tensorflow生成tflite檔案不可用

keep_prob = tf.placeholder(dtype=tf.float32, shape=[], name='keep_prob')

    images = tf.placeholder(dtype=tf.float32, shape=[None, 64, 64, 1], name='image_batch')

    labels = tf.placeholder(dtype=tf.int64, shape=[None], name='label_batch')

    is_training = tf.placeholder(dtype=tf.bool, shape=[], name='train_flag')

    with tf.device('/cpu:0'):

        with slim.arg_scope([slim.conv2d, slim.fully_connected],

                            normalizer_fn=slim.batch_norm,

                            normalizer_params={'is_training': is_training}):

            conv3_1 = slim.conv2d(images, 64, [3, 3], 1, padding='SAME', scope='conv3_1')

            max_pool_1 = slim.max_pool2d(conv3_1, [2, 2], [2, 2], padding='SAME', scope='pool1')

            conv3_2 = slim.conv2d(max_pool_1, 128, [3, 3], padding='SAME', scope='conv3_2')

            max_pool_2 = slim.max_pool2d(conv3_2, [2, 2], [2, 2], padding='SAME', scope='pool2')

            conv3_3 = slim.conv2d(max_pool_2, 256, [3, 3], padding='SAME', scope='conv3_3')

            max_pool_3 = slim.max_pool2d(conv3_3, [2, 2], [2, 2], padding='SAME', scope='pool3')

            conv3_4 = slim.conv2d(max_pool_3, 512, [3, 3], padding='SAME', scope='conv3_4')

            conv3_5 = slim.conv2d(conv3_4, 512, [3, 3], padding='SAME', scope='conv3_5')

            max_pool_4 = slim.max_pool2d(conv3_5, [2, 2], [2, 2], padding='SAME', scope='pool4')



            flatten = slim.flatten(max_pool_4)

            fc1 = slim.fully_connected(slim.dropout(flatten, keep_prob), 1024,

                                       activation_fn=tf.nn.relu, scope='fc1')

            logits = slim.fully_connected(slim.dropout(fc1, keep_prob), FLAGS.charset_size, activation_fn=None,

                                          scope='fc2')

        loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=labels))

        accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(logits, 1), labels), tf.float32))



        update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)

        if update_ops:

            updates = tf.group(*update_ops)

            loss = control_flow_ops.with_dependencies([updates], loss)



        global_step = tf.get_variable("step", [], initializer=tf.constant_initializer(0.0), trainable=False)

        optimizer = tf.train.AdamOptimizer(learning_rate=0.1)

        train_op = slim.learning.create_train_op(loss, optimizer, global_step=global_step)

        probabilities = tf.nn.softmax(logits)



        tf.summary.scalar('loss', loss)

        tf.summary.scalar('accuracy', accuracy)

        merged_summary_op = tf.summary.merge_all()

        predicted_val_top_k, predicted_index_top_k = tf.nn.top_k(probabilities, k=top_k)

        accuracy_in_top_k = tf.reduce_mean(tf.cast(tf.nn.in_top_k(probabilities, labels, top_k), tf.float32))

output_node_names = "fc2/BatchNorm/Reshape_1"

saver = tf.train.import_meta_graph(input_checkpoint + '.meta', clear_devices=False)

    with tf.Session() as sess:

        saver.restore(sess, input_checkpoint)  # 恢復圖并得到資料

        output_graph_def = graph_util.convert_variables_to_constants(  # 模型持久化，將變數值固定

            sess=sess,

            input_graph_def=sess.graph_def,  # 等于:sess.graph_def

            output_node_names=output_node_names.split(","))  # 如果有多個輸出節點，以逗號隔開



        all_output_node_names = [n.name for n in output_graph_def.node]

        print(str(all_output_node_names))



        pb_file1 = './output_graph-7.pb'

        with tf.gfile.GFile(pb_file1, "wb") as f:  # 保存模型

            f.write(output_graph_def.SerializeToString())  # 序列化輸出

def get_tflite_file(pb_file_path):

    input_names = ["image_batch", "train_flag", "keep_prob"]

    output_names = ["fc2/BatchNorm/Reshape_1"]

    input_tensor = {"image_batch": [1, float(64), float(64), 1],  "train_flag": [True],  "keep_prob": [4]}

    convert = tf.lite.TFLiteConverter.from_frozen_graph(pb_file_path, input_arrays=input_names, output_arrays=output_names,input_shapes=input_tensor)

    convert.post_training_quantize = True

    convert.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS, tf.lite.OpsSet.SELECT_TF_OPS]

    convert.allow_custom_ops = True

    convert.default_ranges_stats = (0, 255)

    tflite_model = convert.convert()

    open("../bak/output_graph-1.tflite", "wb").write(tflite_model)

標籤：腳本語言(Perl/Python)

上一篇：如何處理TypeError: 'type' object is not subscriptable

下一篇：求大佬指路小白。