I think the issue comes from the evaluation metric. The following code include
a modified rmse metric that flatten the pred and label vector. I think it's bug
that the eval metric fails when the predictions are not in a flat setting, I'll
open a PR to get it fixed.
```
data.A <- mx.nd.random.normal(shape = c(3,3,1,10))
data.A.2 <- mx.nd.random.normal(shape = c(3,3,1,10))
batch_size <- 5
train_iter = mx.io.arrayiter(data = as.array(data.A),
label = as.array(data.A.2),
batch.size = batch_size)
data <- mx.symbol.Variable('data')
label <- mx.symbol.Variable('label')
conv_1 <- mx.symbol.Convolution(data= data, kernel = c(1,1), num_filter = 4,
name="conv_1")
conv_act_1 <- mx.symbol.Activation(data= conv_1, act_type = "relu",
name="conv_act_1")
flat <- mx.symbol.flatten(data = conv_act_1, name="flatten")
fcl_1 <- mx.symbol.FullyConnected(data = flat, num_hidden = 9, name="fc_1")
fcl_2 <- mx.symbol.reshape(fcl_1, shape=c(3, 3, 1, batch_size))
NN_Model <- mx.symbol.LinearRegressionOutput(data=fcl_2 , label=label,
name="lro")
fcl_2$infer.shape(list(data = c(3,3,1,batch_size)))
NN_Model$infer.shape(list(data = c(3,3,1,batch_size)))
mx.metric.rmse <- mx.metric.custom("rmse", function(label, pred) {
pred <- mx.nd.reshape(pred, shape = -1)
label <- mx.nd.reshape(label, shape = -1)
res <- mx.nd.sqrt(mx.nd.mean(mx.nd.square(label-pred)))
return(as.array(res))
})
mx.set.seed(99)
autoencoder <- mx.model.FeedForward.create(
NN_Model,
X = train_iter,
initializer = mx.init.uniform(0.01),
ctx=mx.cpu(),
num.round=5,
eval.metric = mx.metric.rmse,
optimizer = mx.opt.create("sgd"),
verbose = TRUE)
```
[ Full content available at:
https://github.com/apache/incubator-mxnet/issues/12431 ]
This message was relayed via gitbox.apache.org for [email protected]
