MAML 源代码解释说明 (一)

元学习系列文章

1. optimization based meta-learning
   1. 《Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks》 论文翻译笔记
   2. 元学习方向 optimization based meta learning 之 MAML论文详细解读
   3. MAML 源代码解释说明 (一)：本篇博客
   4. MAML 源代码解释说明 (二)
   5. 元学习之《On First-Order Meta-Learning Algorithms》论文详细解读
   6. 元学习之《OPTIMIZATION AS A MODEL FOR FEW-SHOT LEARNING》论文详细解读
2. metric based meta-learning: 待更新…
3. model based meta-learning: 待更新…

前言

此篇是对 MAML 源代码的解释，作者开源了论文代码，但是代码中注释很少，刚开始不容易理清思路，所以对代码中的关键部分进行了解释说明，核心是 construct_model() 函数，里面包含了 MAML 的训练过程，看代码实现能够更清楚地理解作者的思想。

maml.py

maml.py

""" Code for the MAML algorithm and network definitions. """
from __future__ import print_function
import numpy as np
import sys
import tensorflow as tf
try:
    import special_grads
except KeyError as e:
    print('WARN: Cannot define MaxPoolGrad, likely already defined for this version of tensorflow: %s' % e,
          file=sys.stderr)

from tensorflow.python.platform import flags
from utils import mse, xent, conv_block, normalize

FLAGS = flags.FLAGS

class MAML:
    def __init__(self, dim_input=1, dim_output=1, test_num_updates=5):
        """ must call construct_model() after initializing MAML! """
        self.dim_input = dim_input
        self.dim_output = dim_output
        self.update_lr = FLAGS.update_lr
        self.meta_lr = tf.placeholder_with_default(FLAGS.meta_lr, ())
        self.classification = False
        self.test_num_updates = test_num_updates
        if FLAGS.datasource == 'sinusoid':
            self.dim_hidden = [40, 40]
            self.loss_func = mse
            self.forward = self.forward_fc
            self.construct_weights = self.construct_fc_weights
        elif FLAGS.datasource == 'omniglot' or FLAGS.datasource == 'miniimagenet':
            self.loss_func = xent
            self.classification = True
            if FLAGS.conv:
                self.dim_hidden = FLAGS.num_filters
                self.forward = self.forward_conv
                self.construct_weights = self.construct_conv_weights
            else:
                self.dim_hidden = [256, 128, 64, 64]
                self.forward=self.forward_fc
                self.construct_weights = self.construct_fc_weights
            if FLAGS.datasource == 'miniimagenet':
                self.channels = 3
            else:
                self.channels = 1
            self.img_size = int(np.sqrt(self.dim_input/self.channels))
        else:
            raise ValueError('Unrecognized data source.')

    # ************************* 模型训练图的构建过程，此函数是核心代码 *************************
    def construct_model(self, input_tensors=None, prefix='metatrain_'):
        # a: training data for inner gradient, b: test data for meta gradient
        if input_tensors is None:
            self.inputa = tf.placeholder(tf.float32)
            self.inputb = tf.placeholder(tf.float32)
            self.labela = tf.placeholder(tf.float32)
            self.labelb = tf.placeholder(tf.float32)
        else:
            self.inputa = input_tensors['inputa']
            self.inputb = input_tensors['inputb']
            self.labela = input_tensors['labela']
            self.labelb = input_tensors['labelb']

        # 训练过程计算图
        with tf.variable_scope('model', reuse=None) as training_scope:
            # 如果不是第一次执行训练图， self 中则存在 self.weights 变量，那么所有的 tasks 都会共享这组 weights
            if 'weights' in dir(self):
                training_scope.reuse_variables()
                weights = self.weights
            else:
                # Define the weights
                # 第一次执行， weights 不在 dir(self) 中，则进行手动初始化
                self.weights = weights = self.construct_weights()

            # outputbs[i] and lossesb[i] is the output and loss after i+1 gradient updates
            lossesa, outputas, lossesb, outputbs = [], [], [], []
            accuraciesa, accuraciesb = [], []
            num_updates = max(self.test_num_updates, FLAGS.num_updates)
            outputbs = [[]]*num_updates
            lossesb = [[]]*num_updates
            accuraciesb = [[]]*num_updates

            def task_metalearn(inp, reuse=True):
                """ Perform gradient descent for one task in the meta-batch.
                    meta batch 个 task，并行执行 task_metalearn, 每个 task_metalearn 处理一个具体 task 的训练任务
                """
                inputa, inputb, labela, labelb = inp
                task_outputbs, task_lossesb = [], []

                if self.classification:
                    task_accuraciesb = []
                # inputa: [inner_batch, 1], task_outputa: [inner_batch, 1]
                task_outputa = self.forward(inputa, weights, reuse=reuse)  # only reuse on the first iter
                task_lossa = self.loss_func(task_outputa, labela)

                grads = tf.gradients(task_lossa, list(weights.values()))
                if FLAGS.stop_grad:
                    grads = [tf.stop_gradient(grad) for grad in grads]
                # w1: g1, w2: g2
                gradients = dict(zip(weights.keys(), grads))
                # w1: w1 - α*g1, w2: w2 - α*g2,
                fast_weights = dict(zip(weights.keys(), [weights[key] - self.update_lr*gradients[key] for key in weights.keys()]))
                # 使用更新后的 w, 在 inputb task 上再计算一次 meta
                output = self.forward(inputb, fast_weights, reuse=True)
                task_outputbs.append(output)
                task_lossesb.append(self.loss_func(output, labelb))

                # task 内部进行 num_updates 次更新，上面更新了一次，所以这里是 num_updates-1
                for j in range(num_updates - 1):
                    loss = self.loss_func(self.forward(inputa, fast_weights, reuse=True), labela)
                    grads = tf.gradients(loss, list(fast_weights.values()))
                    if FLAGS.stop_grad:
                        grads = [tf.stop_gradient(grad) for grad in grads]
                    gradients = dict(zip(fast_weights.keys(), grads))
                    fast_weights = dict(zip(fast_weights.keys(), [fast_weights[key] - self.update_lr*gradients[key] for key in fast_weights.keys()]))
                    output = self.forward(inputb, fast_weights, reuse=True)
                    task_outputbs.append(output)
                    task_lossesb.append(self.loss_func(output, labelb))

                # inputa 是训练集，inputb 是 task 的测试集
                # task_outputa 是第一次前向计算在 inputa 数据的输出，task_lossa 是基于 task_outputa 在参数 weight 上计算的 loss
                # task_outputbs 是每次梯度更新后参数在 inputb 数据上的输出，task_lossesb 是基于每个 task_outputb 计算出的 loss
                task_output = [task_outputa, task_outputbs, task_lossa, task_lossesb]

                if self.classification:
                    task_accuracya = tf.contrib.metrics.accuracy(tf.argmax(tf.nn.softmax(task_outputa), 1), tf.argmax(labela, 1))
                    for j in range(num_updates):
                        task_accuraciesb.append(tf.contrib.metrics.accuracy(tf.argmax(tf.nn.softmax(task_outputbs[j]), 1), tf.argmax(labelb, 1)))
                    task_output.extend([task_accuracya, task_accuraciesb])

                return task_output

            if FLAGS.norm is not 'None':
                # to initialize the batch norm vars, might want to combine this, and not run idx 0 twice.
                unused = task_metalearn((self.inputa[0], self.inputb[0], self.labela[0], self.labelb[0]), False)

            # [task_outputa, task_outputbs, task_lossa, task_lossesb]
            out_dtype = [tf.float32, [tf.float32]*num_updates, tf.float32, [tf.float32]*num_updates]
            if self.classification:
                out_dtype.extend([tf.float32, [tf.float32]*num_updates])
            result = tf.map_fn(task_metalearn, elems=(self.inputa, self.inputb, self.labela, self.labelb), dtype=out_dtype, parallel_iterations=FLAGS.meta_batch_size)
            if self.classification:
                outputas, outputbs, lossesa, lossesb, accuraciesa, accuraciesb = result
            else:
                outputas, outputbs, lossesa, lossesb = result

        ## Performance & Optimization
        ## 汇总 loss 函数作为 meta 的训练节点
        if 'train' in prefix:
            # lossesa 是 meta_batch_size 个具体任务在 inputa 数据上的第一次前向的 loss，
            self.total_loss1 = total_loss1 = tf.reduce_sum(lossesa) / tf.to_float(FLAGS.meta_batch_size)
            # lossesb[j] 是第 j 次更新时，meta_batch_size 个任务在 inputb 数据上的 loss
            self.total_losses2 = total_losses2 = [tf.reduce_sum(lossesb[j]) / tf.to_float(FLAGS.meta_batch_size) for j in range(num_updates)]
            # after the map_fn
            self.outputas, self.outputbs = outputas, outputbs
            if self.classification:
                self.total_accuracy1 = total_accuracy1 = tf.reduce_sum(accuraciesa) / tf.to_float(FLAGS.meta_batch_size)
                self.total_accuracies2 = total_accuracies2 = [tf.reduce_sum(accuraciesb[j]) / tf.to_float(FLAGS.meta_batch_size) for j in range(num_updates)]
            # pretrain 使用 inputa 数据上的第一次 loss 和，pretrain 相当于迁移学习的预训练
            self.pretrain_op = tf.train.AdamOptimizer(self.meta_lr).minimize(total_loss1)

            # metatrain, 使用 inputb 数据上最后一次前向计算出的平均 loss 作为优化目标
            if FLAGS.metatrain_iterations > 0:
                optimizer = tf.train.AdamOptimizer(self.meta_lr)
                # metatrain_op 最小化目标是每个 task 最后一次前向计算出的 loss 的平均值
                self.gvs = gvs = optimizer.compute_gradients(self.total_losses2[FLAGS.num_updates-1])
                if FLAGS.datasource == 'miniimagenet':
                    gvs = [(tf.clip_by_value(grad, -10, 10), var) for grad, var in gvs]
                self.metatrain_op = optimizer.apply_gradients(gvs)
        else:
            self.metaval_total_loss1 = total_loss1 = tf.reduce_sum(lossesa) / tf.to_float(FLAGS.meta_batch_size)
            self.metaval_total_losses2 = total_losses2 = [tf.reduce_sum(lossesb[j]) / tf.to_float(FLAGS.meta_batch_size) for j in range(num_updates)]
            if self.classification:
                self.metaval_total_accuracy1 = total_accuracy1 = tf.reduce_sum(accuraciesa) / tf.to_float(FLAGS.meta_batch_size)
                self.metaval_total_accuracies2 = total_accuracies2 =[tf.reduce_sum(accuraciesb[j]) / tf.to_float(FLAGS.meta_batch_size) for j in range(num_updates)]

        ## Summaries
        # total_loss1 是各个任务在 meta-update 之前的平均 loss
        tf.summary.scalar(prefix+'Pre-update loss', total_loss1)
        if self.classification:
            tf.summary.scalar(prefix+'Pre-update accuracy', total_accuracy1)

        for j in range(num_updates):
            tf.summary.scalar(prefix+'Post-update loss, step ' + str(j+1), total_losses2[j])
            if self.classification:
                tf.summary.scalar(prefix+'Post-update accuracy, step ' + str(j+1), total_accuracies2[j])

    ### Network construction functions (fc networks and conv networks)
    # 构建网络模型全连接层的参数
    def construct_fc_weights(self):
        weights = {}
        weights['w1'] = tf.Variable(tf.truncated_normal([self.dim_input, self.dim_hidden[0]], stddev=0.01))
        weights['b1'] = tf.Variable(tf.zeros([self.dim_hidden[0]]))
        for i in range(1,len(self.dim_hidden)):
            weights['w'+str(i+1)] = tf.Variable(tf.truncated_normal([self.dim_hidden[i-1], self.dim_hidden[i]], stddev=0.01))
            weights['b'+str(i+1)] = tf.Variable(tf.zeros([self.dim_hidden[i]]))
        weights['w'+str(len(self.dim_hidden)+1)] = tf.Variable(tf.truncated_normal([self.dim_hidden[-1], self.dim_output], stddev=0.01))
        weights['b'+str(len(self.dim_hidden)+1)] = tf.Variable(tf.zeros([self.dim_output]))
        return weights

    # 执行网络模型的前向计算过程
    def forward_fc(self, inp, weights, reuse=False):
        hidden = normalize(tf.matmul(inp, weights['w1']) + weights['b1'], activation=tf.nn.relu, reuse=reuse, scope='0')
        for i in range(1,len(self.dim_hidden)):
            hidden = normalize(tf.matmul(hidden, weights['w'+str(i+1)]) + weights['b'+str(i+1)], activation=tf.nn.relu, reuse=reuse, scope=str(i+1))
        return tf.matmul(hidden, weights['w'+str(len(self.dim_hidden)+1)]) + weights['b'+str(len(self.dim_hidden)+1)]

    # 构建卷积层的参数
    def construct_conv_weights(self):
        weights = {}

        dtype = tf.float32
        conv_initializer = tf.contrib.layers.xavier_initializer_conv2d(dtype=dtype)
        fc_initializer = tf.contrib.layers.xavier_initializer(dtype=dtype)
        k = 3

        weights['conv1'] = tf.get_variable('conv1', [k, k, self.channels, self.dim_hidden], initializer=conv_initializer, dtype=dtype)
        weights['b1'] = tf.Variable(tf.zeros([self.dim_hidden]))
        weights['conv2'] = tf.get_variable('conv2', [k, k, self.dim_hidden, self.dim_hidden], initializer=conv_initializer, dtype=dtype)
        weights['b2'] = tf.Variable(tf.zeros([self.dim_hidden]))
        weights['conv3'] = tf.get_variable('conv3', [k, k, self.dim_hidden, self.dim_hidden], initializer=conv_initializer, dtype=dtype)
        weights['b3'] = tf.Variable(tf.zeros([self.dim_hidden]))
        weights['conv4'] = tf.get_variable('conv4', [k, k, self.dim_hidden, self.dim_hidden], initializer=conv_initializer, dtype=dtype)
        weights['b4'] = tf.Variable(tf.zeros([self.dim_hidden]))
        if FLAGS.datasource == 'miniimagenet':
            # assumes max pooling
            weights['w5'] = tf.get_variable('w5', [self.dim_hidden*5*5, self.dim_output], initializer=fc_initializer)
            weights['b5'] = tf.Variable(tf.zeros([self.dim_output]), name='b5')
        else:
            weights['w5'] = tf.Variable(tf.random_normal([self.dim_hidden, self.dim_output]), name='w5')
            weights['b5'] = tf.Variable(tf.zeros([self.dim_output]), name='b5')
        return weights

    # 执行卷积层的前向计算
    def forward_conv(self, inp, weights, reuse=False, scope=''):
        # reuse is for the normalization parameters.
        channels = self.channels
        inp = tf.reshape(inp, [-1, self.img_size, self.img_size, channels])

        hidden1 = conv_block(inp, weights['conv1'], weights['b1'], reuse, scope+'0')
        hidden2 = conv_block(hidden1, weights['conv2'], weights['b2'], reuse, scope+'1')
        hidden3 = conv_block(hidden2, weights['conv3'], weights['b3'], reuse, scope+'2')
        hidden4 = conv_block(hidden3, weights['conv4'], weights['b4'], reuse, scope+'3')
        if FLAGS.datasource == 'miniimagenet':
            # last hidden layer is 6x6x64-ish, reshape to a vector
            hidden4 = tf.reshape(hidden4, [-1, np.prod([int(dim) for dim in hidden4.get_shape()[1:]])])
        else:
            hidden4 = tf.reduce_mean(hidden4, [1, 2])

        return tf.matmul(hidden4, weights['w5']) + weights['b5']