使用Tensorflow.js计算损失的梯度

问题描述 投票:2回答:1

我试图使用Tensorflow.js计算损失的梯度,与网络的可训练权重相关,以便将这些梯度应用于我的网络权重。在python中,可以使用tf.gradients()函数轻松完成,这需要两个表示dx和dy的最小输入。但是,我无法重现Tensorflow.js中的行为。我不确定我对损失梯度的理解是否有重量是错误的,或者我的代码是否包含错误。

我花了一些时间分析tfjs-node包的核心代码,以了解在调用函数tf.model.fit()时它是如何完成的,但到目前为止收效甚微。

let model = build_model(); //Two stacked dense layers followed by two parallel dense layers for the output
let loss = compute_loss(...); //This function returns a tf.Tensor of shape [1] containing the mean loss for the batch.
const f = () => loss;
const grad = tf.variableGrads(f);
grad(model.getWeights());

model.getWeights()函数返回一个tf.variable()数组,所以我假设函数会为每一层计算dL​​ / dW,我可以稍后将其应用于我的网络权重,但是,这并不是我的情况。得到这个错误:

Error: Cannot compute gradient of y=f(x) with respect to x. Make sure that the f you passed encloses all operations that lead from x to y.

我不太明白这个错误意味着什么。我应该如何使用Tensorflow.js计算丢失的梯度(类似于Python中的tf.gradients())?

编辑:这是计算损失的函数:

function compute_loss(done, new_state, memory, agent, gamma=0.99) {
    let reward_sum = 0.;
    if(done) {
        reward_sum = 0.;
    } else {
        reward_sum = agent.call(tf.oneHot(new_state, 12).reshape([1, 9, 12]))
                    .values.flatten().get(0);
    }

    let discounted_rewards = [];
    let memory_reward_rev = memory.rewards;
    for(let reward of memory_reward_rev.reverse()) {
        reward_sum = reward + gamma * reward_sum;
        discounted_rewards.push(reward_sum);
    }
    discounted_rewards.reverse();

    let onehot_states = [];
    for(let state of memory.states) {
        onehot_states.push(tf.oneHot(state, 12));
    }
    let init_onehot = onehot_states[0];

    for(let i=1; i<onehot_states.length;i++) {
        init_onehot = init_onehot.concat(onehot_states[i]);
    }

    let log_val = agent.call(
        init_onehot.reshape([memory.states.length, 9, 12])
    );

    let disc_reward_tensor = tf.tensor(discounted_rewards);
    let advantage = disc_reward_tensor.reshapeAs(log_val.values).sub(log_val.values);
    let value_loss = advantage.square();
    log_val.values.print();

    let policy = tf.softmax(log_val.logits);
    let logits_cpy = log_val.logits.clone();

    let entropy = policy.mul(logits_cpy.mul(tf.scalar(-1))); 
    entropy = entropy.sum();

    let memory_actions = [];
    for(let i=0; i< memory.actions.length; i++) {
        memory_actions.push(new Array(2000).fill(0));
        memory_actions[i][memory.actions[i]] = 1;
    }
    memory_actions = tf.tensor(memory_actions);
    let policy_loss = tf.losses.softmaxCrossEntropy(memory_actions.reshape([memory.actions.length, 2000]), log_val.logits);

    let value_loss_copy = value_loss.clone();
    let entropy_mul = (entropy.mul(tf.scalar(0.01))).mul(tf.scalar(-1));
    let total_loss_1 = value_loss_copy.mul(tf.scalar(0.5, dtype='float32'));

    let total_loss_2 = total_loss_1.add(policy_loss);
    let total_loss = total_loss_2.add(entropy_mul);
    total_loss.print();
    return total_loss.mean();

}

编辑2:

我设法使用compute_loss作为model.compile()中指定的损失函数。但是,它需要它只需要两个输入(预测,标签),所以它不适合我,因为我想输入多个参数。

我真的迷失了这个问题。

javascript node.js deep-learning tensorflow.js
1个回答
1
投票

错误说明了一切。你的问题与tf.variableGrads有关。 loss应该是使用所有可用的tf张量运算符计算的标量。 loss不应该返回你问题中指出的张量。

这是一个应该是什么损失的例子:

const a = tf.variable(tf.tensor1d([3, 4]));
const b = tf.variable(tf.tensor1d([5, 6]));
const x = tf.tensor1d([1, 2]);

const f = () => a.mul(x.square()).add(b.mul(x)).sum(); // f is a function
// df/da = x ^ 2, df/db = x 
const {value, grads} = tf.variableGrads(f); // gradient of f as respect of each variable

Object.keys(grads).forEach(varName => grads[varName].print());

/!\请注意,梯度是根据使用tf.variable创建的变量计算的

更新:

你不是应该计算渐变。这是修复。

function compute_loss(done, new_state, memory, agent, gamma=0.99) {
    const f = () => { let reward_sum = 0.;
    if(done) {
        reward_sum = 0.;
    } else {
        reward_sum = agent.call(tf.oneHot(new_state, 12).reshape([1, 9, 12]))
                    .values.flatten().get(0);
    }

    let discounted_rewards = [];
    let memory_reward_rev = memory.rewards;
    for(let reward of memory_reward_rev.reverse()) {
        reward_sum = reward + gamma * reward_sum;
        discounted_rewards.push(reward_sum);
    }
    discounted_rewards.reverse();

    let onehot_states = [];
    for(let state of memory.states) {
        onehot_states.push(tf.oneHot(state, 12));
    }
    let init_onehot = onehot_states[0];

    for(let i=1; i<onehot_states.length;i++) {
        init_onehot = init_onehot.concat(onehot_states[i]);
    }

    let log_val = agent.call(
        init_onehot.reshape([memory.states.length, 9, 12])
    );

    let disc_reward_tensor = tf.tensor(discounted_rewards);
    let advantage = disc_reward_tensor.reshapeAs(log_val.values).sub(log_val.values);
    let value_loss = advantage.square();
    log_val.values.print();

    let policy = tf.softmax(log_val.logits);
    let logits_cpy = log_val.logits.clone();

    let entropy = policy.mul(logits_cpy.mul(tf.scalar(-1))); 
    entropy = entropy.sum();

    let memory_actions = [];
    for(let i=0; i< memory.actions.length; i++) {
        memory_actions.push(new Array(2000).fill(0));
        memory_actions[i][memory.actions[i]] = 1;
    }
    memory_actions = tf.tensor(memory_actions);
    let policy_loss = tf.losses.softmaxCrossEntropy(memory_actions.reshape([memory.actions.length, 2000]), log_val.logits);

    let value_loss_copy = value_loss.clone();
    let entropy_mul = (entropy.mul(tf.scalar(0.01))).mul(tf.scalar(-1));
    let total_loss_1 = value_loss_copy.mul(tf.scalar(0.5, dtype='float32'));

    let total_loss_2 = total_loss_1.add(policy_loss);
    let total_loss = total_loss_2.add(entropy_mul);
    total_loss.print();
    return total_loss.mean().asScalar();
}

return tf.variableGrads(f);
}

请注意,您可以快速遇到内存消耗问题。最好围绕与tf.tidy区分的功能来处理张量。

© www.soinside.com 2019 - 2024. All rights reserved.