NeuralNetwork-Examples
NeuralNetwork-Examples copied to clipboard
Published
20 hours ago •
ieee8023
ieee8023
Unexpected results when transforming from Theano to Keras
Hi, I used Keras to re-implement CountCeption model but got unexpected results and did not know how to fix it. Here is my model:
from keras.layers import Convolution2D, Input, Flatten, Activation
from keras.layers.merge import concatenate
from keras.layers.normalization import BatchNormalization
from keras.models import Model
from keras.layers.convolutional import ZeroPadding2D
from keras import optimizers
from keras.losses import mean_squared_error, mean_absolute_error, categorical_crossentropy
from keras.layers.advanced_activations import LeakyReLU
import keras.backend as K
import numpy as np
class CountCeption(object):
def __init__(self, frame_size, stride, learning_rate = 1e-4, padding=0, ef = 1):
self.frame_size = frame_size
self.stride = stride
self.padding = padding
self.learning_rate = learning_rate
self.ef = ef
self.build_model()
def ConvFactory(self, inp, num_filter, filter_size, pad=0):
inp = ZeroPadding2D(padding=(pad, pad), data_format='channels_last')(inp)
inp = Convolution2D(num_filter, filter_size, strides=(self.stride, self.stride),
padding='valid', data_format='channels_last', kernel_initializer='glorot_uniform')(inp)
inp = LeakyReLU(alpha=0.01)(inp)
return BatchNormalization()(inp)
def Inception(self, inp, first_channel, second_channel):
conv1x1 = self.ConvFactory(inp, num_filter= first_channel, filter_size= 1, pad= 0)
conv3x3 = self.ConvFactory(inp, num_filter= second_channel, filter_size= 3, pad= 1)
return concatenate([conv1x1, conv3x3])
def build_model(self):
self.inp = Input(name='Input', shape= (self.frame_size, self.frame_size,3), dtype='float32')
self.net = self.ConvFactory(self.inp, filter_size=3, num_filter=64, pad=self.padding)
self.net = self.Inception(self.net, 16, 16)
self.net = self.Inception(self.net, 16, 32)
self.net = self.ConvFactory(self.net, filter_size=14, num_filter=16)
self.net = self.Inception(self.net, 112, 48)
self.net = self.Inception(self.net, 64, 32)
self.net = self.Inception(self.net, 40, 40)
self.net = self.Inception(self.net, 32, 96)
self.net = self.ConvFactory(self.net, filter_size=18, num_filter=32)
self.net = self.ConvFactory(self.net, filter_size=1, num_filter=64)
self.net = self.ConvFactory(self.net, filter_size=1, num_filter=64)
self.out = self.ConvFactory(self.net, filter_size=1, num_filter=1)
self.model = Model(inputs=self.inp, outputs=self.out)
self.model.summary()
self.count = K.sum(self.out, axis=(1,2,3))/self.ef
self.predict = K.function([self.inp, K.learning_phase()], [self.out, self.count])
def compile_model(self):
self.label = K.placeholder(ndim= 4, dtype='float32')
self.label_count = K.placeholder(ndim= 2, dtype='float32')
self.loss_pixel = K.mean(K.abs(self.label-self.out))
self.loss_count = K.mean(K.abs(self.label_count - self.count))
self.optimizer = optimizers.Adam(lr=self.learning_rate)
# self.optimizer = optimizers.SGD(lr=self.learning_rate)
self.update = self.optimizer.get_updates(self.model.trainable_weights, [], loss = self.loss_pixel)
self.test_step = K.function([self.inp, self.label, self.label_count, K.learning_phase()], \
[self.loss_pixel, self.loss_count])
self.train_step = K.function([self.inp, self.label, self.label_count, K.learning_phase()], \
[self.loss_pixel, self.loss_count, self.out], updates = self.update)
>> model = CountCeption(256, stride=1, padding=32)
Layer (type) Output Shape Param # Connected to
====================================================================================================
Input (InputLayer) (None, 256, 256, 3) 0
____________________________________________________________________________________________________
zero_padding2d_37 (ZeroPadding2D (None, 320, 320, 3) 0 Input[0][0]
____________________________________________________________________________________________________
conv2d_37 (Conv2D) (None, 318, 318, 64) 1792 zero_padding2d_37[0][0]
____________________________________________________________________________________________________
leaky_re_lu_37 (LeakyReLU) (None, 318, 318, 64) 0 conv2d_37[0][0]
____________________________________________________________________________________________________
batch_normalization_37 (BatchNor (None, 318, 318, 64) 256 leaky_re_lu_37[0][0]
____________________________________________________________________________________________________
zero_padding2d_38 (ZeroPadding2D (None, 318, 318, 64) 0 batch_normalization_37[0][0]
____________________________________________________________________________________________________
zero_padding2d_39 (ZeroPadding2D (None, 320, 320, 64) 0 batch_normalization_37[0][0]
____________________________________________________________________________________________________
conv2d_38 (Conv2D) (None, 318, 318, 16) 1040 zero_padding2d_38[0][0]
____________________________________________________________________________________________________
conv2d_39 (Conv2D) (None, 318, 318, 16) 9232 zero_padding2d_39[0][0]
____________________________________________________________________________________________________
leaky_re_lu_38 (LeakyReLU) (None, 318, 318, 16) 0 conv2d_38[0][0]
____________________________________________________________________________________________________
leaky_re_lu_39 (LeakyReLU) (None, 318, 318, 16) 0 conv2d_39[0][0]
____________________________________________________________________________________________________
batch_normalization_38 (BatchNor (None, 318, 318, 16) 64 leaky_re_lu_38[0][0]
____________________________________________________________________________________________________
batch_normalization_39 (BatchNor (None, 318, 318, 16) 64 leaky_re_lu_39[0][0]
____________________________________________________________________________________________________
concatenate_13 (Concatenate) (None, 318, 318, 32) 0 batch_normalization_38[0][0]
batch_normalization_39[0][0]
____________________________________________________________________________________________________
zero_padding2d_40 (ZeroPadding2D (None, 318, 318, 32) 0 concatenate_13[0][0]
____________________________________________________________________________________________________
zero_padding2d_41 (ZeroPadding2D (None, 320, 320, 32) 0 concatenate_13[0][0]
____________________________________________________________________________________________________
conv2d_40 (Conv2D) (None, 318, 318, 16) 528 zero_padding2d_40[0][0]
____________________________________________________________________________________________________
conv2d_41 (Conv2D) (None, 318, 318, 32) 9248 zero_padding2d_41[0][0]
____________________________________________________________________________________________________
leaky_re_lu_40 (LeakyReLU) (None, 318, 318, 16) 0 conv2d_40[0][0]
____________________________________________________________________________________________________
leaky_re_lu_41 (LeakyReLU) (None, 318, 318, 32) 0 conv2d_41[0][0]
____________________________________________________________________________________________________
batch_normalization_40 (BatchNor (None, 318, 318, 16) 64 leaky_re_lu_40[0][0]
____________________________________________________________________________________________________
batch_normalization_41 (BatchNor (None, 318, 318, 32) 128 leaky_re_lu_41[0][0]
____________________________________________________________________________________________________
concatenate_14 (Concatenate) (None, 318, 318, 48) 0 batch_normalization_40[0][0]
batch_normalization_41[0][0]
____________________________________________________________________________________________________
zero_padding2d_42 (ZeroPadding2D (None, 318, 318, 48) 0 concatenate_14[0][0]
____________________________________________________________________________________________________
conv2d_42 (Conv2D) (None, 305, 305, 16) 150544 zero_padding2d_42[0][0]
____________________________________________________________________________________________________
leaky_re_lu_42 (LeakyReLU) (None, 305, 305, 16) 0 conv2d_42[0][0]
____________________________________________________________________________________________________
batch_normalization_42 (BatchNor (None, 305, 305, 16) 64 leaky_re_lu_42[0][0]
____________________________________________________________________________________________________
zero_padding2d_43 (ZeroPadding2D (None, 305, 305, 16) 0 batch_normalization_42[0][0]
____________________________________________________________________________________________________
zero_padding2d_44 (ZeroPadding2D (None, 307, 307, 16) 0 batch_normalization_42[0][0]
____________________________________________________________________________________________________
conv2d_43 (Conv2D) (None, 305, 305, 112) 1904 zero_padding2d_43[0][0]
____________________________________________________________________________________________________
conv2d_44 (Conv2D) (None, 305, 305, 48) 6960 zero_padding2d_44[0][0]
____________________________________________________________________________________________________
leaky_re_lu_43 (LeakyReLU) (None, 305, 305, 112) 0 conv2d_43[0][0]
____________________________________________________________________________________________________
leaky_re_lu_44 (LeakyReLU) (None, 305, 305, 48) 0 conv2d_44[0][0]
____________________________________________________________________________________________________
batch_normalization_43 (BatchNor (None, 305, 305, 112) 448 leaky_re_lu_43[0][0]
____________________________________________________________________________________________________
batch_normalization_44 (BatchNor (None, 305, 305, 48) 192 leaky_re_lu_44[0][0]
____________________________________________________________________________________________________
concatenate_15 (Concatenate) (None, 305, 305, 160) 0 batch_normalization_43[0][0]
batch_normalization_44[0][0]
____________________________________________________________________________________________________
zero_padding2d_45 (ZeroPadding2D (None, 305, 305, 160) 0 concatenate_15[0][0]
____________________________________________________________________________________________________
zero_padding2d_46 (ZeroPadding2D (None, 307, 307, 160) 0 concatenate_15[0][0]
____________________________________________________________________________________________________
conv2d_45 (Conv2D) (None, 305, 305, 64) 10304 zero_padding2d_45[0][0]
____________________________________________________________________________________________________
conv2d_46 (Conv2D) (None, 305, 305, 32) 46112 zero_padding2d_46[0][0]
____________________________________________________________________________________________________
leaky_re_lu_45 (LeakyReLU) (None, 305, 305, 64) 0 conv2d_45[0][0]
____________________________________________________________________________________________________
leaky_re_lu_46 (LeakyReLU) (None, 305, 305, 32) 0 conv2d_46[0][0]
____________________________________________________________________________________________________
batch_normalization_45 (BatchNor (None, 305, 305, 64) 256 leaky_re_lu_45[0][0]
____________________________________________________________________________________________________
batch_normalization_46 (BatchNor (None, 305, 305, 32) 128 leaky_re_lu_46[0][0]
____________________________________________________________________________________________________
concatenate_16 (Concatenate) (None, 305, 305, 96) 0 batch_normalization_45[0][0]
batch_normalization_46[0][0]
____________________________________________________________________________________________________
zero_padding2d_47 (ZeroPadding2D (None, 305, 305, 96) 0 concatenate_16[0][0]
____________________________________________________________________________________________________
zero_padding2d_48 (ZeroPadding2D (None, 307, 307, 96) 0 concatenate_16[0][0]
____________________________________________________________________________________________________
conv2d_47 (Conv2D) (None, 305, 305, 40) 3880 zero_padding2d_47[0][0]
____________________________________________________________________________________________________
conv2d_48 (Conv2D) (None, 305, 305, 40) 34600 zero_padding2d_48[0][0]
____________________________________________________________________________________________________
leaky_re_lu_47 (LeakyReLU) (None, 305, 305, 40) 0 conv2d_47[0][0]
____________________________________________________________________________________________________
leaky_re_lu_48 (LeakyReLU) (None, 305, 305, 40) 0 conv2d_48[0][0]
____________________________________________________________________________________________________
batch_normalization_47 (BatchNor (None, 305, 305, 40) 160 leaky_re_lu_47[0][0]
____________________________________________________________________________________________________
batch_normalization_48 (BatchNor (None, 305, 305, 40) 160 leaky_re_lu_48[0][0]
____________________________________________________________________________________________________
concatenate_17 (Concatenate) (None, 305, 305, 80) 0 batch_normalization_47[0][0]
batch_normalization_48[0][0]
____________________________________________________________________________________________________
zero_padding2d_49 (ZeroPadding2D (None, 305, 305, 80) 0 concatenate_17[0][0]
____________________________________________________________________________________________________
zero_padding2d_50 (ZeroPadding2D (None, 307, 307, 80) 0 concatenate_17[0][0]
____________________________________________________________________________________________________
conv2d_49 (Conv2D) (None, 305, 305, 32) 2592 zero_padding2d_49[0][0]
____________________________________________________________________________________________________
conv2d_50 (Conv2D) (None, 305, 305, 96) 69216 zero_padding2d_50[0][0]
____________________________________________________________________________________________________
leaky_re_lu_49 (LeakyReLU) (None, 305, 305, 32) 0 conv2d_49[0][0]
____________________________________________________________________________________________________
leaky_re_lu_50 (LeakyReLU) (None, 305, 305, 96) 0 conv2d_50[0][0]
____________________________________________________________________________________________________
batch_normalization_49 (BatchNor (None, 305, 305, 32) 128 leaky_re_lu_49[0][0]
____________________________________________________________________________________________________
batch_normalization_50 (BatchNor (None, 305, 305, 96) 384 leaky_re_lu_50[0][0]
____________________________________________________________________________________________________
concatenate_18 (Concatenate) (None, 305, 305, 128) 0 batch_normalization_49[0][0]
batch_normalization_50[0][0]
____________________________________________________________________________________________________
zero_padding2d_51 (ZeroPadding2D (None, 305, 305, 128) 0 concatenate_18[0][0]
____________________________________________________________________________________________________
conv2d_51 (Conv2D) (None, 288, 288, 32) 1327136 zero_padding2d_51[0][0]
____________________________________________________________________________________________________
leaky_re_lu_51 (LeakyReLU) (None, 288, 288, 32) 0 conv2d_51[0][0]
____________________________________________________________________________________________________
batch_normalization_51 (BatchNor (None, 288, 288, 32) 128 leaky_re_lu_51[0][0]
____________________________________________________________________________________________________
zero_padding2d_52 (ZeroPadding2D (None, 288, 288, 32) 0 batch_normalization_51[0][0]
____________________________________________________________________________________________________
conv2d_52 (Conv2D) (None, 288, 288, 64) 2112 zero_padding2d_52[0][0]
____________________________________________________________________________________________________
leaky_re_lu_52 (LeakyReLU) (None, 288, 288, 64) 0 conv2d_52[0][0]
____________________________________________________________________________________________________
batch_normalization_52 (BatchNor (None, 288, 288, 64) 256 leaky_re_lu_52[0][0]
____________________________________________________________________________________________________
zero_padding2d_53 (ZeroPadding2D (None, 288, 288, 64) 0 batch_normalization_52[0][0]
____________________________________________________________________________________________________
conv2d_53 (Conv2D) (None, 288, 288, 64) 4160 zero_padding2d_53[0][0]
____________________________________________________________________________________________________
leaky_re_lu_53 (LeakyReLU) (None, 288, 288, 64) 0 conv2d_53[0][0]
____________________________________________________________________________________________________
batch_normalization_53 (BatchNor (None, 288, 288, 64) 256 leaky_re_lu_53[0][0]
____________________________________________________________________________________________________
zero_padding2d_54 (ZeroPadding2D (None, 288, 288, 64) 0 batch_normalization_53[0][0]
____________________________________________________________________________________________________
conv2d_54 (Conv2D) (None, 288, 288, 1) 65 zero_padding2d_54[0][0]
____________________________________________________________________________________________________
leaky_re_lu_54 (LeakyReLU) (None, 288, 288, 1) 0 conv2d_54[0][0]
____________________________________________________________________________________________________
batch_normalization_54 (BatchNor (None, 288, 288, 1) 4 leaky_re_lu_54[0][0]
====================================================================================================
Total params: 1,684,565
Trainable params: 1,682,995
Non-trainable params: 1,570
_______________________________
And here come the losses:
Found device 0 with properties:
name: GeForce GTX TITAN Z
major: 3 minor: 5 memoryClockRate (GHz) 0.8755
pciBusID 0000:84:00.0
Total memory: 5.94GiB
Free memory: 5.86GiB
2017-10-12 21:14:52.882773: I tensorflow/core/common_runtime/gpu/gpu_device.cc:961] DMA: 0
2017-10-12 21:14:52.882787: I tensorflow/core/common_runtime/gpu/gpu_device.cc:971] 0: Y
2017-10-12 21:14:52.882805: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1030] Creating TensorFlow device (/gpu:0) -> (device: 0, name: GeForce GTX TITAN Z, pci bus id: 0000:84:00.0) Weights loaded!
2017-10-12 21:15:04.972817: I tensorflow/core/common_runtime/gpu/pool_allocator.cc:247] PoolAllocator: After 2024 get requests, put_count=1992 evicted_count=1000 eviction_rate=0.502008 and unsatisfied allocation rate=0.559289
2017-10-12 21:15:04.972892: I tensorflow/core/common_runtime/gpu/pool_allocator.cc:259] Raising pool_size_limit_ from 100 to 110
EP 0/100 | TRAIN_LOSS: 2.074034, 167.909124 | VALID_LOSS: 0.000000, 0.000000
EP 5/100 | TRAIN_LOSS: 1.574912, 165.120924 | VALID_LOSS: 0.000000, 0.000000
EP 10/100 | TRAIN_LOSS: 0.81238, 164.328942 | VALID_LOSS: 0.000000, 0.000000
.
.
.
EP 80/100 | TRAIN_LOSS: 0.24034, 26.891274 | VALID_LOSS: 0.000000, 0.000000
EP 85/100 | TRAIN_LOSS: 0.23912, 26.889122 | VALID_LOSS: 0.000000, 0.000000
EP 90/100 | TRAIN_LOSS: 0.23120, 26.871982 | VALID_LOSS: 0.000000, 0.000000
The valid loss is 0 because I do not calculate it (for timing purpose). Actually I calculated all the losses at the first time I ran the code and I saw that the valid loss decreased as low as the training loss.
When I test the model, it turns out that it always predict the same count (160) although the heat maps returned are different and quite closed to the label.
Could anyone please explain what is wrong with my code :D
