Tensorflow.js 模型维度不匹配答案

【问题标题】：Tensorflow.js Model Dimension MismatchTensorflow.js 模型维度不匹配
【发布时间】：2021-03-30 21:47:57
【问题描述】：

我是 tensoflow.js 的新手，我在创建 CNN 模型时遇到问题，因为尺寸不匹配。我有一个使用 tf.browser.fromPixels(image); 的 3d 数组。

但是当我尝试训练我的人工智能时它不会启动，我收到了消息：Uncaught (in promise) Error: Error when checking target: expected dense_Dense2 to have shape [6,6,4], but got array with shape [6,6,3].

这里是完整的代码：

image = new Image(32, 32);
data = tf.browser.fromPixels(image); //to get pixel array

model = tf.sequential();
encoder = tf.layers.dense({units: 4, batchInputShape:[6, 6, 3], activation: 'relu', kernelInitializer:"randomNormal", biasInitializer:"ones"});
decoder = tf.layers.dense({units: 4, activation: 'relu'});
model.add(encoder);
model.add(decoder);
model.compile({loss:'meanSquaredError', optimizer:tf.train.adam()});

async function botTraining(model, epochs = 60){ //train ai 60 epochs
    history = 
    await model.fit(data, data,{ epochs: epochs + 1,
        callbacks:{
            onEpochEnd: async(epoch, logs) =>{
                console.log("Epoch:" + epoch + " Loss:" + logs.loss);
            }
        }
    });
}

【问题讨论】：

标签： javascript tensorflow conv-neural-network png autoencoder

【解决方案1】：

好的，这是我的代码：

image = new Image(32, 32);
data = tf.browser.fromPixels(image); //to get pixel array

model = tf.sequential();
encoder = tf.layers.dense({units: 4, batchInputShape:[6, 6, 3], activation: 'relu', kernelInitializer:"randomNormal", biasInitializer:"ones"});
decoder = tf.layers.dense({units: 4, activation: 'relu'});
model.add(encoder);
model.add(decoder);
model.compile({loss:'meanSquaredError', optimizer:tf.train.adam()});

async function botTraining(model, epochs = 60){ //train ai 60 epochs
    history = 
    await model.fit(data, data,{ epochs: epochs + 1,
        callbacks:{
            onEpochEnd: async(epoch, logs) =>{
                console.log("Epoch:" + epoch + " Loss:" + logs.loss);
            }
        }
    });
}

我发现我的“bachInputSize”需要设置为与decoder 单位相同

这是固定代码：

image = new Image(4, 4);
data = tf.browser.fromPixels(image);

model = tf.sequential();
encoder = tf.layers.dense({units: 3, batchInputShape:[null, null, 3], activation: 'relu', kernelInitializer:"randomNormal", biasInitializer:"ones"});
decoder = tf.layers.dense({units: 3, activation: 'relu'});
model.add(encoder);
model.add(decoder);
model.compile({loss:'meanSquaredError', optimizer:tf.train.adam()});

async function botTraining(model, epochs = 60){
    history = 
    await model.fit(data, data,{ epochs: epochs + 1,
        callbacks:{
            onEpochEnd: async(epoch, logs) =>{
                console.log("Epoch:" + epoch + " Loss:" + logs.loss);
            }
        }
    });
}

【讨论】：