Golang Tensorflow批处理图像输入

Question

Problem Statement: Could not able to make batch image processing in GO Tensorflow.

I have been going through following URL on GoLang Tensorflow. https://github.com/tensorflow/tensorflow/blob/master/tensorflow/go/example_inception_inference_test.go

I am Facing problem while making batch of images for input to model. Check this line https://github.com/tensorflow/tensorflow/blob/master/tensorflow/go/example_inception_inference_test.go#L199

Any help will be appreciated!

result, err := classifier.Session.Run(
        map[tf.Output]*tf.Tensor{
            inputTensor.Output(0): imageTensor,
        },
        []tf.Output{
            outputTensorOne.Output(0),
            outputTensorTwo.Output(0),
        },
        nil, /*targets*/
    )

// How to make that imageTensor a batch of images in GO Tensorflow.

Answer 1

The Go API isn't as comprehensive and to make things easier there are a few additions that could be made. However, given the current API, it is possible to construct the batched tensor using something like this:

var buf bytes.Buffer
for i, img := range images {
  bytes, err := gocv.IMEncode(gocv.JPEGFileExt, img)
  if err != nil {
    fmt.Println("Error")
  }
  tensor, err = tf.NewTensor(string(bytes))
  if err != nil {
    fmt.Println("Error")
  }
  normalized, err := session.Run(
    map[tf.Output]*Tensor: { input: tensor },
    []tf.Output{output},
    nil)
  if _, err := normalized[0].WriteContentsTo(&buf); err != nil {
    // Handle error
  }
}

batchShape := []int64{len(images), 224, 224, 3}
batch, err := tf.ReadTensor(tf.Float, batchShape, &buf)
if err != nil {
  // Handle error
}

// Now feed "batch" to the model

Another alternative would be to do this batching in the graph by constructing a graph that packs multiple single-image tensors together in a batch using the Pack operation).

Hope that helps.

(P.S. Seems like you also asked this in a GitHub issue and the same answer is posited there: https://github.com/tensorflow/tensorflow/issues/25440)