TensorFlow API
定義張量
import tensorflow as tf import numpy as np if __name__ == '__main__': a = tf.constant([[1, 2], [3, 4]]) print(a) b = a a = tf.ones((2, 2)) print(a) a = tf.zeros((2, 2)) print(a) a = tf.eye(2, 2) print(a) a = tf.zeros_like(b) print(a) a = tf.ones_like(b) print(a) # 生成0到10，中間間隔1的隊列 a = tf.constant(np.arange(0, 11, 1)) print(a) # 生成2到10，中間間隔4的隊列 a = tf.linspace(2, 10, 4) print(a) # 生成0～6的隊列 a = tf.range(7) print(a)
運行結果
tf.Tensor(
[[1 2]
[3 4]], shape=(2, 2), dtype=int32)
tf.Tensor(
[[1. 1.]
[1. 1.]], shape=(2, 2), dtype=float32)
tf.Tensor(
[[0. 0.]
[0. 0.]], shape=(2, 2), dtype=float32)
tf.Tensor(
[[1. 0.]
[0. 1.]], shape=(2, 2), dtype=float32)
tf.Tensor(
[[0 0]
[0 0]], shape=(2, 2), dtype=int32)
tf.Tensor(
[[1 1]
[1 1]], shape=(2, 2), dtype=int32)
tf.Tensor([ 0 1 2 3 4 5 6 7 8 9 10], shape=(11,), dtype=int64)
tf.Tensor([ 2. 4.66666667 7.33333333 10. ], shape=(4,), dtype=float64)
tf.Tensor([0 1 2 3 4 5 6], shape=(7,), dtype=int32)
算術運算
import tensorflow as tf if __name__ == '__main__': a = tf.constant([[1, 2, 3], [4, 5, 6]]) b = tf.constant([1, 4, 7]) print(a + b) print(tf.add(a, b)) print(a - b) print(tf.subtract(a, b)) print(a * b) print(tf.multiply(a, b)) print(a / b) print(tf.divide(a, b))
運行結果
tf.Tensor(
[[ 2 6 10]
[ 5 9 13]], shape=(2, 3), dtype=int32)
tf.Tensor(
[[ 2 6 10]
[ 5 9 13]], shape=(2, 3), dtype=int32)
tf.Tensor(
[[ 0 -2 -4]
[ 3 1 -1]], shape=(2, 3), dtype=int32)
tf.Tensor(
[[ 0 -2 -4]
[ 3 1 -1]], shape=(2, 3), dtype=int32)
tf.Tensor(
[[ 1 8 21]
[ 4 20 42]], shape=(2, 3), dtype=int32)
tf.Tensor(
[[ 1 8 21]
[ 4 20 42]], shape=(2, 3), dtype=int32)
tf.Tensor(
[[1. 0.5 0.42857143]
[4. 1.25 0.85714286]], shape=(2, 3), dtype=float64)
tf.Tensor(
[[1. 0.5 0.42857143]
[4. 1.25 0.85714286]], shape=(2, 3), dtype=float64)
矩陣乘法
import tensorflow as tf if __name__ == '__main__': a = tf.constant([[1, 2, 3], [4, 5, 6]]) b = tf.constant([[2, 4], [11, 13], [7, 9]]) print(tf.matmul(a, b))
運行結果
tf.Tensor(
[[ 45 57]
[105 135]], shape=(2, 2), dtype=int32)
冪運算
import tensorflow as tf if __name__ == '__main__': a = tf.constant([[1, 2, 3], [4, 5, 6]]) print(tf.pow(a, 2)) print(a**2) a = tf.constant([2], dtype=tf.float32) print(tf.exp(a))
運行結果
tf.Tensor(
[[ 1 4 9]
[16 25 36]], shape=(2, 3), dtype=int32)
tf.Tensor(
[[ 1 4 9]
[16 25 36]], shape=(2, 3), dtype=int32)
tf.Tensor([7.389056], shape=(1,), dtype=float32)
稀疏張量
import tensorflow as tf if __name__ == '__main__': a = tf.SparseTensor(indices=[[0, 2], [1, 0], [1, 2]], values=[3., 4., 5.], dense_shape=[2, 4]) print(a) print(tf.sparse.to_dense(a))
運行結果
SparseTensor(indices=tf.Tensor(
[[0 2]
[1 0]
[1 2]], shape=(3, 2), dtype=int64), values=tf.Tensor([3. 4. 5.], shape=(3,), dtype=float32), dense_shape=tf.Tensor([2 4], shape=(2,), dtype=int64))
tf.Tensor(
[[0. 0. 3. 0.]
[4. 0. 5. 0.]], shape=(2, 4), dtype=float32)
這裏跟PyTorch不同的是序號定義的不同，PyTorch是上下定義位置，而Tensorflow是左右定義位置。
開方運算
import tensorflow as tf if __name__ == '__main__': a = tf.constant([[1, 2, 3], [4, 5, 6]], dtype=tf.float32) print(tf.sqrt(a))
運行結果
tf.Tensor(
[[1. 1.4142135 1.7320508]
[2. 2.236068 2.4494898]], shape=(2, 3), dtype=float32)
對數運算
import tensorflow as tf if __name__ == '__main__': a = tf.constant([1, 2, 3], dtype=tf.float32) print(tf.math.log(a)) # 以2爲底的對數運算 print(tf.math.log(a) / tf.math.log(tf.constant([2], dtype=tf.float32)))
運行結果
tf.Tensor([0. 0.6931472 1.0986123], shape=(3,), dtype=float32)
tf.Tensor([0. 1. 1.5849625], shape=(3,), dtype=float32)
廣播機制
import tensorflow as tf import numpy as np if __name__ == '__main__': a = tf.constant(np.random.rand(2, 1, 1)) print(a) b = tf.constant(np.random.rand(3)) print(b) print(a + b)
運行結果
tf.Tensor(
[[[0.01537639]]
[[0.82803137]]], shape=(2, 1, 1), dtype=float64)
tf.Tensor([0.85110735 0.80508366 0.94771198], shape=(3,), dtype=float64)
tf.Tensor(
[[[0.86648374 0.82046006 0.96308837]]
[[1.67913872 1.63311503 1.77574335]]], shape=(2, 1, 3), dtype=float64)
取整/取餘
import tensorflow as tf import numpy as np if __name__ == '__main__': a = tf.constant(np.random.rand(2, 2)) a *= 10 print(a) print(tf.floor(a)) print(tf.math.ceil(a)) print(tf.round(a)) print(a % 2) print(tf.math.mod(a, 2))
運行結果
tf.Tensor(
[[7.44209334 6.03067307]
[0.33836814 2.66141984]], shape=(2, 2), dtype=float64)
tf.Tensor(
[[7. 6.]
[0. 2.]], shape=(2, 2), dtype=float64)
tf.Tensor(
[[8. 7.]
[1. 3.]], shape=(2, 2), dtype=float64)
tf.Tensor(
[[7. 6.]
[0. 3.]], shape=(2, 2), dtype=float64)
tf.Tensor(
[[1.44209334 0.03067307]
[0.33836814 0.66141984]], shape=(2, 2), dtype=float64)
tf.Tensor(
[[1.44209334 0.03067307]
[0.33836814 0.66141984]], shape=(2, 2), dtype=float64)
比較運算
import tensorflow as tf if __name__ == '__main__': a = tf.constant([[1, 2, 3], [4, 5, 6]]) b = tf.constant([[1, 4, 7], [6, 5, 7]]) c = tf.random.uniform((2, 4)) d = a print(tf.equal(a, b)) # >= print(tf.greater_equal(a, b)) # > print(tf.greater(a, b)) # <= print(tf.less_equal(a, b)) # < print(tf.less(a, b)) # != print(tf.not_equal(a, b))
運行結果
tf.Tensor(
[[ True False False]
[False True False]], shape=(2, 3), dtype=bool)
tf.Tensor(
[[ True False False]
[False True False]], shape=(2, 3), dtype=bool)
tf.Tensor(
[[False False False]
[False False False]], shape=(2, 3), dtype=bool)
tf.Tensor(
[[ True True True]
[ True True True]], shape=(2, 3), dtype=bool)
tf.Tensor(
[[False True True]
[ True False True]], shape=(2, 3), dtype=bool)
tf.Tensor(
[[False True True]
[ True False True]], shape=(2, 3), dtype=bool)
排序
import tensorflow as tf if __name__ == '__main__': a = tf.constant([1, 4, 4, 3, 5]) print(tf.sort(a)) # 升序的序號 print(tf.argsort(a)) print(tf.sort(a, direction='DESCENDING')) # 降序的序號 print(tf.argsort(a, direction='DESCENDING')) b = tf.constant([[1, 4, 4, 3, 5], [2, 3, 1, 3, 5]]) print(tf.sort(b)) # 在第一個維度進行升序排序 print(tf.sort(b, axis=0))
運行結果
tf.Tensor([1 3 4 4 5], shape=(5,), dtype=int32)
tf.Tensor([0 3 1 2 4], shape=(5,), dtype=int32)
tf.Tensor([5 4 4 3 1], shape=(5,), dtype=int32)
tf.Tensor([4 1 2 3 0], shape=(5,), dtype=int32)
tf.Tensor(
[[1 3 4 4 5]
[1 2 3 3 5]], shape=(2, 5), dtype=int32)
tf.Tensor(
[[1 3 1 3 5]
[2 4 4 3 5]], shape=(2, 5), dtype=int32)
Top-K
import tensorflow as tf if __name__ == '__main__': a = tf.constant([[1, 4, 6, 3, 5], [2, 3, 4, 4, 5]], dtype=tf.float32) print(tf.math.top_k(a, k=2)) # 預測出來的值是否在4和1的位置上 print(tf.math.in_top_k([4, 1], a, k=2))
運行結果
TopKV2(values=<tf.Tensor: shape=(2, 2), dtype=float32, numpy=
array([[6., 5.],
[5., 4.]], dtype=float32)>, indices=<tf.Tensor: shape=(2, 2), dtype=int32, numpy=
array([[2, 4],
[4, 2]], dtype=int32)>)
tf.Tensor([ True False], shape=(2,), dtype=bool)
這裏跟PyTorch不同的是，它不支持不同維度上的Top-K.
三角函數
import tensorflow as tf if __name__ == '__main__': a = tf.constant([0, 0, 0], dtype=tf.float32) print(tf.cos(a))
運行結果
tf.Tensor([1. 1. 1.], shape=(3,), dtype=float32)
統計學函數
import tensorflow as tf import numpy as np from scipy import stats if __name__ == '__main__': a = tf.constant(np.random.rand(2, 2)) print(a) # 求均值 print(tf.reduce_mean(a)) # 獲取每一行的最大值 print(tf.reduce_max(a, [1])) # 獲取每一列的最大值 print(tf.reduce_max(a, [0])) # 對第一個維度求均值 print(tf.reduce_mean(a, axis=0)) # 求和 print(tf.math.reduce_sum(a)) # 對第一個維度求和 print(tf.math.reduce_sum(a, axis=0)) # 累積 print(tf.math.reduce_prod(a)) # 對第一個維度求累積 print(tf.math.reduce_prod(a, axis=0)) # 求第一個維度的最大值索引 print(tf.argmax(a, axis=0)) # 求第一個維度的最小值索引 print(tf.argmin(a, axis=0)) # 計算標準差 print(tf.math.reduce_std(a)) # 計算方差 print(tf.math.reduce_variance(a)) # 獲取中數(中間的數) print(tf.constant(np.median(a.numpy()))) # 獲取衆數(出現次數最多的數) print(tf.constant(stats.mode(a.numpy())[0][0])) a *= 10 # 打印直方圖 print(tf.histogram_fixed_width(a, [0, tf.math.reduce_max(a)], nbins=6)) a = tf.constant(np.random.randint(0, 10, [10]), dtype=tf.int32) print(a) print(tf.math.bincount(a))
運行結果
tf.Tensor(
[[0.18079731 0.83275504]
[0.93849049 0.69468691]], shape=(2, 2), dtype=float64)
tf.Tensor(0.6616824381588154, shape=(), dtype=float64)
tf.Tensor([0.83275504 0.93849049], shape=(2,), dtype=float64)
tf.Tensor([0.93849049 0.83275504], shape=(2,), dtype=float64)
tf.Tensor([0.5596439 0.76372097], shape=(2,), dtype=float64)
tf.Tensor(2.646729752635262, shape=(), dtype=float64)
tf.Tensor([1.1192878 1.52744195], shape=(2,), dtype=float64)
tf.Tensor(0.09815857168640853, shape=(), dtype=float64)
tf.Tensor([0.16967656 0.57850403], shape=(2,), dtype=float64)
tf.Tensor([1 0], shape=(2,), dtype=int64)
tf.Tensor([0 1], shape=(2,), dtype=int64)
tf.Tensor(0.29078700667359864, shape=(), dtype=float64)
tf.Tensor(0.08455708325019148, shape=(), dtype=float64)
tf.Tensor(0.7637209747951317, shape=(), dtype=float64)
tf.Tensor([0.18079731 0.69468691], shape=(2,), dtype=float64)
tf.Tensor([0 1 0 0 1 2], shape=(6,), dtype=int32)
tf.Tensor([9 8 5 5 1 9 7 3 9 9], shape=(10,), dtype=int32)
tf.Tensor([0 1 0 1 0 2 0 1 1 4], shape=(10,), dtype=int32)
隨機抽樣
import tensorflow as tf import numpy as np if __name__ == '__main__': tf.random.set_seed(1) np.random.seed(1) mean = tf.constant(np.random.rand(1, 2), dtype=tf.float32) std = tf.constant(np.random.rand(1, 2), dtype=tf.float32) # 正態分佈 print(tf.random.normal((1, 2), mean=mean, stddev=std))
運行結果
tf.Tensor([[0.41689605 1.1876557 ]], shape=(1, 2), dtype=float32)
範數運算
import tensorflow as tf import numpy as np if __name__ == '__main__': a = tf.constant(np.random.rand(2, 1)) b = tf.constant(np.random.rand(2, 1)) print(a) print(b) # L1範數 print(tf.reduce_sum(tf.abs(a - b))) # L2範數 print(tf.sqrt(tf.reduce_sum(tf.square(a - b), axis=0))) print(tf.norm(a))
運行結果
tf.Tensor(
[[0.90844106]
[0.57310612]], shape=(2, 1), dtype=float64)
tf.Tensor(
[[0.49385858]
[0.8327303 ]], shape=(2, 1), dtype=float64)
tf.Tensor(0.6742066607716513, shape=(), dtype=float64)
tf.Tensor([0.48916597], shape=(1,), dtype=float64)
tf.Tensor(1.0741116263657322, shape=(), dtype=float64)
張量裁剪
import tensorflow as tf import numpy as np if __name__ == '__main__': a = tf.constant(np.random.rand(2, 2)) * 10 print(a) # 將小於2的變成2，大於5的變成5，2～5之間的不變 a = tf.clip_by_value(a, 2, 5) print(a)
運行結果
tf.Tensor(
[[0.70826017 7.9352412 ]
[2.81880047 0.15188473]], shape=(2, 2), dtype=float64)
tf.Tensor(
[[2. 5. ]
[2.81880047 2. ]], shape=(2, 2), dtype=float64)
張量的索引與數據篩選
import tensorflow as tf import numpy as np if __name__ == '__main__': a = tf.constant(np.random.rand(4, 4)) b = tf.constant(np.random.rand(4, 4)) print(a) print(b) out = tf.where(a > 0.5, a, b) print(out) # 此處返回的是a大於b的所有座標，但是跟PyTorch表現形式不同 out = tf.where(a > b) print(out) # 挑選a的第一個維度的第0行、第3行、第2行構建出一個新的tensor out = tf.gather(a, indices=tf.constant([0, 3, 2]), axis=0) print(out) # 挑選a的第二個維度的第0列、第3列、第2列構建出一個新的tensor out = tf.gather(a, indices=tf.constant([0, 3, 2]), axis=1) print(out) a = tf.linspace(1, 16, 16) a = tf.reshape(a, (4, 4)) print(a) # 此處跟PyTorch的gather不一樣 out = tf.gather(a, indices=tf.constant([[0, 1, 1, 1], [0, 1, 2, 2], [0, 1, 3, 3]]), axis=0) print(out) mask = tf.greater(a, 8) print(mask) out = tf.boolean_mask(a, mask) print(out) a = tf.reshape(a, (-1,)) out = tf.gather(a, indices=tf.constant([0, 15, 13, 10])) print(out) a = tf.constant([[0, 1, 2, 0], [2, 3, 0, 1]]) out = tf.where(a > 0) print(out)
運行結果
tf.Tensor(
[[0.36259518 0.19380163 0.44899191 0.28385357]
[0.515987 0.80924924 0.2813977 0.61434081]
[0.77888583 0.11589285 0.47789409 0.02359696]
[0.16616797 0.19040272 0.80915301 0.49108375]], shape=(4, 4), dtype=float64)
tf.Tensor(
[[0.38603178 0.6567274 0.93351692 0.97103008]
[0.11047731 0.44910036 0.66476197 0.223172 ]
[0.01461463 0.76318377 0.44559227 0.93314565]
[0.92961906 0.58895644 0.0390715 0.78634479]], shape=(4, 4), dtype=float64)
tf.Tensor(
[[0.38603178 0.6567274 0.93351692 0.97103008]
[0.515987 0.80924924 0.66476197 0.61434081]
[0.77888583 0.76318377 0.44559227 0.93314565]
[0.92961906 0.58895644 0.80915301 0.78634479]], shape=(4, 4), dtype=float64)
tf.Tensor(
[[1 0]
[1 1]
[1 3]
[2 0]
[2 2]
[3 2]], shape=(6, 2), dtype=int64)
tf.Tensor(
[[0.36259518 0.19380163 0.44899191 0.28385357]
[0.16616797 0.19040272 0.80915301 0.49108375]
[0.77888583 0.11589285 0.47789409 0.02359696]], shape=(3, 4), dtype=float64)
tf.Tensor(
[[0.36259518 0.28385357 0.44899191]
[0.515987 0.61434081 0.2813977 ]
[0.77888583 0.02359696 0.47789409]
[0.16616797 0.49108375 0.80915301]], shape=(4, 3), dtype=float64)
tf.Tensor(
[[ 1. 2. 3. 4.]
[ 5. 6. 7. 8.]
[ 9. 10. 11. 12.]
[13. 14. 15. 16.]], shape=(4, 4), dtype=float64)
tf.Tensor(
[[[ 1. 2. 3. 4.]
[ 5. 6. 7. 8.]
[ 5. 6. 7. 8.]
[ 5. 6. 7. 8.]]
[[ 1. 2. 3. 4.]
[ 5. 6. 7. 8.]
[ 9. 10. 11. 12.]
[ 9. 10. 11. 12.]]
[[ 1. 2. 3. 4.]
[ 5. 6. 7. 8.]
[13. 14. 15. 16.]
[13. 14. 15. 16.]]], shape=(3, 4, 4), dtype=float64)
tf.Tensor(
[[False False False False]
[False False False False]
[ True True True True]
[ True True True True]], shape=(4, 4), dtype=bool)
tf.Tensor([ 9. 10. 11. 12. 13. 14. 15. 16.], shape=(8,), dtype=float64)
tf.Tensor([ 1. 16. 14. 11.], shape=(4,), dtype=float64)
tf.Tensor(
[[0 1]
[0 2]
[1 0]
[1 1]
[1 3]], shape=(5, 2), dtype=int64)
張量的組合與拼接
import tensorflow as tf if __name__ == '__main__': a = tf.zeros((2, 4)) b = tf.ones((2, 4)) # 在第一個維度上進行拼接，此時不會增加維度，只會增加該維度的條數 out = tf.concat((a, b), axis=0) print(out) a = tf.linspace(1, 6, 6) a = tf.reshape(a, (2, 3)) b = tf.linspace(7, 12, 6) b = tf.reshape(b, (2, 3)) print(a) print(b) # 將a和b看成獨立元素進行拼接，此時會增加一個維度，該維度表示a和b的2個元素 out = tf.stack((a, b), axis=0) print(out) # 將a的每一行跟b的相同序號的一行拼接，拼接後形成一個維度 out = tf.stack((a, b), axis=1) print(out) print(out[:, 0, :]) print(out[:, 1, :]) # 將a的每一行每一列跟b的相同序號的行相同序號的列拼接，拼接後形成一個維度 out = tf.stack((a, b), axis=2) print(out) print(out[:, :, 0]) print(out[:, :, 1])
運行結果
tf.Tensor(
[[0. 0. 0. 0.]
[0. 0. 0. 0.]
[1. 1. 1. 1.]
[1. 1. 1. 1.]], shape=(4, 4), dtype=float32)
tf.Tensor(
[[1. 2. 3.]
[4. 5. 6.]], shape=(2, 3), dtype=float64)
tf.Tensor(
[[ 7. 8. 9.]
[10. 11. 12.]], shape=(2, 3), dtype=float64)
tf.Tensor(
[[[ 1. 2. 3.]
[ 4. 5. 6.]]
[[ 7. 8. 9.]
[10. 11. 12.]]], shape=(2, 2, 3), dtype=float64)
tf.Tensor(
[[[ 1. 2. 3.]
[ 7. 8. 9.]]
[[ 4. 5. 6.]
[10. 11. 12.]]], shape=(2, 2, 3), dtype=float64)
tf.Tensor(
[[1. 2. 3.]
[4. 5. 6.]], shape=(2, 3), dtype=float64)
tf.Tensor(
[[ 7. 8. 9.]
[10. 11. 12.]], shape=(2, 3), dtype=float64)
tf.Tensor(
[[[ 1. 7.]
[ 2. 8.]
[ 3. 9.]]
[[ 4. 10.]
[ 5. 11.]
[ 6. 12.]]], shape=(2, 3, 2), dtype=float64)
tf.Tensor(
[[1. 2. 3.]
[4. 5. 6.]], shape=(2, 3), dtype=float64)
tf.Tensor(
[[ 7. 8. 9.]
[10. 11. 12.]], shape=(2, 3), dtype=float64)
張量切片
import tensorflow as tf import numpy as np if __name__ == '__main__': a = tf.constant(np.random.rand(4, 4)) print(a) # 不支持不等分切片 out = tf.split(a, 2, axis=0) print(out) out = tf.split(a, 2, axis=1) print(out) out = tf.split(a, [2, 1, 1], axis=0) print(out)
運行結果
tf.Tensor(
[[0.76125837 0.55359673 0.5226298 0.162057 ]
[0.70181616 0.21919613 0.29682429 0.87271386]
[0.65784256 0.11186252 0.46019658 0.68551491]
[0.93408276 0.7907662 0.62974255 0.16056717]], shape=(4, 4), dtype=float64)
[<tf.Tensor: shape=(2, 4), dtype=float64, numpy=
array([[0.76125837, 0.55359673, 0.5226298 , 0.162057 ],
[0.70181616, 0.21919613, 0.29682429, 0.87271386]])>, <tf.Tensor: shape=(2, 4), dtype=float64, numpy=
array([[0.65784256, 0.11186252, 0.46019658, 0.68551491],
[0.93408276, 0.7907662 , 0.62974255, 0.16056717]])>]
[<tf.Tensor: shape=(4, 2), dtype=float64, numpy=
array([[0.76125837, 0.55359673],
[0.70181616, 0.21919613],
[0.65784256, 0.11186252],
[0.93408276, 0.7907662 ]])>, <tf.Tensor: shape=(4, 2), dtype=float64, numpy=
array([[0.5226298 , 0.162057 ],
[0.29682429, 0.87271386],
[0.46019658, 0.68551491],
[0.62974255, 0.16056717]])>]
[<tf.Tensor: shape=(2, 4), dtype=float64, numpy=
array([[0.76125837, 0.55359673, 0.5226298 , 0.162057 ],
[0.70181616, 0.21919613, 0.29682429, 0.87271386]])>, <tf.Tensor: shape=(1, 4), dtype=float64, numpy=array([[0.65784256, 0.11186252, 0.46019658, 0.68551491]])>, <tf.Tensor: shape=(1, 4), dtype=float64, numpy=array([[0.93408276, 0.7907662 , 0.62974255, 0.16056717]])>]
張量變形
import tensorflow as tf import numpy as np if __name__ == '__main__': a = tf.constant(np.random.rand(2, 3)) print(a) out = tf.reshape(a, (3, 2)) print(out) print(tf.transpose(a)) a = tf.constant(np.random.rand(1, 2, 3)) print(a) # 這裏跟PyTorch不同，需要列出所有的維度，才能交換 out = tf.transpose(a, (1, 0, 2)) print(out) out = tf.squeeze(a) print(out) b = tf.constant(np.random.rand(1, 2, 1)) print(b) # 只去掉第一個維度的1，保留第三個維度的1 out = tf.squeeze(b, [0]) print(out) out = tf.expand_dims(a, -1) print(out) out = tf.split(a, 2, axis=1) print(out) out = tf.reverse(a, axis=[1]) print(out) out = tf.reverse(a, axis=[2]) print(out) out = tf.reverse(a, axis=[1, 2]) print(out)
運行結果
tf.Tensor(
[[0.58693252 0.40994406 0.4536791 ]
[0.98581972 0.02112171 0.13718469]], shape=(2, 3), dtype=float64)
tf.Tensor(
[[0.58693252 0.40994406]
[0.4536791 0.98581972]
[0.02112171 0.13718469]], shape=(3, 2), dtype=float64)
tf.Tensor(
[[0.58693252 0.98581972]
[0.40994406 0.02112171]
[0.4536791 0.13718469]], shape=(3, 2), dtype=float64)
tf.Tensor(
[[[0.0294984 0.39577176 0.41719977]
[0.16434196 0.06959612 0.89783393]]], shape=(1, 2, 3), dtype=float64)
tf.Tensor(
[[[0.0294984 0.39577176 0.41719977]]
[[0.16434196 0.06959612 0.89783393]]], shape=(2, 1, 3), dtype=float64)
tf.Tensor(
[[0.0294984 0.39577176 0.41719977]
[0.16434196 0.06959612 0.89783393]], shape=(2, 3), dtype=float64)
tf.Tensor(
[[[0.48809593]
[0.65678944]]], shape=(1, 2, 1), dtype=float64)
tf.Tensor(
[[0.48809593]
[0.65678944]], shape=(2, 1), dtype=float64)
tf.Tensor(
[[[[0.0294984 ]
[0.39577176]
[0.41719977]]
[[0.16434196]
[0.06959612]
[0.89783393]]]], shape=(1, 2, 3, 1), dtype=float64)
[<tf.Tensor: shape=(1, 1, 3), dtype=float64, numpy=array([[[0.0294984 , 0.39577176, 0.41719977]]])>, <tf.Tensor: shape=(1, 1, 3), dtype=float64, numpy=array([[[0.16434196, 0.06959612, 0.89783393]]])>]
tf.Tensor(
[[[0.16434196 0.06959612 0.89783393]
[0.0294984 0.39577176 0.41719977]]], shape=(1, 2, 3), dtype=float64)
tf.Tensor(
[[[0.41719977 0.39577176 0.0294984 ]
[0.89783393 0.06959612 0.16434196]]], shape=(1, 2, 3), dtype=float64)
tf.Tensor(
[[[0.89783393 0.06959612 0.16434196]
[0.41719977 0.39577176 0.0294984 ]]], shape=(1, 2, 3), dtype=float64)
tensorflow沒有支持旋轉的API。
張量填充
import tensorflow as tf import numpy as np if __name__ == '__main__': a = tf.fill((2, 3), 10) print(a) a = tf.constant(np.random.rand(2, 2)) print(a) # 在tensor的上下左各填充一行0，右邊不填充 out = tf.pad(a, [[1, 1], [1, 0]]) print(out)
運行結果
tf.Tensor(
[[10 10 10]
[10 10 10]], shape=(2, 3), dtype=int32)
tf.Tensor(
[[0.65031861 0.45838591]
[0.54348782 0.06645889]], shape=(2, 2), dtype=float64)
tf.Tensor(
[[0. 0. 0. ]
[0. 0.65031861 0.45838591]
[0. 0.54348782 0.06645889]
[0. 0. 0. ]], shape=(4, 3), dtype=float64)
網格
import tensorflow as tf import numpy as np if __name__ == '__main__': a = tf.constant(np.random.rand(2, 2)) b = tf.constant(np.random.rand(3, 2)) print(a) print(b) a_count = a.shape[0] * a.shape[1] b_count = b.shape[0] * b.shape[1] print(a_count, b_count) # 生成兩個b_count*a_count的張量(這裏是6*4)，第一個張量 # 包含6行a的扁平向量，第二個張量包含6行4列 # b的每一個值，其中每行的值相同，不同行的值不同 out = tf.meshgrid(a, b) print(out)
運行結果
tf.Tensor(
[[0.71037233 0.70864467]
[0.72272255 0.97050206]], shape=(2, 2), dtype=float64)
tf.Tensor(
[[0.9679168 0.60009048]
[0.43580961 0.31357009]
[0.23997045 0.63421268]], shape=(3, 2), dtype=float64)
4 6
[<tf.Tensor: shape=(6, 4), dtype=float64, numpy=
array([[0.71037233, 0.70864467, 0.72272255, 0.97050206],
[0.71037233, 0.70864467, 0.72272255, 0.97050206],
[0.71037233, 0.70864467, 0.72272255, 0.97050206],
[0.71037233, 0.70864467, 0.72272255, 0.97050206],
[0.71037233, 0.70864467, 0.72272255, 0.97050206],
[0.71037233, 0.70864467, 0.72272255, 0.97050206]])>, <tf.Tensor: shape=(6, 4), dtype=float64, numpy=
array([[0.9679168 , 0.9679168 , 0.9679168 , 0.9679168 ],
[0.60009048, 0.60009048, 0.60009048, 0.60009048],
[0.43580961, 0.43580961, 0.43580961, 0.43580961],
[0.31357009, 0.31357009, 0.31357009, 0.31357009],
[0.23997045, 0.23997045, 0.23997045, 0.23997045],
[0.63421268, 0.63421268, 0.63421268, 0.63421268]])>]
保留最大最小值
import tensorflow as tf import numpy as np if __name__ == '__main__': a = tf.constant(np.random.rand(2, 3)) b = tf.constant(np.random.rand(2, 3)) print(a) print(b) # 保留a、b每一個位置上的最小值，形成一個新的張量 out = tf.minimum(a, b) print(out) out = tf.where(a < b, a, b) print(out) # 保留a、b每一個位置上的最大值，形成一個新的張量 out = tf.maximum(a, b) print(out) out = tf.where(a > b, a, b) print(out)
運行結果
tf.Tensor(
[[0.21136504 0.79508077 0.08270648]
[0.63864327 0.45865036 0.92851595]], shape=(2, 3), dtype=float64)
tf.Tensor(
[[0.60009355 0.06415244 0.79583999]
[0.57438331 0.0202818 0.87532482]], shape=(2, 3), dtype=float64)
tf.Tensor(
[[0.21136504 0.06415244 0.08270648]
[0.57438331 0.0202818 0.87532482]], shape=(2, 3), dtype=float64)
tf.Tensor(
[[0.21136504 0.06415244 0.08270648]
[0.57438331 0.0202818 0.87532482]], shape=(2, 3), dtype=float64)
tf.Tensor(
[[0.60009355 0.79508077 0.79583999]
[0.63864327 0.45865036 0.92851595]], shape=(2, 3), dtype=float64)
tf.Tensor(
[[0.60009355 0.79508077 0.79583999]
[0.63864327 0.45865036 0.92851595]], shape=(2, 3), dtype=float64)
亂序
import tensorflow as tf import numpy as np if __name__ == '__main__': a = tf.constant(np.random.rand(2, 3)) print(a) a = tf.reshape(a, (-1,)) # 亂序 out = tf.random.shuffle(a) out = tf.reshape(out, (2, 3)) print(out)
運行結果
tf.Tensor(
[[0.20761759 0.51233245 0.67144284]
[0.91502356 0.47983641 0.00408644]], shape=(2, 3), dtype=float64)
tf.Tensor(
[[0.91502356 0.67144284 0.00408644]
[0.51233245 0.47983641 0.20761759]], shape=(2, 3), dtype=float64)
保留唯一值
import tensorflow as tf if __name__ == '__main__': a = tf.constant([1, 2, 3, 1, 2, 3]) # 保留唯一值 out = tf.unique(a)[0] print(out)
運行結果
tf.Tensor([1 2 3], shape=(3,), dtype=int32)
根據位置座標填充
import tensorflow as tf import numpy as np if __name__ == '__main__': a = tf.constant(np.random.rand(5, 4), dtype=tf.float32) print(a) dim = tf.where(a > 0.5) print(dim) value = tf.ones((dim.shape[0],)) print(value) # 根據dim的位置座標，將value列表填充回a out = tf.tensor_scatter_nd_update(a, dim, value) print(out)
運行結果
tf.Tensor(
[[0.20862171 0.26360217 0.96401286 0.70376277]
[0.07841657 0.12179881 0.47482252 0.7394226 ]
[0.85304624 0.54386026 0.7446324 0.8679093 ]
[0.59722424 0.25879326 0.07276706 0.44899505]
[0.2745695 0.9253298 0.908037 0.04632197]], shape=(5, 4), dtype=float32)
tf.Tensor(
[[0 2]
[0 3]
[1 3]
[2 0]
[2 1]
[2 2]
[2 3]
[3 0]
[4 1]
[4 2]], shape=(10, 2), dtype=int64)
tf.Tensor([1. 1. 1. 1. 1. 1. 1. 1. 1. 1.], shape=(10,), dtype=float32)
tf.Tensor(
[[0.20862171 0.26360217 1. 1. ]
[0.07841657 0.12179881 0.47482252 1. ]
[1. 1. 1. 1. ]
[1. 0.25879326 0.07276706 0.44899505]
[0.2745695 1. 1. 0.04632197]], shape=(5, 4), dtype=float32)
一維向量也是可以的
import tensorflow as tf import numpy as np if __name__ == '__main__': a = tf.constant(np.random.rand(9,), dtype=tf.float32) print(a) dim = tf.where(a > 0.5) print(dim) value = tf.ones((dim.shape[0],)) print(value) # 根據dim的位置座標，將value列表填充回a out = tf.tensor_scatter_nd_update(a, dim, value) print(out)
運行結果
tf.Tensor(
[0.9599079 0.9158476 0.59267145 0.69443715 0.4816758 0.71821004
0.10235249 0.20957421 0.9197179 ], shape=(9,), dtype=float32)
tf.Tensor(
[[0]
[1]
[2]
[3]
[5]
[8]], shape=(6, 1), dtype=int64)
tf.Tensor([1. 1. 1. 1. 1. 1.], shape=(6,), dtype=float32)
tf.Tensor(
[1. 1. 1. 1. 0.4816758 1.
0.10235249 0.20957421 1. ], shape=(9,), dtype=float32)
擴展維度，複製數值
import tensorflow as tf import numpy as np if __name__ == '__main__': a = tf.constant(np.random.rand(3, 2), dtype=tf.float32) print(a) # 對第一個維度擴展2倍，這裏從3行變成6行；對第二個維度擴展3倍 # 這裏從2列變成6列，值都是複製 out = tf.tile(a, (2, 3)) print(out)
運行結果
tf.Tensor(
[[0.32347408 0.2620785 ]
[0.27956298 0.67863226]
[0.4664628 0.3838015 ]], shape=(3, 2), dtype=float32)
tf.Tensor(
[[0.32347408 0.2620785 0.32347408 0.2620785 0.32347408 0.2620785 ]
[0.27956298 0.67863226 0.27956298 0.67863226 0.27956298 0.67863226]
[0.4664628 0.3838015 0.4664628 0.3838015 0.4664628 0.3838015 ]
[0.32347408 0.2620785 0.32347408 0.2620785 0.32347408 0.2620785 ]
[0.27956298 0.67863226 0.27956298 0.67863226 0.27956298 0.67863226]
[0.4664628 0.3838015 0.4664628 0.3838015 0.4664628 0.3838015 ]], shape=(6, 6), dtype=float32)
根據概率挑選樣本索引
import tensorflow as tf if __name__ == '__main__': a = tf.constant([[3, 2, 4, 5, 3], [2, 3, 4, 2, 1]], dtype=tf.float32) for i in range(5): # 根據概率來取索引數，其中第一行[3, 2, 4, 5, 3]中拿到第一個3的索引的概率爲3/17 # 拿到2的索引的概率爲2/17，4的索引的概率爲4/17 # 5表示在這一行中根據概率取5次 out = tf.random.categorical(a, 5) print(out)
運行結果
tf.Tensor(
[[3 3 3 4 3]
[2 1 2 0 2]], shape=(2, 5), dtype=int64)
tf.Tensor(
[[3 2 3 3 3]
[3 2 1 3 1]], shape=(2, 5), dtype=int64)
tf.Tensor(
[[2 2 0 3 3]
[2 2 2 2 1]], shape=(2, 5), dtype=int64)
tf.Tensor(
[[3 4 3 2 4]
[0 1 2 0 2]], shape=(2, 5), dtype=int64)
tf.Tensor(
[[4 3 3 1 2]
[2 2 1 1 2]], shape=(2, 5), dtype=int64)
數據抽取
import tensorflow as tf if __name__ == '__main__': a = tf.constant([[3, 2, 4, 5, 3], [2, 3, 4, 2, 1]], dtype=tf.float32) # 從第0行第1列開始抽取，抽取的範圍爲2行，3列 out = tf.slice(a, [0, 1], [2, 3]) print(out)
運行結果
tf.Tensor(
[[2. 4. 5.]
[3. 4. 2.]], shape=(2, 3), dtype=float32)
圖像裁剪
import tensorflow as tf from tensorflow.keras import preprocessing import numpy as np import matplotlib.pyplot as plt if __name__ == '__main__': img0 = preprocessing.image.load_img("/Users/admin/Documents/6565.jpeg", color_mode='rgb') img0 = tf.constant(np.asarray(img0)) img1 = preprocessing.image.load_img("/Users/admin/Documents/7788.jpg", color_mode='rgb') img1 = tf.constant(np.asarray(img1)) img1 = tf.image.resize(img1, (500, 600)) img1 = tf.cast(img1, tf.uint8) img = tf.stack([img0, img1], axis=0) # 圖像裁剪，image的格式爲(batch_size,高,寬,通道數)的圖片，boxes表示裁剪區域，格式爲[y1, x1, y2, x2] # 這裏有3個裁剪區域，box_indices表示這些裁剪區域是哪張圖片的 # crop_size表示裁剪後的尺寸都resize到該尺寸，原裁剪尺寸不被保留 out = tf.image.crop_and_resize(image=img, boxes=[[0.5, 0.6, 0.9, 0.8], [0.2, 0.6, 1.3, 0.9], [0.3, 0.7, 1.5, 1.]], box_indices=[0, 1, 0], crop_size=(100, 100)) img1 = tf.cast(out[0], tf.uint8) img2 = tf.cast(out[1], tf.uint8) img3 = tf.cast(out[2], tf.uint8) fig, ax = plt.subplots(1, 3, figsize=(20, 10)) ax[0].imshow(img1) ax[1].imshow(img2) ax[2].imshow(img3) plt.show()
運行結果
波士頓房價預測
import tensorflow as tf from sklearn import datasets if __name__ == '__main__': boston = datasets.load_boston() X = tf.constant(boston.data, dtype=tf.float32) y = tf.constant(boston.target, dtype=tf.float32) y = tf.expand_dims(y, -1) data = tf.concat((X, y), axis=-1) print(data)
運行結果
tf.Tensor(
[[6.3200e-03 1.8000e+01 2.3100e+00 ... 3.9690e+02 4.9800e+00 2.4000e+01]
[2.7310e-02 0.0000e+00 7.0700e+00 ... 3.9690e+02 9.1400e+00 2.1600e+01]
[2.7290e-02 0.0000e+00 7.0700e+00 ... 3.9283e+02 4.0300e+00 3.4700e+01]
...
[6.0760e-02 0.0000e+00 1.1930e+01 ... 3.9690e+02 5.6400e+00 2.3900e+01]
[1.0959e-01 0.0000e+00 1.1930e+01 ... 3.9345e+02 6.4800e+00 2.2000e+01]
[4.7410e-02 0.0000e+00 1.1930e+01 ... 3.9690e+02 7.8800e+00 1.1900e+01]], shape=(506, 14), dtype=float32)
import tensorflow as tf from sklearn import datasets from tensorflow.keras import models, layers, losses, optimizers if __name__ == '__main__': boston = datasets.load_boston() X = tf.constant(boston.data, dtype=tf.float32) y = tf.constant(boston.target, dtype=tf.float32) y = tf.expand_dims(y, -1) data = tf.concat((X, y), axis=-1) print(data) y = tf.squeeze(y) X_train = X[:496] y_train = y[:496] X_test = X[496:] y_test = y[496:] class Net(models.Model): def __init__(self, n_output): super(Net, self).__init__() self.hidden = layers.Dense(100, activation='relu') self.pred = layers.Dense(n_output) def call(self, x): out = self.hidden(x) out = self.pred(out) return out net = Net(1) optimizer = optimizers.Adam(learning_rate=0.01) loss_func = losses.MeanSquaredError() for epoch in range(10000): with tf.GradientTape() as tape: pred = net(X_train) pred = tf.squeeze(pred) loss = loss_func(y_train, pred) * 0.001 grads = tape.gradient(loss, net.variables) optimizer.apply_gradients(zip(grads, net.variables)) print("item:{},loss:{}".format(epoch, loss)) print(pred[:10]) print(y_train[:10]) pred_test = net.predict(X_test) pred_test = tf.squeeze(pred_test) print(pred_test[:10]) print(y_test[:10])
運行結果(最終訓練結果)
item:9999,loss:0.004027204588055611
tf.Tensor(
[26.32511 23.44358 31.221886 33.211582 33.850327 28.128717 22.353136
21.29065 16.69853 20.120735], shape=(10,), dtype=float32)
tf.Tensor([24. 21.6 34.7 33.4 36.2 28.7 22.9 27.1 16.5 18.9], shape=(10,), dtype=float32)
tf.Tensor(
[15.420125 19.307917 21.476751 18.167856 20.689434 23.90838 19.904911
22.959293 21.854155 18.169727], shape=(10,), dtype=float32)
tf.Tensor([19.7 18.3 21.2 17.5 16.8 22.4 20.6 23.9 22. 11.9], shape=(10,), dtype=float32)
手寫數字集
import tensorflow as tf from tensorflow.keras import datasets, models, layers, losses, Sequential, optimizers if __name__ == '__main__': (X_train, y_train), (X_test, y_test) = datasets.mnist.load_data() X_train = tf.constant(X_train, dtype=tf.float32) / 255 X_train = tf.reshape(X_train, (X_train.shape[0], 28, 28, 1)) y_train = tf.constant(y_train, dtype=tf.int32) X_test = tf.constant(X_test, dtype=tf.float32) / 255 X_test = tf.reshape(X_test, (X_test.shape[0], 28, 28, 1)) y_test = tf.constant(y_test, dtype=tf.int32) data_train = tf.data.Dataset.from_tensor_slices((X_train, y_train)) data_test = tf.data.Dataset.from_tensor_slices((X_test, y_test)) data_train = data_train.shuffle(10000).batch(64) data_test = data_test.shuffle(10000).batch(64) class CNN(models.Model): def __init__(self): super(CNN, self).__init__() self.conv = Sequential([ layers.Conv2D(32, (5, 5), padding='same'), layers.BatchNormalization(), layers.ReLU(), layers.MaxPool2D((2, 2)) ]) self.fc = layers.Dense(10) def call(self, x): out = self.conv(x) out = tf.reshape(out, (out.shape[0], -1)) out = self.fc(out) return out cnn = CNN() loss_func = losses.SparseCategoricalCrossentropy(from_logits=True) optimizer = optimizers.Adam(learning_rate=0.01) for epoch in range(10): for i, (images, labels) in enumerate(data_train): with tf.GradientTape() as tape: out = cnn(images) loss = loss_func(labels, out) grads = tape.gradient(loss, cnn.trainable_variables) optimizer.apply_gradients(zip(grads, cnn.trainable_variables)) print("epoch is {}, ite is {}/{}, loss is {}".format(epoch + 1, i, len(X_train) // 64, loss.numpy())) loss_test = 0 accuracy = 0 for i, (images, labels) in enumerate(data_test): out = cnn(images) loss_test += loss_func(labels, out) pred = tf.argmax(out, axis=-1) pred = tf.cast(pred, tf.int32) accuracy += tf.where(pred == labels).shape[0] accuracy = accuracy / len(X_test) loss_test = loss_test / (len(X_test) // 64) print("epoch is {}, accuracy is {}, loss test is {}".format(epoch + 1, accuracy, loss_test.numpy()))
運行結果
epoch is 1, ite is 937/937, loss is 0.03964223340153694
epoch is 1, accuracy is 0.9831, loss test is 0.053593263030052185
epoch is 2, ite is 937/937, loss is 0.0005187737406231463
epoch is 2, accuracy is 0.9808, loss test is 0.06631167978048325
epoch is 3, ite is 937/937, loss is 0.25279638171195984
epoch is 3, accuracy is 0.9853, loss test is 0.04976803809404373
epoch is 4, ite is 937/937, loss is 0.005918027367442846
epoch is 4, accuracy is 0.9814, loss test is 0.0674174427986145
epoch is 5, ite is 937/937, loss is 0.001395822037011385
epoch is 5, accuracy is 0.9845, loss test is 0.054165858775377274
epoch is 6, ite is 937/937, loss is 0.009619451127946377
epoch is 6, accuracy is 0.9809, loss test is 0.07322004437446594
epoch is 7, ite is 937/937, loss is 0.0014850541483610868
epoch is 7, accuracy is 0.9829, loss test is 0.07367455959320068
epoch is 8, ite is 937/937, loss is 0.000294497178401798
epoch is 8, accuracy is 0.9829, loss test is 0.07780919224023819
epoch is 9, ite is 937/937, loss is 1.4779652701690793e-05
epoch is 9, accuracy is 0.9843, loss test is 0.06709001213312149
epoch is 10, ite is 937/937, loss is 0.00040841943700797856
epoch is 10, accuracy is 0.9829, loss test is 0.07521752268075943
Cifar10圖像分類
- VggNet網絡結構
import tensorflow as tf from tensorflow.keras import datasets, models, Sequential, layers, losses, optimizers if __name__ == '__main__': (X_train, y_train), (X_test, y_test) = datasets.cifar10.load_data() X_train = tf.constant(X_train, dtype=tf.float32) / 255 y_train = tf.constant(y_train, dtype=tf.int32) X_test = tf.constant(X_test, dtype=tf.float32) / 255 y_test = tf.constant(y_test, dtype=tf.int32) data_train = tf.data.Dataset.from_tensor_slices((X_train, y_train)) data_test = tf.data.Dataset.from_tensor_slices((X_test, y_test)) data_train = data_train.shuffle(10000).batch(64) data_test = data_test.shuffle(10000).batch(64) class Vggbase(models.Model): def __init__(self): super(Vggbase, self).__init__() self.conv1 = Sequential([ layers.Conv2D(64, (3, 3), strides=1, padding='same'), layers.BatchNormalization(), layers.ReLU() ]) self.max_pooling1 = layers.MaxPool2D((2, 2), strides=2) self.conv2_1 = Sequential([ layers.Conv2D(128, (3, 3), strides=1, padding='same'), layers.BatchNormalization(), layers.ReLU() ]) self.conv2_2 = Sequential([ layers.Conv2D(128, (3, 3), strides=1, padding='same'), layers.BatchNormalization(), layers.ReLU() ]) self.max_pooling2 = layers.MaxPool2D((2, 2), strides=2) self.conv3_1 = Sequential([ layers.Conv2D(256, (3, 3), strides=1, padding='same'), layers.BatchNormalization(), layers.ReLU() ]) self.conv3_2 = Sequential([ layers.Conv2D(256, (3, 3), strides=1, padding='same'), layers.BatchNormalization(), layers.ReLU() ]) self.max_pooling3 = layers.MaxPool2D((2, 2), strides=2, padding='same') self.conv4_1 = Sequential([ layers.Conv2D(512, (3, 3), strides=1, padding='same'), layers.BatchNormalization(), layers.ReLU() ]) self.conv4_2 = Sequential([ layers.Conv2D(512, (3, 3), strides=1, padding='same'), layers.BatchNormalization(), layers.ReLU() ]) self.max_pooling4 = layers.MaxPool2D((2, 2), strides=2) self.fc = layers.Dense(10) def call(self, x): batchsize = x.shape[0] out = self.conv1(x) out = self.max_pooling1(out) out = self.conv2_1(out) out = self.conv2_2(out) out = self.max_pooling2(out) out = self.conv3_1(out) out = self.conv3_2(out) out = self.max_pooling3(out) out = self.conv4_1(out) out = self.conv4_2(out) out = self.max_pooling4(out) out = tf.reshape(out, (batchsize, -1)) out = self.fc(out) return out epoch_num = 200 lr = 0.001 net = Vggbase() loss_func = losses.SparseCategoricalCrossentropy(from_logits=True) scheduler = optimizers.schedules.ExponentialDecay(initial_learning_rate=lr, decay_steps=3905, decay_rate=0.9) optimizer = optimizers.Adam(learning_rate=scheduler) for epoch in range(epoch_num): for i, (images, labels) in enumerate(data_train): with tf.GradientTape() as tape: outputs = net(images) labels = tf.squeeze(labels, axis=-1) loss = loss_func(labels, outputs) grads = tape.gradient(loss, net.trainable_variables) optimizer.apply_gradients(zip(grads, net.trainable_variables)) pred = tf.argmax(outputs, axis=-1) pred = tf.cast(pred, tf.int32) correct = tf.where(pred == labels).shape[0] print("epoch is ", epoch, "step ", i, "loss is: ", loss.numpy(), "mini-batch correct is: ", 100.0 * correct / 64)
運行結果(部分)
epoch is 14 step 82 loss is: 0.08098086 mini-batch correct is: 95.3125
epoch is 14 step 83 loss is: 0.031877242 mini-batch correct is: 100.0
epoch is 14 step 84 loss is: 0.028943349 mini-batch correct is: 98.4375
epoch is 14 step 85 loss is: 0.064095914 mini-batch correct is: 96.875
epoch is 14 step 86 loss is: 0.16567983 mini-batch correct is: 96.875
epoch is 14 step 87 loss is: 0.04256191 mini-batch correct is: 98.4375
epoch is 14 step 88 loss is: 0.028884312 mini-batch correct is: 100.0
epoch is 14 step 89 loss is: 0.032220896 mini-batch correct is: 100.0
epoch is 14 step 90 loss is: 0.025648516 mini-batch correct is: 100.0
- ResNet網絡結構
import tensorflow as tf from tensorflow.keras import datasets, models, Sequential, layers, losses, optimizers if __name__ == '__main__': (X_train, y_train), (X_test, y_test) = datasets.cifar10.load_data() X_train = tf.constant(X_train, dtype=tf.float32) / 255 y_train = tf.constant(y_train, dtype=tf.int32) X_test = tf.constant(X_test, dtype=tf.float32) / 255 y_test = tf.constant(y_test, dtype=tf.int32) data_train = tf.data.Dataset.from_tensor_slices((X_train, y_train)) data_test = tf.data.Dataset.from_tensor_slices((X_test, y_test)) data_train = data_train.shuffle(10000).batch(64) data_test = data_test.shuffle(10000).batch(64) class ResBlock(models.Model): def __init__(self, out_channel, stride=1): super(ResBlock, self).__init__() self.layer = Sequential([ layers.Conv2D(out_channel, (3, 3), strides=stride, padding='same'), layers.BatchNormalization(), layers.ReLU(), layers.Conv2D(out_channel, (3, 3), strides=1, padding='same'), layers.BatchNormalization() ]) self.shortcut = Sequential([]) if stride > 1: self.shortcut = Sequential([ layers.Conv2D(out_channel, (3, 3), strides=stride, padding='same'), layers.BatchNormalization() ]) def call(self, x): out1 = self.layer(x) out2 = self.shortcut(x) out = out1 + out2 out = layers.ReLU()(out) return out class ResNet(models.Model): def make_layer(self, block, out_channel, stride, num_block): layers_list = [] for i in range(num_block): if i == 0: in_stride = stride else: in_stride = 1 layers_list.append(block(out_channel, in_stride)) return Sequential(layers_list) def __init__(self): super(ResNet, self).__init__() self.conv1 = Sequential([ layers.Conv2D(32, (3, 3), strides=1, padding='same'), layers.BatchNormalization(), layers.ReLU() ]) self.layer1 = self.make_layer(ResBlock, 64, 2, 2) self.layer2 = self.make_layer(ResBlock, 128, 2, 2) self.layer3 = self.make_layer(ResBlock, 256, 2, 2) self.layer4 = self.make_layer(ResBlock, 512, 2, 2) self.fc = layers.Dense(10) def call(self, x): out = self.conv1(x) out = self.layer1(out) out = self.layer2(out) out = self.layer3(out) out = self.layer4(out) out = layers.AvgPool2D((2, 2))(out) out = tf.reshape(out, (out.shape[0], -1)) out = self.fc(out) return out epoch_num = 200 lr = 0.001 net = ResNet() loss_func = losses.SparseCategoricalCrossentropy(from_logits=True) scheduler = optimizers.schedules.ExponentialDecay(initial_learning_rate=lr, decay_steps=3905, decay_rate=0.9) optimizer = optimizers.Adam(learning_rate=scheduler) for epoch in range(epoch_num): for i, (images, labels) in enumerate(data_train): with tf.GradientTape() as tape: outputs = net(images) labels = tf.squeeze(labels, axis=-1) loss = loss_func(labels, outputs) grads = tape.gradient(loss, net.trainable_variables) optimizer.apply_gradients(zip(grads, net.trainable_variables)) pred = tf.argmax(outputs, axis=-1) pred = tf.cast(pred, tf.int32) correct = tf.where(pred == labels).shape[0] print("epoch is ", epoch, "step ", i, "loss is: ", loss.numpy(), "mini-batch correct is: ", 100.0 * correct / 64)
- MobileNet網絡結構
import tensorflow as tf from tensorflow.keras import datasets, models, Sequential, layers, losses, optimizers if __name__ == '__main__': (X_train, y_train), (X_test, y_test) = datasets.cifar10.load_data() X_train = tf.constant(X_train, dtype=tf.float32) / 255 y_train = tf.constant(y_train, dtype=tf.int32) X_test = tf.constant(X_test, dtype=tf.float32) / 255 y_test = tf.constant(y_test, dtype=tf.int32) data_train = tf.data.Dataset.from_tensor_slices((X_train, y_train)) data_test = tf.data.Dataset.from_tensor_slices((X_test, y_test)) data_train = data_train.shuffle(10000).batch(64) data_test = data_test.shuffle(10000).batch(64) class MobileNet(models.Model): def conv_dw(self, out_channel, stride): return Sequential([ layers.DepthwiseConv2D((3, 3), strides=stride, padding='same', use_bias=False), layers.BatchNormalization(), layers.ReLU(), layers.Conv2D(out_channel, (1, 1), strides=1, padding='same', use_bias=False), layers.BatchNormalization(), layers.ReLU() ]) def __init__(self): super(MobileNet, self).__init__() self.conv1 = Sequential([ layers.Conv2D(32, (3, 3), strides=1, padding='same'), layers.BatchNormalization(), layers.ReLU() ]) self.conv_dw2 = self.conv_dw(32, 1) self.conv_dw3 = self.conv_dw(64, 2) self.conv_dw4 = self.conv_dw(64, 1) self.conv_dw5 = self.conv_dw(128, 2) self.conv_dw6 = self.conv_dw(128, 1) self.conv_dw7 = self.conv_dw(256, 2) self.conv_dw8 = self.conv_dw(256, 1) self.conv_dw9 = self.conv_dw(512, 2) self.fc = layers.Dense(10) def call(self, x): out = self.conv1(x) out = self.conv_dw2(out) out = self.conv_dw3(out) out = self.conv_dw4(out) out = self.conv_dw5(out) out = self.conv_dw6(out) out = self.conv_dw7(out) out = self.conv_dw8(out) out = self.conv_dw9(out) out = layers.GlobalAveragePooling2D()(out) out = self.fc(out) return out epoch_num = 200 lr = 0.001 loss_func = losses.SparseCategoricalCrossentropy(from_logits=True) scheduler = optimizers.schedules.ExponentialDecay(initial_learning_rate=lr, decay_steps=3905, decay_rate=0.9) optimizer = optimizers.Adam(learning_rate=scheduler) net = MobileNet() net.build(input_shape=(None, 32, 32, 3)) print(net.summary()) net.compile(loss=loss_func, optimizer=optimizer, metrics=['accuracy']) net.fit(data_train, batch_size=64, epochs=epoch_num)
運行結果(部分)
Model: "mobile_net"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
sequential (Sequential) (None, 32, 32, 32) 1024
_________________________________________________________________
sequential_1 (Sequential) (None, 32, 32, 32) 1568
_________________________________________________________________
sequential_2 (Sequential) (None, 16, 16, 64) 2720
_________________________________________________________________
sequential_3 (Sequential) (None, 16, 16, 64) 5184
_________________________________________________________________
sequential_4 (Sequential) (None, 8, 8, 128) 9536
_________________________________________________________________
sequential_5 (Sequential) (None, 8, 8, 128) 18560
_________________________________________________________________
sequential_6 (Sequential) (None, 4, 4, 256) 35456
_________________________________________________________________
sequential_7 (Sequential) (None, 4, 4, 256) 69888
_________________________________________________________________
sequential_8 (Sequential) (None, 2, 2, 512) 136448
_________________________________________________________________
dense (Dense) multiple 5130
=================================================================
Total params: 285,514
Trainable params: 280,650
Non-trainable params: 4,864
_________________________________________________________________
None
Epoch 1/200
782/782 [==============================] - 114s 144ms/step - loss: 1.6013 - accuracy: 0.4112
Epoch 2/200
782/782 [==============================] - 114s 146ms/step - loss: 1.2068 - accuracy: 0.5633
Epoch 3/200
782/782 [==============================] - 114s 146ms/step - loss: 1.0066 - accuracy: 0.6420
- InceptionNet網絡結構
import tensorflow as tf from tensorflow.keras import datasets, models, Sequential, layers, losses, optimizers if __name__ == '__main__': (X_train, y_train), (X_test, y_test) = datasets.cifar10.load_data() X_train = tf.constant(X_train, dtype=tf.float32) / 255 y_train = tf.constant(y_train, dtype=tf.int32) X_test = tf.constant(X_test, dtype=tf.float32) / 255 y_test = tf.constant(y_test, dtype=tf.int32) data_train = tf.data.Dataset.from_tensor_slices((X_train, y_train)) data_test = tf.data.Dataset.from_tensor_slices((X_test, y_test)) data_train = data_train.shuffle(10000).batch(64) data_test = data_test.shuffle(10000).batch(64) class BaseInception(models.Model): def ConvBNRelu(self, out_channel, kernel_size): return Sequential([ layers.Conv2D(out_channel, (kernel_size, kernel_size), strides=1, padding='same'), layers.BatchNormalization(), layers.ReLU() ]) def __init__(self, out_channel_list, reduce_channel_list): super(BaseInception, self).__init__() self.branch1_conv = self.ConvBNRelu(out_channel_list[0], 1) self.branch2_conv1 = self.ConvBNRelu(reduce_channel_list[0], 1) self.branch2_conv2 = self.ConvBNRelu(out_channel_list[1], 3) self.branch3_conv1 = self.ConvBNRelu(reduce_channel_list[1], 1) self.branch3_conv2 = self.ConvBNRelu(out_channel_list[2], 5) self.branch4_pool = layers.MaxPool2D((3, 3), strides=1, padding='same') self.branch4_conv = self.ConvBNRelu(out_channel_list[3], 3) def call(self, x): out1 = self.branch1_conv(x) out2 = self.branch2_conv1(x) out2 = self.branch2_conv2(out2) out3 = self.branch3_conv1(x) out3 = self.branch3_conv2(out3) out4 = self.branch4_pool(x) out4 = self.branch4_conv(out4) out = tf.concat([out1, out2, out3, out4], axis=1) return out class InceptionNet(models.Model): def __init__(self): super(InceptionNet, self).__init__() self.block1 = Sequential([ layers.Conv2D(64, (3, 3), strides=1, padding='same'), layers.BatchNormalization(), layers.ReLU() ]) self.block2 = Sequential([ layers.Conv2D(128, (3, 3), strides=2, padding='same'), layers.BatchNormalization(), layers.ReLU() ]) self.block3 = Sequential([ BaseInception(out_channel_list=[64, 64, 64, 64], reduce_channel_list=[16, 16]), layers.MaxPool2D((3, 3), strides=2, padding='same') ]) self.block4 = Sequential([ BaseInception(out_channel_list=[96, 96, 96, 96], reduce_channel_list=[32, 32]), layers.MaxPool2D((3, 3), strides=2, padding='same') ]) self.fc = layers.Dense(10) def call(self, x): out = self.block1(x) out = self.block2(out) out = self.block3(out) out = self.block4(out) out = layers.AvgPool2D((2, 2))(out) out = tf.reshape(out, (out.shape[0], -1)) out = self.fc(out) return out epoch_num = 200 lr = 0.001 net = InceptionNet() loss_func = losses.SparseCategoricalCrossentropy(from_logits=True) scheduler = optimizers.schedules.ExponentialDecay(initial_learning_rate=lr, decay_steps=3905, decay_rate=0.9) optimizer = optimizers.Adam(learning_rate=scheduler) for epoch in range(epoch_num): for i, (images, labels) in enumerate(data_train): with tf.GradientTape() as tape: outputs = net(images) labels = tf.squeeze(labels, axis=-1) loss = loss_func(labels, outputs) grads = tape.gradient(loss, net.trainable_variables) optimizer.apply_gradients(zip(grads, net.trainable_variables)) pred = tf.argmax(outputs, axis=-1) pred = tf.cast(pred, tf.int32) correct = tf.where(pred == labels).shape[0] print("epoch is ", epoch, "step ", i, "loss is: ", loss.numpy(), "mini-batch correct is: ", 100.0 * correct / 64)
- 使用預訓練模型
import tensorflow as tf from tensorflow.keras import datasets, models, losses, optimizers, applications if __name__ == '__main__': (X_train, y_train), (X_test, y_test) = datasets.cifar10.load_data() X_train = tf.constant(X_train, dtype=tf.float32) / 255 y_train = tf.constant(y_train, dtype=tf.int32) X_test = tf.constant(X_test, dtype=tf.float32) / 255 y_test = tf.constant(y_test, dtype=tf.int32) data_train = tf.data.Dataset.from_tensor_slices((X_train, y_train)) data_test = tf.data.Dataset.from_tensor_slices((X_test, y_test)) data_train = data_train.shuffle(10000).batch(64) data_test = data_test.shuffle(10000).batch(64) class ResNet50(models.Model): def __init__(self): super(ResNet50, self).__init__() self.model = applications.ResNet50(input_shape=(32, 32, 3), classes=10, weights=None) def call(self, x): out = self.model(x) return out epoch_num = 200 lr = 0.001 net = ResNet50() loss_func = losses.SparseCategoricalCrossentropy(from_logits=True) scheduler = optimizers.schedules.ExponentialDecay(initial_learning_rate=lr, decay_steps=3905, decay_rate=0.9) optimizer = optimizers.Adam(learning_rate=scheduler) for epoch in range(epoch_num): for i, (images, labels) in enumerate(data_train): with tf.GradientTape() as tape: outputs = net(images) labels = tf.squeeze(labels, axis=-1) loss = loss_func(labels, outputs) grads = tape.gradient(loss, net.trainable_variables) optimizer.apply_gradients(zip(grads, net.trainable_variables)) pred = tf.argmax(outputs, axis=-1) pred = tf.cast(pred, tf.int32) correct = tf.where(pred == labels).shape[0] print("epoch is ", epoch, "step ", i, "loss is: ", loss.numpy(), "mini-batch correct is: ", 100.0 * correct / 64)
圖像增強
import tensorflow as tf from tensorflow.keras import preprocessing import numpy as np import matplotlib.pyplot as plt if __name__ == '__main__': trans = preprocessing.image.ImageDataGenerator( featurewise_center=True, rotation_range=75, horizontal_flip=True, height_shift_range=0.2 ) image = plt.imread("/Users/admin/Documents/444.jpeg") image = tf.constant(image) image = tf.reshape(image, (1, 1279, 1896, 3)) result = trans.flow(image) image_out = result.next() image_out = np.reshape(image_out, (1279, 1896, 3)) image_out = image_out.astype(np.uint8) print(image_out) plt.imshow(image_out) plt.show()
運行結果
[[[28 27 25]
[23 22 20]
[20 19 17]
...
[ 8 2 2]
[ 8 2 2]
[ 8 2 2]]
[[30 29 27]
[29 28 26]
[24 23 21]
...
[ 8 2 2]
[ 8 2 2]
[ 8 2 2]]
[[32 31 29]
[30 29 27]
[29 28 26]
...
[ 8 2 2]
[ 8 2 2]
[ 8 2 2]]
...
[[29 22 16]
[28 19 13]
[28 19 12]
...
[19 19 11]
[19 19 11]
[20 17 8]]
[[28 20 13]
[28 19 12]
[29 20 13]
...
[19 19 11]
[19 19 11]
[19 19 11]]
[[28 19 12]
[29 20 13]
[31 22 15]
...
[19 19 11]
[19 19 11]
[19 19 11]]]
