Tensorflow1.0を2.0に書き換える際、minimizeにvar_listとして何を与えれば良いか

やりたいこと

このサイトのナンバープレート認識プログラムの内、train.pyを動かしたいのです。しかしTensorflow1.xで書かれているため、Tensorflow2.xに書き換えている所です。

その内の一つのエラーのとり方がわからないので、教えていただきたいです。

optimizersのminimize関数へvar_listを渡さなければならないところで、何を最小化したいのか理解できないです…????

エラーメッセージ

cmd
1WARNING:tensorflow:From C:\Users\main9\anaconda3\envs\kmutt\lib\site-packages\tensorflow_core\python\ops\resource_variable_ops.py:1635: calling BaseResourceVariable.__init__ (from tensorflow.python.ops.resource_variable_ops) with constraint is deprecated and will be removed in a future version.
2Instructions for updating:
3If using Keras pass *_constraint arguments to layers.
4Traceback (most recent call last):
5  File "train.py", line 269, in <module>
6    initial_weights=initial_weights)
7  File "train.py", line 178, in train
8    train_step = tf.optimizers.Adam(learn_rate).minimize(loss)
9TypeError: minimize() missing 1 required positional argument: 'var_list'

プログラム

該当部分：

Python
1def train(learn_rate, report_steps, batch_size, initial_weights=None):
2    x, y, params = model.get_training_model()
3
4    # y_ = tf.placeholder(tf.float32, [None, 7 * len(common.CHARS) + 1])
5    tf.compat.v1.disable_eager_execution()
6    y_ = tf.compat.v1.placeholder(tf.float32, [None, 7 * len(common.CHARS) + 1])
7
8    digits_loss, presence_loss, loss = get_loss(y, y_)
9    # train_step = tf.train.AdamOptimizer(learn_rate).minimize(loss)
10    train_step = tf.optimizers.Adam(learn_rate).minimize(loss)
11
12    best = tf.argmax(tf.reshape(y[:, 1:], [-1, 7, len(common.CHARS)]), 2)
13    correct = tf.argmax(tf.reshape(y_[:, 1:], [-1, 7, len(common.CHARS)]), 2)
14
15    if initial_weights is not None:
16        assert len(params) == len(initial_weights)
17        assign_ops = [w.assign(v) for w, v in zip(params, initial_weights)]
18
19    init = tf.initialize_all_variables()
20

該当の関数全体（プログラム全体は文字数以内に入り切りませんでした(>人<)）：

Python
1def train(learn_rate, report_steps, batch_size, initial_weights=None):
2    """
3    Train the network.
4
5    The function operates interactively: Progress is reported on stdout, and
6    training ceases upon `KeyboardInterrupt` at which point the learned weights
7    are saved to `weights.npz`, and also returned.
8
9    :param learn_rate:
10        Learning rate to use.
11
12    :param report_steps:
13        Every `report_steps` batches a progress report is printed.
14
15    :param batch_size:
16        The size of the batches used for training.
17
18    :param initial_weights:
19        (Optional.) Weights to initialize the network with.
20
21    :return:
22        The learned network weights.
23
24    """
25    x, y, params = model.get_training_model()
26
27    # y_ = tf.placeholder(tf.float32, [None, 7 * len(common.CHARS) + 1])
28    tf.compat.v1.disable_eager_execution()
29    y_ = tf.compat.v1.placeholder(tf.float32, [None, 7 * len(common.CHARS) + 1])
30
31    digits_loss, presence_loss, loss = get_loss(y, y_)
32    # train_step = tf.train.AdamOptimizer(learn_rate).minimize(loss)
33    train_step = tf.optimizers.Adam(learn_rate).minimize(loss)
34
35    best = tf.argmax(tf.reshape(y[:, 1:], [-1, 7, len(common.CHARS)]), 2)
36    correct = tf.argmax(tf.reshape(y_[:, 1:], [-1, 7, len(common.CHARS)]), 2)
37
38    if initial_weights is not None:
39        assert len(params) == len(initial_weights)
40        assign_ops = [w.assign(v) for w, v in zip(params, initial_weights)]
41
42    init = tf.initialize_all_variables()
43
44    def vec_to_plate(v):
45        return "".join(common.CHARS[i] for i in v)
46
47    def do_report():
48        r = sess.run([best,
49                      correct,
50                      tf.greater(y[:, 0], 0),
51                      y_[:, 0],
52                      digits_loss,
53                      presence_loss,
54                      loss],
55                     feed_dict={x: test_xs, y_: test_ys})
56        num_correct = numpy.sum(
57                        numpy.logical_or(
58                            numpy.all(r[0] == r[1], axis=1),
59                            numpy.logical_and(r[2] < 0.5,
60                                              r[3] < 0.5)))
61        r_short = (r[0][:190], r[1][:190], r[2][:190], r[3][:190])
62        for b, c, pb, pc in zip(*r_short):
63            print("{} {} <-> {} {}".format(vec_to_plate(c), pc,
64                                           vec_to_plate(b), float(pb)) )
65        num_p_correct = numpy.sum(r[2] == r[3])
66
67        print ("B{:3d} {:2.02f}% {:02.02f}% loss: {} "
68               "(digits: {}, presence: {}) |{}|").format(
69            batch_idx,
70            100. * num_correct / (len(r[0])),
71            100. * num_p_correct / len(r[2]),
72            r[6],
73            r[4],
74            r[5],
75            "".join("X "[numpy.array_equal(b, c) or (not pb and not pc)]
76                                           for b, c, pb, pc in zip(*r_short)))
77
78    def do_batch():
79        sess.run(train_step,
80                 feed_dict={x: batch_xs, y_: batch_ys})
81        if batch_idx % report_steps == 0:
82            do_report()
83
84    gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.95)
85    with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
86        sess.run(init)
87        if initial_weights is not None:
88            sess.run(assign_ops)
89
90        test_xs, test_ys = unzip(list(read_data("test/*.png"))[:50])
91
92        try:
93            last_batch_idx = 0
94            last_batch_time = time.time()
95            batch_iter = enumerate(read_batches(batch_size))
96            for batch_idx, (batch_xs, batch_ys) in batch_iter:
97                do_batch()
98                if batch_idx % report_steps == 0:
99                    batch_time = time.time()
100                    if last_batch_idx != batch_idx:
101                        print ("time for 60 batches {}".format(
102                            60 * (last_batch_time - batch_time) /
103                                            (last_batch_idx - batch_idx)) )
104                        last_batch_idx = batch_idx
105                        last_batch_time = batch_time
106
107        except KeyboardInterrupt:
108            last_weights = [p.eval() for p in params]
109            numpy.savez("weights.npz", *last_weights)
110            return last_weights

損失関数：

Python
1def get_loss(y, y_):
2    # Calculate the loss from digits being incorrect.  Don't count loss from
3    # digits that are in non-present plates.
4    digits_loss = tf.nn.softmax_cross_entropy_with_logits(
5                                          tf.reshape(y[:, 1:],
6                                                     [-1, len(common.CHARS)]),
7                                          tf.reshape(y_[:, 1:],
8                                                     [-1, len(common.CHARS)]))
9    digits_loss = tf.reshape(digits_loss, [-1, 7])
10    digits_loss = tf.reduce_sum(digits_loss, 1)
11    digits_loss *= (y_[:, 0] != 0)
12    digits_loss = tf.reduce_sum(digits_loss)
13
14    # Calculate the loss from presence indicator being wrong.
15    presence_loss = tf.nn.sigmoid_cross_entropy_with_logits(
16                                                          y[:, :1], y_[:, :1])
17    presence_loss = 7 * tf.reduce_sum(presence_loss)
18
19    return digits_loss, presence_loss, digits_loss + presence_loss