Tensor 및 훈련 변수 접근하기

- Tensorflow 주요 기능 정리 

: filed name을 지정하지 않고 network를 구성하고

훈련 파라미터들과 주요 tensor 접근 및 데이터 출력해보기

1. 입력 채널 : 1

2. conolution layer[11X11] output channels 10

3. Bias

4. convolution layer[5X5] output channels 20

5. convolution layer[3X3] output channels 50

import tensorflow as tf
import numpy as np
net_inputs = tf.placeholder(tf.float32,[None, None, None,1])
net_ouputs = tf.placeholder(tf.float32,[None, None, None,1])

# input: A Tensor. Must be one of the following types: half, bfloat16, float32, float64. A 4-D tensor. The dimension order is interpreted according to the value of data_format, see below for details.
# filter: A Tensor. Must have the same type as input. A 4-D tensor of shape [filter_height, filter_width, in_channels, out_channels]
# strides: A list of ints. 1-D tensor of length 4. The stride of the sliding window for each dimension of input. The dimension order is determined by the value of data_format, see below for details.
# padding: A string from: "SAME", "VALID". The type of padding algorithm to use.
# use_cudnn_on_gpu: An optional bool. Defaults to True.
# data_format: An optional string from: "NHWC", "NCHW". Defaults to "NHWC". Specify the data format of the input and output data. With the default format "NHWC", the data is stored in the order of: [batch, height, width, channels]. Alternatively, the format could be "NCHW", the data storage order of: [batch, channels, height, width].
# dilations: An optional list of ints. Defaults to [1, 1, 1, 1]. 1-D tensor of length 4. The dilation factor for each dimension of input. If set to k > 1, there will be k-1 skipped cells between each filter element on that dimension. The dimension order is determined by the value of data_format, see above for details. Dilations in the batch and depth dimensions must be 1.
# name: A name for the operation (optional).

# 1st layer : W*X + B
W = tf.Variable( tf.random_normal([11, 11, 1, 10]))
B = tf.Variable( tf.random_normal([10]))
net = tf.nn.conv2d(net_inputs, filter=W, strides = [1,4,4,1], padding='SAME')
net = tf.nn.bias_add(net, B)
net = tf.nn.relu(net)

W = tf.Variable( tf.random_normal([5, 5, 10, 20]))


# 2nd layer : W*X
net = tf.nn.conv2d(net, filter=W, strides = [1,2,2,1], padding='SAME')
net = tf.nn.relu(net)


# 3rd layer : W*X
W = tf.Variable( tf.random_normal([3, 3, 20, 50]))

net = tf.nn.conv2d(net, filter=W, strides = [1,2,2,1], padding='SAME')
net = tf.nn.relu(net)

1. 훈련 파라미터 확인 : Convolution filter와 bias 

코드:

print('trainbale weight 확인')
train_var = [var for var in tf.trainable_variables() ]
for tvar in train_var:
    print(tvar)

결과 : 

2. 모든 tensor 확인 :

- field name을 별도로 지정하지 않았으므로 모든 tensor를 로드하고 이 후 원하는(convolution, bias)layer만

추출해본다.

- name을 별도로 지정했을 경우 " tf.get_default_graph().get_tensor_by_name()"을 이용하여 tensor를 로드가능

코드 : 

# tensor 확인하기
print('tensor 확인')
# tensors = [ tf.get_default_graph().get_tensor_by_name(tensor.name) for tensor in tf.global_variables()]
alltensors = [ tensor.values()[0] for tensor in tf.get_default_graph().get_operations()]
# print(tensors)
for tnsr in alltensors:
    print(tnsr)

결과 :

tensor 확인
Tensor("Placeholder:0", shape=(?, ?, ?, 1), dtype=float32)
Tensor("Placeholder_1:0", shape=(?, ?, ?, 1), dtype=float32)
Tensor("random_normal/shape:0", shape=(4,), dtype=int32)
Tensor("random_normal/mean:0", shape=(), dtype=float32)
Tensor("random_normal/stddev:0", shape=(), dtype=float32)
Tensor("random_normal/RandomStandardNormal:0", shape=(11, 11, 1, 10), dtype=float32)
Tensor("random_normal/mul:0", shape=(11, 11, 1, 10), dtype=float32)
Tensor("random_normal:0", shape=(11, 11, 1, 10), dtype=float32)
Tensor("Variable:0", shape=(11, 11, 1, 10), dtype=float32_ref)
Tensor("Variable/Assign:0", shape=(11, 11, 1, 10), dtype=float32_ref)
Tensor("Variable/read:0", shape=(11, 11, 1, 10), dtype=float32)
Tensor("random_normal_1/shape:0", shape=(1,), dtype=int32)
Tensor("random_normal_1/mean:0", shape=(), dtype=float32)
Tensor("random_normal_1/stddev:0", shape=(), dtype=float32)
Tensor("random_normal_1/RandomStandardNormal:0", shape=(10,), dtype=float32)
Tensor("random_normal_1/mul:0", shape=(10,), dtype=float32)
Tensor("random_normal_1:0", shape=(10,), dtype=float32)
Tensor("Variable_1:0", shape=(10,), dtype=float32_ref)
Tensor("Variable_1/Assign:0", shape=(10,), dtype=float32_ref)
Tensor("Variable_1/read:0", shape=(10,), dtype=float32)
Tensor("Conv2D:0", shape=(?, ?, ?, 10), dtype=float32)
Tensor("BiasAdd:0", shape=(?, ?, ?, 10), dtype=float32)
Tensor("Relu:0", shape=(?, ?, ?, 10), dtype=float32)
Tensor("random_normal_2/shape:0", shape=(4,), dtype=int32)
Tensor("random_normal_2/mean:0", shape=(), dtype=float32)
Tensor("random_normal_2/stddev:0", shape=(), dtype=float32)
Tensor("random_normal_2/RandomStandardNormal:0", shape=(5, 5, 10, 20), dtype=float32)
Tensor("random_normal_2/mul:0", shape=(5, 5, 10, 20), dtype=float32)
Tensor("random_normal_2:0", shape=(5, 5, 10, 20), dtype=float32)
Tensor("Variable_2:0", shape=(5, 5, 10, 20), dtype=float32_ref)
Tensor("Variable_2/Assign:0", shape=(5, 5, 10, 20), dtype=float32_ref)
Tensor("Variable_2/read:0", shape=(5, 5, 10, 20), dtype=float32)
Tensor("Conv2D_1:0", shape=(?, ?, ?, 20), dtype=float32)
Tensor("Relu_1:0", shape=(?, ?, ?, 20), dtype=float32)
Tensor("random_normal_3/shape:0", shape=(4,), dtype=int32)
Tensor("random_normal_3/mean:0", shape=(), dtype=float32)
Tensor("random_normal_3/stddev:0", shape=(), dtype=float32)
Tensor("random_normal_3/RandomStandardNormal:0", shape=(3, 3, 20, 50), dtype=float32)
Tensor("random_normal_3/mul:0", shape=(3, 3, 20, 50), dtype=float32)
Tensor("random_normal_3:0", shape=(3, 3, 20, 50), dtype=float32)
Tensor("Variable_3:0", shape=(3, 3, 20, 50), dtype=float32_ref)
Tensor("Variable_3/Assign:0", shape=(3, 3, 20, 50), dtype=float32_ref)
Tensor("Variable_3/read:0", shape=(3, 3, 20, 50), dtype=float32)
Tensor("Conv2D_2:0", shape=(?, ?, ?, 50), dtype=float32)
Tensor("Relu_2:0", shape=(?, ?, ?, 50), dtype=float32)

3. Convolution / Bias Tensor 확인

코드 : 

convtesnor = []
for tnsr in alltensors:
    # print(tnsr)
    if( 'Conv2D' in tnsr.name or 'BiasAdd' in tnsr.name ):
        convtesnor.append( tnsr )
print('convolution tensor')
for tnsr in convtesnor:
    # print(tf.get_default_graph().get_tensor_by_name(tnsr))
    print(tnsr)

결과 :

convolution tensor
Tensor("Conv2D:0", shape=(?, ?, ?, 10), dtype=float32)
Tensor("BiasAdd:0", shape=(?, ?, ?, 10), dtype=float32)
Tensor("Conv2D_1:0", shape=(?, ?, ?, 20), dtype=float32)
Tensor("Conv2D_2:0", shape=(?, ?, ?, 50), dtype=float32)

4. session 구동시켜서 각 Layer들의 state 확인

 : 입력을 200X200X1 로 설정

코드 :

inputs = np.random.randn(1,200,200,1)
#
sess = tf.Session()
sess.run( tf.global_variables_initializer())

for tnsr in convtesnor:
    print(tnsr)
    currrent_graph = sess.run( tnsr, feed_dict={net_inputs:inputs})
    print( currrent_graph.shape )

결과 :

Tensor("Conv2D:0", shape=(?, ?, ?, 10), dtype=float32)
(1, 50, 50, 10)
Tensor("BiasAdd:0", shape=(?, ?, ?, 10), dtype=float32)
(1, 50, 50, 10)
Tensor("Conv2D_1:0", shape=(?, ?, ?, 20), dtype=float32)
(1, 25, 25, 20)
Tensor("Conv2D_2:0", shape=(?, ?, ?, 50), dtype=float32)
(1, 13, 13, 50)