Computing the gradient of the loss using Tensorflow.js

Question

I am trying to compute the gradient of a loss, with relation to a network\'s trainable weights using Tensorflow.js in order to apply these gradients to my network\'s weight.

Answer 1

The error says it all. Your issue has to do with tf.variableGrads. loss should be a scalar computed using all available tf tensors operators. loss should not return a tensor as indicated in your question.

Here is an example of what loss should be:

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());

/!\ Notice that the gradient is calculated as respect of variables created using tf.variable

Update:

You're not computing the gradients as it should be. Here is the fix.

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);
}

Notice that you can quickly run into a memory consumption issue. It will advisable to surround the function differentiated with tf.tidy to dispose of the tensors.