作為練習，我正在嘗試使用 Tensorflow v1 訓練分類模型，而不使用 keras、sklearn 或任何其他可以大大減輕我生活的庫..大聲笑。
我不斷收到此錯誤，但我無法弄清楚我的代碼有什么問題：
# imports
import numpy as np # linear algebra
import pandas as pd # data processing, CSV file I/O
import os,shutil, cv2, itertools
import tensorflow.compat.v1 as tf
tf.disable_v2_behavior() # use Tensorflow v1
為簡單起見，我將每個影像的大小調整為 64 x 64。
由于有 3 個通道，因此輸入向量 X wiz n 的大小為 64 x 64 x 3 (=12288)。
為了減少計算時間，我只使用了訓練資料中的 3000 張影像。
TRAIN_DATA_DIR_PREFIX = '/kaggle/input/cat-and-dog/training_set/training_set/'
TRAIN_DATA_DOG_DIR = TRAIN_DATA_DIR_PREFIX 'dogs/'
TRAIN_DATA_CAT_DIR = TRAIN_DATA_DIR_PREFIX 'cats/'
N_x = 64 * 64 * 3
TRAIN_DOG = [(TRAIN_DATA_DOG_DIR i, 1) for i in os.listdir(TRAIN_DATA_DOG_DIR) if i.endswith(".jpg")]
TRAIN_DOG = TRAIN_DOG[:1500]
TRAIN_CAT = [(TRAIN_DATA_CAT_DIR i, 0) for i in os.listdir(TRAIN_DATA_CAT_DIR) if i.endswith(".jpg")]
TRAIN_CAT = TRAIN_CAT[:1500]
TRAIN_DATA = TRAIN_DOG TRAIN_CAT
print('TRAIN_DATA :', len(TRAIN_DATA))
def read_image(file_path):
image = cv2.imread(file_path, cv2.IMREAD_COLOR)
return cv2.resize(image, (64, 64),interpolation=cv2.INTER_CUBIC)
def prepare(data):
X = np.ndarray((N_x,len(TRAIN_DATA)), dtype=np.uint8)
print ("X shape is {}".format(X.shape))
Y = np.zeros((1,len(TRAIN_DATA)))
print ("Y shape is {}".format(Y.shape))
for i,(image_path, y) in enumerate(data):
Y[0,i] = y
image = read_image(image_path)
X[:,i] = np.squeeze(image.reshape((N_x,1)))
return X,Y
X,Y = prepare(TRAIN_DATA)
輸出：
TRAIN_DATA : 3000
X shape is (12288, 3000)
Y shape is (1, 3000)
然后我定義我的 tf_train 函式：
def tf_train(X, Y, batch_size=20):
# Dataset (inputs and labels)
D=200
x = tf.placeholder(tf.float32, [batch_size, D])
y_ = tf.placeholder(tf.float32, [batch_size, 1])
# random variable
W = tf.Variable(tf.random_normal([D, 1],stddev=0.1))
z=tf.matmul(x,W)
# map x to inferred output function
# a = tf.sigmoid(tf.matmul(x, W))
# Define loss and optimizer
# cross_entropy = tf.reduce_mean(-(y_*tf.log(a) (1-y_)*tf.log(1-a)))
cross_entropy = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=z, labels=y_))
train_step = tf.train.GradientDescentOptimizer(1e-2).minimize(cross_entropy)
sess = tf.InteractiveSession()
tf.initialize_all_variables().run() #initializes W and a
# Train
for epoch in range (3000):
idx = np.random.choice(len(TRAIN_DATA), batch_size, replace=False)
#idx = np.random.permutation(np.arange(X.shape[1])) [:batch_size]
_, l = sess.run([train_step, cross_entropy], feed_dict={x: X[idx], y_: Y[0,idx]})
if epoch%100 == 0:
print('loss: ' str(l))
然后當我開始訓練我的模型時：
tf_train(X,Y)
我得到以下輸出：
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
/tmp/ipykernel_34/1897119706.py in <module>
----> 1 tf_train(X,Y)
/tmp/ipykernel_34/2072834761.py in tf_train(X, Y, batch_size)
28 idx = np.random.choice(len(TRAIN_DATA), batch_size, replace=False)
29 #idx = np.random.permutation(np.arange(X.shape[1])) [:batch_size]
---> 30 _, l = sess.run([train_step, cross_entropy], feed_dict={x: X[idx], y_: Y[0,idx]})
31 if epoch%100 == 0:
32 print('loss: ' str(l))
/opt/conda/lib/python3.7/site-packages/tensorflow/python/client/session.py in run(self, fetches, feed_dict, options, run_metadata)
966 try:
967 result = self._run(None, fetches, feed_dict, options_ptr,
--> 968 run_metadata_ptr)
969 if run_metadata:
970 proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)
/opt/conda/lib/python3.7/site-packages/tensorflow/python/client/session.py in _run(self, handle, fetches, feed_dict, options, run_metadata)
1165 'Cannot feed value of shape %r for Tensor %r, '
1166 'which has shape %r' %
-> 1167 (np_val.shape, subfeed_t.name, str(subfeed_t.get_shape())))
1168 if not self.graph.is_feedable(subfeed_t):
1169 raise ValueError('Tensor %s may not be fed.' % subfeed_t)
ValueError: Cannot feed value of shape (20, 3000) for Tensor 'Placeholder:0', which has shape '(20, 200)'
為什么我不斷收到此錯誤訊息？
uj5u.com熱心網友回復：
您需要重塑資料，因為目前它被解釋為(12288, 3000)12288 個樣本，每個樣本有 3000 個特征。你可能想要形狀(3000, 12288)甚至(3000, 64, 64, 3)，所以試試這個：
import tensorflow as tf
X = tf.random.normal((12288, 3000))
Y = tf.random.normal((1, 3000))
X = tf.reshape(X, (3000, 64, 64, 3)) # or tf.transpose(X) if you want (3000, 12288)
Y = tf.transpose(Y)
print(X.shape)
print(Y.shape)
(3000, 64, 64, 3)
(3000, 1)
你的代碼導致了實際問題：
D = 200
x = tf.placeholder(tf.float32, [batch_size, D])
D 應為 12288 或 64 x 64 x 3，具體取決于您將如何重塑資料。
轉載請註明出處，本文鏈接：https://www.uj5u.com/houduan/372467.html