大致如下，我建立了一个5层conv，2层fc的深度网络对cifar100进行学习。epoch=50，adam(lr=1e-3)
我将60k的数据，分为50k为train，然后10k为test
其中再把50k中分为train_data 和 validation_data，比例为9：1
当我没有分出来validation数据的时候，运行的很顺畅，但是当我在每一次epoch开始的时候随机打乱数据分为train和validation数据后，就开始出现问题了。当epoch为2时候，在validation上的acc竟然已经开始达到100%了 ，最后在test上的acc却只为1%。我实在是新手，也找不到bug出在哪儿，希望大佬们可以赐教。
代码如下（我是在colab上跑的）
import tensorflow as tf
from tensorflow.keras import layers, optimizers, datasets, Sequential
import os
tf.random.set_seed(2345)
conv_layers = [
# 5 units of conv + max pooling
# unit 1
# 64 means: every conv layer has 64 kernels
layers.Conv2D(64, kernel_size=[3,3], padding='same', activation=tf.nn.relu),
layers.Conv2D(64, kernel_size=[3,3], padding='same', activation=tf.nn.relu),
layers.MaxPool2D(pool_size=[2,2], strides=2, padding='same'),
# unit 2
layers.Conv2D(128, kernel_size=[3,3], padding='same', activation=tf.nn.relu),
layers.Conv2D(128, kernel_size=[3,3], padding='same', activation=tf.nn.relu),
layers.MaxPool2D(pool_size=[2,2], strides=2, padding='same'),
# unit 3
layers.Conv2D(256, kernel_size=[3,3], padding='same', activation=tf.nn.relu),
layers.Conv2D(256, kernel_size=[3,3], padding='same', activation=tf.nn.relu),
layers.MaxPool2D(pool_size=[2,2], strides=2, padding='same'),
# unit 4
layers.Conv2D(512, kernel_size=[3,3], padding='same', activation=tf.nn.relu),
layers.Conv2D(512, kernel_size=[3,3], padding='same', activation=tf.nn.relu),
layers.MaxPool2D(pool_size=[2,2], strides=2, padding='same'),
# unit 5
layers.Conv2D(512, kernel_size=[3,3], padding='same', activation=tf.nn.relu),
layers.Conv2D(512, kernel_size=[3,3], padding='same', activation=tf.nn.relu),
layers.MaxPool2D(pool_size=[2,2], strides=2, padding='same'),
]
fc_layers = [
layers.Dense(256, activation=tf.nn.relu),
layers.Dense(256, activation=tf.nn.relu),
layers.Dense(100, activation=None),
]
def preprocess(x,y):
x = tf.cast(x,dtype=tf.float32)/255. #[0-1]
y = tf.cast(y,dtype=tf.int32)
return x,y
# load data
# x:[50k,32,32,3], x_test:[10k,32,32,3]
(x,y),(x_test,y_test) = datasets.cifar100.load_data()
y = tf.squeeze(y,axis = 1)
y_test = tf.squeeze(y_test,axis = 1)
test_db = tf.data.Dataset.from_tensor_slices((x_test, y_test))
test_db = test_db.map(preprocess).batch(64)
#[b,32,32,3]=>[b,1,1,512]
conv_net = Sequential(conv_layers)
conv_net.build(input_shape=[None,32,32,3])
fc_net = Sequential(fc_layers)
fc_net.build(input_shape=[None,512])
optimizer = optimizers.Adam(lr = 1e-4)
# combine the variables of the conv net and fully connected net
variables = conv_net.trainable_variables + fc_net.trainable_variables
for epoch in range(50):
idx = tf.range(50000)
idx = tf.random.shuffle(idx)
# train_num : validation_num => 0.9: 0.1
x_train, y_train = tf.gather(x, idx[:45000]), tf.gather(y, idx[:45000])
x_val, y_val = tf.gather(x, idx[-5000:]), tf.gather(y, idx[-5000:])
print('train data', x_train.shape, y_train.shape, 'validation data',x_val.shape, y_val.shape)
train_db = tf.data.Dataset.from_tensor_slices((x_train,y_train))
train_db = train_db.map(preprocess).shuffle(10000).batch(64)
val_db = tf.data.Dataset.from_tensor_slices((x_val,y_val))
val_db = val_db.map(preprocess).shuffle(10000).batch(64)
for step, (x,y) in enumerate(train_db):
with tf.GradientTape() as tape:
# [b,32,32,3] => [b,1,1,512]
out = conv_net(x)
# [b,1,1,512] => [b,512] flatten
out = tf.reshape(out,[-1,512])
# [b,512] => [b,100]
logits = fc_net(out)
# y: [b,]=>[b,100]
y_onehot = tf.one_hot(y, depth=100)
# compute loss
loss = tf.losses.categorical_crossentropy(y_onehot,logits, from_logits=True)
loss = tf.reduce_mean(loss)
grads = tape.gradient(loss, variables)
optimizer.apply_gradients(zip(grads, variables))
if step%100 == 0:
print('epoch:',epoch,'step:',step,'loss:',float(loss))
# validation part
total_num = 0
total_correct = 0
for x,y in val_db:
out = conv_net(x)
out = tf.reshape(out,[-1,512])
logits = fc_net(out)
prob = tf.nn.softmax(logits, axis=1)
pred = tf.argmax(prob, axis=1)
pred = tf.cast(pred, dtype=tf.int32)
correct = tf.cast(tf.equal(pred,y),dtype=tf.int32)
correct = tf.reduce_sum(correct)
total_num += x.shape[0]
total_correct += int(correct)
acc = total_correct / total_num
print('in validation data','epoch:',epoch,'acc=',acc)
# use test data to do the prediction
total_num = 0
total_correct = 0
for x,y in test_db:
out = conv_net(x)
out = tf.reshape(out,[-1,512])
logits = fc_net(out)
prob = tf.nn.softmax(logits, axis=1)
pred = tf.argmax(prob, axis=1)
pred = tf.cast(pred, dtype=tf.int32)
correct = tf.cast(tf.equal(pred,y),dtype=tf.int32)
correct = tf.reduce_sum(correct)
total_num += x.shape[0]
total_correct += int(correct)
acc = total_correct / total_num
print('in test data','acc=',acc)