Question

我不确定为什么会出现此语法错误，因为我在另一件事上使用了相同的train方法，但是没有错误。我该如何解决？

import copy
import os
import math
import numpy as np
import scipy
import scipy.io

from six.moves import range

import read_data

@read_data.restartable
def svhn_dataset_generator(dataset_name, batch_size):
    assert dataset_name in ['train', 'test']
    assert batch_size > 0 or batch_size == -1  # -1 for entire dataset

    path = './svhn_mat/' # path to the SVHN dataset you will download in Q1.1
    file_name = '%s_32x32.mat' % dataset_name
    file_dict = scipy.io.loadmat(os.path.join(path, file_name))
    X_all = file_dict['X'].transpose((3, 0, 1, 2))
    y_all = file_dict['y']
    data_len = X_all.shape[0]
    batch_size = batch_size if batch_size > 0 else data_len

    X_all_padded = np.concatenate([X_all, X_all[:batch_size]], axis=0)
    y_all_padded = np.concatenate([y_all, y_all[:batch_size]], axis=0)
    y_all_padded[y_all_padded == 10] = 0

    for slice_i in range(int(math.ceil(data_len / batch_size))):
        idx = slice_i * batch_size
        X_batch = X_all_padded[idx:idx + batch_size]
        y_batch = np.ravel(y_all_padded[idx:idx + batch_size])
        yield X_batch, y_batch



def apply_classification_loss(model_function):
    with tf.Graph().as_default() as g:
        #with tf.device("/gpu:0"):  # use gpu:0 if on GPU
        x_ = tf.placeholder(tf.float32, [None, 32, 32, 3])
        y_ = tf.placeholder(tf.int32, [None])
        y_logits = model_function(x_)

        y_dict = dict(labels=y_, logits=y_logits)
        losses = tf.nn.sparse_softmax_cross_entropy_with_logits(**y_dict)
        cross_entropy_loss = tf.reduce_mean(losses)
        trainer = tf.train.AdamOptimizer(learning_rate=0.001)
        train_op = trainer.minimize(cross_entropy_loss)

        y_pred = tf.argmax(tf.nn.softmax(y_logits), axis=1)
        correct_prediction = tf.equal(tf.cast(y_pred, tf.int32), y_)
        accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

    model_dict = {'graph': g, 'inputs': [x_, y_], 'train_op': train_op,
                  'accuracy': accuracy, 'loss': cross_entropy_loss}

    return model_dict




def train_model(model_dict, dataset_generators, epoch_n, print_every):
    with model_dict['graph'].as_default(), tf.Session() as sess:
        with tf.device("/gpu:0"): 
            sess.run(tf.global_variables_initializer())

            for epoch_i in range(epoch_n):
                for iter_i, data_batch in enumerate(dataset_generators['train']):
                    train_feed_dict = dict(zip(model_dict['inputs'], data_batch))
                    sess.run(model_dict['train_op'], feed_dict=train_feed_dict)

                    if iter_i % print_every == 0:
                        collect_arr = []
                        for test_batch in dataset_generators['test']:
                            test_feed_dict = dict(zip(model_dict['inputs'], test_batch))
                            to_compute = [model_dict['loss'], model_dict['accuracy']]
                            collect_arr.append(sess.run(to_compute, test_feed_dict))
                        averages = np.mean(collect_arr, axis=0)
                        fmt = (epoch_i, iter_i, ) + tuple(averages)
                        print('epoch {:d} iter {:d}, loss: {:.3f}, '
                              'accuracy: {:.3f}'.format(*fmt))


dataset_generators = {
        'train': svhn_dataset_generator('train', 256),
        'test': svhn_dataset_generator('test', 256)
}






def cnn_expanded(x_):
    conv1 = tf.layers.conv2d(
            inputs=x_,
            filters=32,  # number of filters
            kernel_size=[5, 5],
            padding="same",
            activation=tf.nn.relu)

    pool1 = tf.layers.max_pooling2d(inputs=conv1, 
                                    pool_size=[2, 2], 
                                    strides=2)  # convolution stride

    conv2 = tf.layers.conv2d(
            inputs=pool1,
            filters=32, # number of filters
            kernel_size=[9, 9],
            padding="same",
            activation=tf.nn.relu)
    pool2 = tf.layers.max_pooling2d(inputs=conv2, 
                                    pool_size=[2, 2], 
                                    strides=2)  # convolution stride


    conv3 = tf.layers.conv2d(
            inputs=pool1,
            filters=32,
            kernel_size=[9,9],
            padding="same",
            activation=tf.nn.relu)
    pool3 = tf.layers.max_pooling2d(inputs=conv2,
                                   pool_size=[2,2],
                                   strides=2) #convolution stride
    pool_flat = tf.contrib.layers.flatten(pool3, scope='pool3flat')


    #pool_flat = tf.contrib.layers.flatten(pool2, scope='pool2flat')
    dense = tf.layers.dense(inputs=pool_flat, units=500, activation=tf.nn.relu)
    logits = tf.layers.dense(inputs=dense, units=10)
    return logits




model_dict = apply_classification_loss(cnn_expanded)
train_model(model_dict, dataset_generators, epoch_n=50, print_every=20)

错误是：

$ python cnn_expansion.py 
  File "cnn_expansion.py", line 63
    with model_dict['graph'].as_default(), tf.Session() as sess:
                                         ^
SyntaxError: invalid syntax

Answer 1

如果我了解您在做什么，则需要做

with model_dict['graph'].as_default() as g:
    with tf.Session() as sess:

而不是您当前正在做什么。

您正在使用的代码实际上是在试图说

with (model_dict['graph'].as_default(), tf.Session()) as sess

其中

python不知道如何解析，
即使知道如何解析它也不会起作用，因为生成的sess现在是一个元组，所以类似

sess.run（）

会失败

使用model_dict ['graph']。as_default（），tf.Session（）作为sess：^ SyntaxError：无效语法

1 个答案: