1
我已經創建了一個模擬CNN，我試圖在視頻數據集上使用。 我將測試數據設置爲所有幀的全部單個圖像以獲得正面示例，將0設置爲負面示例。我認爲這會很快學會。但它根本不動。 在Windows 10 64bit上使用當前版本的Keras & Tensorflow。CNN學習停滯
第一個問題，我的邏輯錯了嗎？我應該期望這個測試數據的學習能夠快速達到高精度嗎？
我的模型或參數有什麼問題嗎？我一直在嘗試一些改變，但仍然遇到同樣的問題。
樣本量（56）是否太小？
# testing feature extraction model.
import time
import numpy as np, cv2
import sys
import os
import keras
import tensorflow as tf
from keras.models import Sequential
from keras.layers import Dense, Dropout, Activation, Flatten, BatchNormalization
from keras.layers import Conv3D, MaxPooling3D
from keras.optimizers import SGD,rmsprop, adam
from keras import regularizers
from keras.initializers import Constant
from keras.models import Model
#set gpu options
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=.99, allocator_type = 'BFC')
sess = tf.Session(config=tf.ConfigProto(log_device_placement=True, gpu_options=gpu_options))
config = tf.ConfigProto()
batch_size = 5
num_classes = 1
epochs = 50
nvideos = 56
nframes = 55
nchan = 3
nrows = 480
ncols = 640
#load any single image, resize if needed
img = cv2.imread('C:\\Users\\david\\Documents\\AutonomousSS\\single frame.jpg',cv2.IMREAD_COLOR)
img = cv2.resize(img,(640,480))
x_learn = np.random.randint(0,255,(nvideos,nframes,nrows,ncols,nchan),dtype=np.uint8)
y_learn = np.array([[1],[1],[1],[0],[1],[0],[1],[0],[1],[0],
[1],[0],[0],[1],[0],[0],[1],[0],[1],[0],
[1],[0],[1],[1],[0],[1],[0],[0],[1],[1],
[1],[0],[1],[0],[1],[0],[1],[0],[1],[0],
[0],[1],[0],[0],[1],[0],[1],[0],[1],[0],
[1],[1],[0],[1],[0],[0]],np.uint8)
#each sample, each frame is either the single image for postive examples or 0 for negative examples.
for i in range (nvideos):
if y_learn[i] == 0 :
x_learn[i]=0
else:
x_learn[i,:nframes]=img
#build model
m_loss = 'mean_squared_error'
m_opt = SGD(lr=0.001, decay=1e-6, momentum=0.9, nesterov=True)
m_met = 'acc'
model = Sequential()
# 1st layer group
model.add(Conv3D(32, (3, 3,3), activation='relu',padding="same", name="conv1a", strides=(3, 3, 3),
kernel_initializer = 'glorot_normal',
trainable=False,
input_shape=(nframes,nrows,ncols,nchan)))
#model.add(BatchNormalization(axis=1))
model.add(Conv3D(32, (3, 3, 3), trainable=False, strides=(1, 1, 1), padding="same", name="conv1b", activation="relu"))
#model.add(BatchNormalization(axis=1))
model.add(MaxPooling3D(padding="valid", trainable=False, pool_size=(1, 5, 5), name="pool1", strides=(2, 2, 2)))
# 2nd layer group
model.add(Conv3D(128, (3, 3, 3), trainable=False, strides=(1, 1, 1), padding="same", name="conv2a", activation="relu"))
model.add(Conv3D(128, (3, 3, 3), trainable=False, strides=(1, 1, 1), padding="same", name="conv2b", activation="relu"))
#model.add(BatchNormalization(axis=1))
model.add(MaxPooling3D(padding="valid", trainable=False, pool_size=(1, 5, 5), name="pool2", strides=(2, 2, 2)))
# 3rd layer group
model.add(Conv3D(256, (3, 3, 3), trainable=False, strides=(1, 1, 1), padding="same", name="conv3a", activation="relu"))
model.add(Conv3D(256, (3, 3, 3), trainable=False, strides=(1, 1, 1), padding="same", name="conv3b", activation="relu"))
#model.add(BatchNormalization(axis=1))
model.add(MaxPooling3D(padding="valid", trainable=False, pool_size=(1, 5, 5), name="pool3", strides=(2, 2, 2)))
# 4th layer group
model.add(Conv3D(512, (3, 3, 3), trainable=False, strides=(1, 1, 1), padding="same", name="conv4a", activation="relu"))
model.add(Conv3D(512, (3, 3, 3), trainable=False, strides=(1, 1, 1), padding="same", name="conv4b", activation="relu"))
#model.add(BatchNormalization(axis=1))
model.add(MaxPooling3D(padding="valid", trainable=False, pool_size=(1, 5, 5), name="pool4", strides=(2, 2, 2)))
model.add(Flatten(name='flatten',trainable=False))
model.add(Dense(512,activation='relu', trainable=True,name='den0'))
model.add(Dense(num_classes,activation='softmax',name='den1'))
print (model.summary())
#compile model
model.compile(loss=m_loss,
optimizer=m_opt,
metrics=[m_met])
print ('compiled')
#set callbacks
from keras import backend as K
K.set_learning_phase(0) #set learning phase
tb = keras.callbacks.TensorBoard(log_dir=sample_root_path+'logs', histogram_freq=0,
write_graph=True, write_images=False)
tb.set_model(model)
reduce_lr = keras.callbacks.ReduceLROnPlateau(monitor='loss', factor=0.2,verbose=1,
patience=2, min_lr=0.000001)
reduce_lr.set_model(model)
ear_stop = keras.callbacks.EarlyStopping(monitor='loss', min_delta=0, patience=4, verbose=1, mode='auto')
ear_stop.set_model(model)
#fit
history = model.fit(x_learn, y_learn,
batch_size=batch_size,
callbacks=[reduce_lr,tb, ear_stop],
verbose=1,
validation_split=0.1,
shuffle = True,
epochs=epochs)
score = model.evaluate(x_learn, y_learn, batch_size=batch_size)
print(str(model.metrics_names) + ": " + str(score))
像往常一樣，感謝您的任何和所有幫助。
相加的輸出...
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv1a (Conv3D) (None, 19, 160, 214, 32) 2624
_________________________________________________________________
conv1b (Conv3D) (None, 19, 160, 214, 32) 27680
_________________________________________________________________
pool1 (MaxPooling3D) (None, 10, 78, 105, 32) 0
_________________________________________________________________
conv2a (Conv3D) (None, 10, 78, 105, 128) 110720
_________________________________________________________________
conv2b (Conv3D) (None, 10, 78, 105, 128) 442496
_________________________________________________________________
pool2 (MaxPooling3D) (None, 5, 37, 51, 128) 0
_________________________________________________________________
conv3a (Conv3D) (None, 5, 37, 51, 256) 884992
_________________________________________________________________
conv3b (Conv3D) (None, 5, 37, 51, 256) 1769728
_________________________________________________________________
pool3 (MaxPooling3D) (None, 3, 17, 24, 256) 0
_________________________________________________________________
conv4a (Conv3D) (None, 3, 17, 24, 512) 3539456
_________________________________________________________________
conv4b (Conv3D) (None, 3, 17, 24, 512) 7078400
_________________________________________________________________
pool4 (MaxPooling3D) (None, 2, 7, 10, 512) 0
_________________________________________________________________
flatten (Flatten) (None, 71680) 0
_________________________________________________________________
den0 (Dense) (None, 512) 36700672
_________________________________________________________________
den1 (Dense) (None, 1) 513
=================================================================
Total params: 50,557,281
Trainable params: 36,701,185
Non-trainable params: 13,856,096
_________________________________________________________________
None
compiled
Train on 50 samples, validate on 6 samples
Epoch 1/50
50/50 [==============================] - 20s - loss: 0.5000 - acc: 0.5000 - val_loss: 0.5000 - val_acc: 0.5000
Epoch 2/50
50/50 [==============================] - 16s - loss: 0.5000 - acc: 0.5000 - val_loss: 0.5000 - val_acc: 0.5000
Epoch 3/50
50/50 [==============================] - 16s - loss: 0.5000 - acc: 0.5000 - val_loss: 0.5000 - val_acc: 0.5000
Epoch 4/50
45/50 [==========================>...] - ETA: 1s - loss: 0.5111 - acc: 0.4889
Epoch 00003: reducing learning rate to 0.00020000000949949026.
50/50 [==============================] - 16s - loss: 0.5000 - acc: 0.5000 - val_loss: 0.5000 - val_acc: 0.5000
Epoch 5/50
50/50 [==============================] - 16s - loss: 0.5000 - acc: 0.5000 - val_loss: 0.5000 - val_acc: 0.5000
Epoch 6/50
45/50 [==========================>...] - ETA: 1s - loss: 0.5111 - acc: 0.4889
Epoch 00005: reducing learning rate to 4.0000001899898055e-05.
50/50 [==============================] - 16s - loss: 0.5000 - acc: 0.5000 - val_loss: 0.5000 - val_acc: 0.5000
Epoch 7/50
50/50 [==============================] - 16s - loss: 0.5000 - acc: 0.5000 - val_loss: 0.5000 - val_acc: 0.5000
Epoch 8/50
45/50 [==========================>...] - ETA: 1s - loss: 0.4889 - acc: 0.5111
Epoch 00007: reducing learning rate to 8.000000525498762e-06.
50/50 [==============================] - 16s - loss: 0.5000 - acc: 0.5000 - val_loss: 0.5000 - val_acc: 0.5000
Epoch 9/50
50/50 [==============================] - 16s - loss: 0.5000 - acc: 0.5000 - val_loss: 0.5000 - val_acc: 0.5000
Epoch 00008: early stopping
56/56 [==============================] - 12s
['loss', 'acc']: [0.50000001516725334, 0.5000000127724239]
您能否提供關於您的目標的更多細節，您想要訓練的最終數據，特別是您爲什麼要嘗試在單個圖像上進行訓練？並且，由於您將所有圖層設置爲不可訓練（除了最後一個密集圖層）：您是否加載了任何預訓練的重量？我沒有看到你導入像VGG或Inception這樣的Keras應用程序，否則就加載任何權重。 – petezurich
最終目標是訓練某一動作。這是幀之間的一系列移動。上面的測試只是一個測試例子。我得到相同的行爲，正面的例子是1）所有幀1，2）幀是隨機的，3）幀是真實的視頻序列。我不認爲在Keras應用程序中建立將會有所幫助。 – DSP209
好的，謝謝澄清一些觀點。據我瞭解，你想調整預訓模型。如果是這樣的話：你如何加載重量？凱拉斯應用程序正是一種很好的方法。例如，參見Keras的這個教程：https：//blog.keras.io/building-powerful-image-classification-models-using-very-little-data.html否則你的網絡不會學到任何東西，因爲幾乎所有的圖層都是設置爲不可訓練。從你的代碼：'... trainable = False ...' – petezurich