 - mse_angle: 88.4168 - val_loss: 5.8768 - val_mse_angle: 99.4826
Epoch 43/50
16/16 [==============================] - 1s 82ms/step - loss: 5.9887 - mse_angle: 90.3191 - val_loss: 5.9197 - val_mse_angle: 96.8611
Epoch 44/50
16/16 [==============================] - 1s 82ms/step - loss: 5.9889 - mse_angle: 87.8867 - val_loss: 5.8738 - val_mse_angle: 96.6875
Epoch 45/50
16/16 [==============================] - 1s 83ms/step - loss: 5.9694 - mse_angle: 92.4437 - val_loss: 5.8639 - val_mse_angle: 98.7222
Epoch 46/50
16/16 [==============================] - 1s 82ms/step - loss: 5.9560 - mse_angle: 89.9125 - val_loss: 5.8387 - val_mse_angle: 82.4965
Epoch 47/50
16/16 [==============================] - 1s 82ms/step - loss: 5.9468 - mse_angle: 89.7066 - val_loss: 5.9525 - val_mse_angle: 87.1632
Epoch 48/50
16/16 [==============================] - 1s 83ms/step - loss: 6.0111 - mse_angle: 89.5977 - val_loss: 5.9091 - val_mse_angle: 96.6875
Epoch 49/50
16/16 [==============================] - 1s 82ms/step - loss: 5.9648 - mse_angle: 89.0430 - val_loss: 5.9656 - val_mse_angle: 92.8368
Epoch 50/50
16/16 [==============================] - 1s 82ms/step - loss: 5.9234 - mse_angle: 91.1891 - val_loss: 5.9717 - val_mse_angle: 99.2257
for image 0 angle: 312, pred: 46
for image 1 angle: 202, pred: 46
for image 2 angle: 235, pred: 46
for image 3 angle: 286, pred: 46
for image 4 angle: 226, pred: 46
for image 5 angle: 76, pred: 46
for image 6 angle: 91, pred: 46
for image 7 angle: 91, pred: 46
for image 8 angle: 97, pred: 46
for image 9 angle: 263, pred: 46
Код: Выделить всё
import numpy as np
from keras import backend as K
from keras.layers.convolutional import Conv2D, MaxPooling2D
from keras.layers import Input, Dense, Flatten
from keras.models import Model
from keras.optimizers import Adam
from keras.preprocessing import image as keras_image
from keras.utils import Sequence
from keras.utils.np_utils import to_categorical
from PIL import Image
import math
from random import randint
import os
from numpy import argmax
from create_text_images import create_data
def get_dataset(directory, name):
"""
Resize the pictures in the directory and return as a numpy array.
"""
X_train = []
for i, img_name in enumerate(os.listdir(directory)):
img_path = os.path.join(directory, img_name)
with Image.open(img_path) as img:
img = img.resize((262, 262))
x = keras_image.img_to_array(img)
X_train.append(x)
X_train = np.array(X_train)
return X_train
def rotate_pictures(X_images):
"""
Randomly rotate the picture, then crop it to size 224x224.
Return the image as x normalized /255
and the rotation (converted to 360 categories) as y.
"""
X_train, y_train = [], []
for i, img in enumerate(X_images):
img = keras_image.array_to_img(img)
rotation = randint(0, 359)
img = img.rotate(rotation, resample=Image.BICUBIC)
w, h = img.size
img = img.crop(((w//2 - 112),(h//2 - 112), (w//2 + 112), (h//2 + 112)))
x = keras_image.img_to_array(img)/255.0
X_train.append(x)
y_train.append(rotation)
y_train = to_categorical(y_train, num_classes=360)
X_train = np.array(X_train)
y_train = np.array(y_train)
return X_train, y_train
class data_generator(Sequence):
"""
On initiation, create x and y data with the rotated pictures and their rotation.
If the dataset is 'train', then rotate original pictures again after every epoch.
"""
def __init__(self, images, name, batch_size):
self.images = images
self.name = name
self.x, self.y = rotate_pictures(self.images)
self.batch_size = batch_size
self.indices = np.arange(self.x.shape[0])
self.on_epoch_end()
def __len__(self):
return math.ceil(self.x.shape[0] / self.batch_size)
def __getitem__(self, idx):
inds = self.indices[idx * self.batch_size:(idx + 1) * self.batch_size]
batch_x = self.x[inds]
batch_y = self.y[inds]
return batch_x, batch_y
def on_epoch_end(self):
if self.name == "train":
self.x, self.y = rotate_pictures(self.images)
def mse_angle(y_true, y_pred):
"""
Calculate the mean difference between the true angles
and the predicted angles. Each angle is represented
as a binary vector.
"""
a = K.argmax(y_true)
b = K.argmax(y_pred)
diff = 180 - abs(abs(a - b) - 180)
return K.mean(K.cast(K.abs(diff), K.floatx()))
train_dir = "train/"
val_dir = "val/"
test_dir = "test/"
number_of_epochs = 50
number_of_classes = 360
input_shape = (224, 224, 3)
activation_fn = 'softmax'
batch_size = 32
create_data(train_dir, 500)
X_train = get_dataset(train_dir, "train")
train_generator = data_generator(X_train, "train", batch_size)
create_data(val_dir, 50)
X_val = get_dataset(val_dir, "val")
val_generator = data_generator(X_val, "val", batch_size)
create_data(test_dir, 10)
X_test = get_dataset(test_dir, "test")
X_test, y_test = rotate_pictures(X_test)
input_tensor = Input(shape=input_shape)
x = Conv2D(32, (3, 3), activation='relu')(input_tensor)
x = MaxPooling2D((2,2), strides=(2,2))(x)
x = Conv2D(64, (3, 3), activation='relu')(x)
x = MaxPooling2D((2,2), strides=(2,2))(x)
x = Conv2D(128, (3, 3), activation='relu')(x)
x = MaxPooling2D((2,2), strides=(2,2))(x)
x = Conv2D(128, (3, 3), activation='relu')(x)
x = MaxPooling2D((2,2), strides=(2,2))(x)
x = Flatten()(x)
x = Dense(512, activation='relu')(x)
output_tensor = Dense(number_of_classes, activation=activation_fn)(x)
model = Model(input_tensor, output_tensor)
model.summary()
model.compile(
loss='categorical_crossentropy',
optimizer=Adam(lr=0.1),
metrics=[mse_angle]
)
history = model.fit(
train_generator,
epochs=number_of_epochs,
validation_data=val_generator
)
model.save_weights('model_weights.h5')
predictions = model.predict(X_test)
for i, prediction in enumerate(predictions):
angle = argmax(y_test[i])
pred = argmax(prediction)
print("for image {0} angle: {1}, pred: {2}".format(i, angle, pred))
Код: Выделить всё
import random
import string
from PIL import Image, ImageDraw, ImageFont
def get_random_string(stringLength):
characters = 10*string.ascii_letters + 100*' ' + string.punctuation*2 + string.digits
return ''.join(random.choice(characters) for i in range(stringLength))
def get_random_text(lines_min, lines_max, char_min, char_max, newline_min, newline_max):
lines = ''
for line in range(random.randint(lines_min, lines_max+1)):
lines += get_random_string(random.randint( char_min, char_max+1))
lines += '\n' * random.randint(newline_min, newline_max+1)
return lines
def create_random_image(directory, file_name, paragraphs_min, paragraphs_max, fontsize_min, fontsize_max,
lines_min, lines_max, char_min, char_max, newline_min, newline_max):
img = Image.new('RGB', (876, 876), color = 'white')
img.alpha_channel = False
d = ImageDraw.Draw(img)
for i in range(random.randint(paragraphs_min, paragraphs_max+1)):
fnt = ImageFont.truetype('Roboto-Black.ttf', random.randint(fontsize_min, fontsize_max+1))
d.text((50,100+random.uniform(300, 500)*i),
get_random_text(lines_min, lines_max, char_min, char_max, newline_min, newline_max),
fill='black', font=fnt)
img.save('{0}/{1}.png'.format(directory, file_name))
def create_data(directory, count):
for i in range(0, count):
create_random_image(directory, i, 3, 6, 30, 70, 1, 3, 10, 100, 1, 3)
Изменение: удалены две неиспользуемые строки кодов.
Подробнее здесь: https://stackoverflow.com/questions/587 ... the-same-c
