cs20_5-2
1. Assignment_1
1.1 1d解答
-
tf.where()
-
我的实践代码:
def test_4_2(): x = tf.constant([[True, False], [True, False]]) x_where = tf.where(x) # Returns locations(coordinates ) of true values in a boolean tensor. # x中有两个True,所以返回了2个1-d tensor [],[] # 先确定几个True以确定返回几个坐标 # 又因为x是2-d tensor, 所以定位True elem也要2-d indeies(行和列),即[0,0], [1,0] # 再确定坐标是什么 print(sess.run(x_where)) y = tf.constant( [ [ [True, False],[True, False] ], [ [False, True],[False, True] ], [ [False, False],[False, True] ] ] ) y_where = tf.where(y) # 5个True elme,所以返回5个坐标;又因为y是3-dim,所以坐标是 3-d tensor(x,y,z) print(sess.run(y_where)) # [0,0,0], [0,1,0], [1,0,1], [1,1,1], [2,1,1] test_4_2()
-
参考:
[1] https://blog.csdn.net/A_a_ron/article/details/79048446 (讲得不错)
[2] https://blog.csdn.net/ustbbsy/article/details/79564828
[3] http://www.tensorfly.cn/tfdoc/api_docs/python/math_ops.html#where (还不错的第三方文档)
[4] https://www.tensorflow.org/api_docs/python/tf/where (tf官方文档基本没有用,没有例子只有干巴巴的文字解释,晦涩难懂)
-
-
tf.gater()
-
我的示例代码
def test_4_3(): temp = tf.range(0, 10) * 10 + tf.constant(1, shape=[10]) temp2 = tf.gather(temp, [1, 5, 9]) # 相当于根据indeies提取元素 with tf.Session() as sess: print(sess.run(temp)) print(sess.run(temp2)) test_4_3() def test_4_4(): a = tf.Variable([[1, 2, 3, 4, 5], [6, 7, 8, 9, 10], [11, 12, 13, 14, 15]]) index_a = tf.Variable([0, 2]) b = tf.Variable([1, 2, 3, 4, 5, 6, 7, 8, 9, 10]) index_b = tf.Variable([2, 4, 6, 8]) with tf.Session() as sess: sess.run(tf.global_variables_initializer()) # 变量要先初始化 print(sess.run(tf.gather(a, index_a))) print(sess.run(tf.gather(b, index_b))) # [[ 1 2 3 4 5],[11 12 13 14 15]] # [3 5 7 9] test_4_4()
-
参考:
[1] https://blog.csdn.net/guotong1988/article/details/53172882
-
-
tf.greater
-
我的示例代码
# 判断函数。首先张量x和张量y的尺寸要相同,输出的tf.greater(x, y)也是一个和x,y尺寸相同的张量。如果x的某个元素比y中对应位置的元素大,则tf.greater(x, y)对应位置返回True,否则返回False。与此类似的函数还有tf.greater_equal def test_4_5(): x = tf.Variable([[1, 2, 3], [6, 7, 8], [11, 12, 13]]) y = tf.Variable([[0, 1, 2], [5, 6, 7], [10, 11, 12]]) x1 = tf.Variable([[1, 2, 3], [6, 7, 8], [11, 12, 13]]) y1 = tf.Variable([[10, 1, 2], [15, 6, 7], [10, 21, 12]]) with tf.Session() as sess: sess.run(tf.global_variables_initializer()) print(sess.run(tf.greater(x, y))) print(sess.run(tf.greater(x1, y1))) # [[ True True True],[ True True True],[ True True True]] # [[False True True],[False True True],[ True False True]] test_4_5()
-
参考:
[1] https://blog.csdn.net/Cyiano/article/details/76087747 (例子简单易懂)
-
-
解决A1_1d:
def test_4(): x = tf.constant([29.05088806, 27.61298943, 31.19073486, 29.35532951, 30.97266006, 26.67541885, 38.08450317, 20.74983215, 34.94445419, 34.45999146, 29.06485367, 36.01657104, 27.88236427, 20.56035233, 30.20379066, 29.51215172, 33.71149445, 28.59134293, 36.05556488, 28.66994858], name="x") y = tf.ones(shape=x.shape, dtype=tf.float32) # 构造与x相同shape的tensor y = tf.multiply(y, tf.constant(30.0)) print("y: ", sess.run(y)) x_bool = tf.greater(x,y,name="x_bool") # 获得一个bool tensor x_index = tf.where(x_bool) # 获得 x>30的elme index out = tf.gather(x,x_index) # 根据x>30的indexies再从x中取elems print("test4-x: ", sess.run(x)) print("test4-x>30: ", sess.run(out)) test_4()
-
拓展:tf其他几个常见函数:
-
tf.concat
# 把一组向量从某一维上拼接起来,很向numpy中的Concatenate,官网例子: t1 = [[1, 2, 3], [4, 5, 6]] t2 = [[7, 8, 9], [10, 11, 12]] tf.concat([t1, t2], 0) ==> [[1, 2, 3], [4, 5, 6], [7, 8, 9], [10, 11, 12]] tf.concat([t1, t2], 1) ==> [[1, 2, 3, 7, 8, 9], [4, 5, 6, 10, 11, 12]] # tensor t3 with shape [2, 3] # tensor t4 with shape [2, 3] tf.shape(tf.concat([t3, t4], 0)) ==> [4, 3] tf.shape(tf.concat([t3, t4], 1)) ==> [2, 6] # 其实,如果是list类型的话也是可以的,只要是形似Tensor,最后tf.concat返回的还是Tensor类型
-
tf.gather_nd
# tf.gather(): 类似于数组的索引,可以把向量中某些索引值提取出来,得到新的向量,适用于要提取的索引为不连续的情况。这个函数似乎只适合在一维的情况下使用。 # gather_nd同上,但允许在多维上进行索引,例子只展示了一种很简单的用法,更复杂的用法可见官网。 import tensorflow as tf a = tf.Variable([[1,2,3,4,5], [6,7,8,9,10], [11,12,13,14,15]]) index_a = tf.Variable([[0,2], [0,4], [2,2]]) with tf.Session() as sess: sess.run(tf.global_variables_initializer()) print(sess.run(tf.gather_nd(a, index_a))) # [ 3 5 13]
-
tf.cast
# 转换数据类型。 a = tf.constant([0, 2, 0, 4, 2, 2], dtype='int32') print(a) # <tf.Tensor 'Const_1:0' shape=(6,) dtype=int32> b = tf.cast(a, 'float32') print(b) # <tf.Tensor 'Cast:0' shape=(6,) dtype=float32>
-
tf.expand_dims & tf.squeeze
# 增加 / 压缩张量的维度。 a = tf.constant([0, 2, 0, 4, 2, 2], dtype='int32') print(a) # <tf.Tensor 'Const_1:0' shape=(6,) dtype=int32> b = tf.expand_dims(a, 0) print(b) # <tf.Tensor 'ExpandDims:0' shape=(1, 6) dtype=int32> print(tf.squeeze(b, 0)) # <tf.Tensor 'Squeeze:0' shape=(6,) dtype=int32>
-
参考:
[1] https://blog.csdn.net/Cyiano/article/details/76087747 (例子简单易懂)
[2] https://www.tensorflow.org/versions/r1.1/api_docs/python/tf/concat (明白了:官网文档的某些api有详细例子,某些没有就很坑)
-
1.2 其他题解答
-
tf.diag
-
示例代码
def test5(): # 这也是1e diag = tf.range(1,7) # [1,7) print("test_5: ", sess.run(tf.diag(diagonal=diag))) test5() def test5_2(): diag = tf.constant([1,2,3,5]) print("test_5_2: ", sess.run(tf.diag(diagonal=diag))) test5_2() # 其实diagonal还能是 2-d/3-d tensor, 详见下方link/官方api
-
参考:
[1] https://blog.csdn.net/baidu_15113429/article/details/78082282
-
-
tf.matrix_determinant
-
示例
def test6(): x = tf.random_normal([10,10], mean=0.5, stddev=1.5) print("test6, x: ", sess.run(x)) print("test6, det(x): ", sess.run(tf.matrix_determinant(x))) test6() def test6_2(): A1 = [[1, 1, 1], [1, -1, -1], [5, -2, 2]] A = tf.constant(A1, tf.float32) print(sess.run(A)) d = tf.matrix_determinant(A) print("det(d): ", sess.run(d)) test6_2()
-
参考
[1] https://segmentfault.com/a/1190000014591047 (包含很多矩阵的基本操作)
-
-
tf.unique
-
代码示例
def test7(): x = tf.constant([5, 2, 3, 5, 10, 6, 2, 3, 4, 2, 1, 1, 0, 9]) y, idx = tf.unique(x) # y是去重之后的1-d tensor, idx是原来x的index,只不过 # x中重复的元素的index被换成该重复元素第一次出现的index print("test7: idx: {}, x: {}".format(sess.run(y), sess.run(idx))) test7()
-
参考
[1] https://www.tensorflow.org/api_docs/python/tf/unique (终于碰到一个api的官方文档有例子了)
-
-
tf.reduce_mean
-
代码:
x = tf.constant([[1., 1.], [2., 2.]]) tf.reduce_mean(x) # 1.5 # 对整个tensor求均值 tf.reduce_mean(x, 0) # [1.5, 1.5] # 第一个维度,即对每列求均值 tf.reduce_mean(x, 1) # [1., 2.] # 第二个维度,即对每行求均值 # 注意:tf.square(x)返回和x一样shape的tensor, 不改变shape仅做elme-wise的op在tf中很常见
-
参考:
[1] https://blog.csdn.net/qq_32166627/article/details/52734387 (同理 tf.reduce_xxx)
-
-
最后一个作业,基于tf.reduce_mean,sum, tf.cond,有点像实现huber_loss
def test8(): x, y = tf.random_normal([300]), tf.random_normal([300]) print("test8, \n x:{} \n y:{}".format(sess.run(x), sess.run(y))) residual = x - y # 返回一个tensor(shape同x,y) mean_residual = tf.reduce_mean(residual) # def f1(): return tf.reduce_mean(tf.square(residual)) # 默认使用上述的全局residual def f2(): return tf.reduce_sum(tf.abs(residual)) return tf.cond(mean_residual < 0, f1, f2) res = test8() print("test8, res is:{}".format(sess.run(res)))
-
上述这些小题目的官方solutions:
https://github.com/chiphuyen/stanford-tensorflow-tutorials/blob/master/assignments/01/q1_sol.py
1.3 LR_MNIST
- 以前写过,有时间再写一次
1.4 LR_notMNIST
- 一个像MNIST的新数据集:notMNIST
- 有时间写一下
1.5 实现97%以上的MNIST的acc
2. CNN部分实践
2.1 一个toy example
-
网络图:
-
一些准备
- ubuntu 16.04 + conda + py3.6 + opencv: https://www.jianshu.com/p/a1db7533b5da
- python/tf中的
-1 or None
大多数情况下指的是:程序自动推断这个值是多少(-1 or None只是一个“自动推断”的标识符) if i % int(data.train.num_examples/batch_size) == 0:
表示刚执行完一个epoch- 数据集问题已解决(通过kaggle),但是代码风格不太喜欢(有点旧了),抽时间再完成这个
2.2 参考
-
[2] https://cv-tricks.com/tensorflow-tutorial/training-convolutional-neural-network-for-image-classification/ (参考的这个)
3. Assignment_2
- 参见7_style_transfer