Convert opencv's `mat` to tfgo's `tensor` about tfgo HOT 7 CLOSED

Hi, I'm not familiar with gocv, but I guess you just need a way to extract the raw byte representation from the gocv.Mat type, and create the new tensor. Be careful, that you want to extract not the RGB representation, but a compressed representation (e.g. the PNG image, thus the png headers + the png content).

Anyway, I do recommend to user the TensorFlow primitives instead of mixing gocv with tfgo.

from tfgo.

Thanks. gocv is basically the OpenCV in Go. I have the following preprocessing steps using gocv in Go and cv2 in python. Doing these, the r_imgs in both cases are exactly the same. Are these steps reproducible using tfgo? For example, I couldn't find how to save the image to file.

img := gocv.IMRead(filename, gocv.IMReadColor)
defer img.Close()

c_img := gocv.NewMat()
defer c_img.Close()
gocv.CvtColor(img, &c_img, gocv.ColorBGRToRGB)

r_img := gocv.NewMat()
defer r_img.Close()
gocv.Resize(c_img, &r_img, image.Point{112, 112}, 0, 0, gocv.InterpolationLinear)

gocv.IMWrite("./resized.png", r_img)

and the python equivalent:

img = cv2.imread(filename)
c_img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
r_img = cv2.resize(c_img, (112, 112))
cv2.imwrite(f"./resized.png", r_img)

from tfgo.

Yes, have a look at the example code in the README.
Using the "github.com/galeone/tfgo/image" package, you can read the image, resize it, get a tensor out of it, and save it.

https://github.com/galeone/tfgo/blob/master/image/image.go

from tfgo.

Great. Thanks for the help.

from tfgo.

Using the Value() from the galeone/tfgo/image package, I'm getting the following error:
Cannot use 'r_img.Value()' (type tf.Output) as the type *tf.Tensor

This is how I feed the image to the model:

root := tg.NewRoot()
img := image.Read(root, filename, 3)
r_img := img.CropAndResize(image.Box{Start: image.Point{X: 0.0, Y: 0.0},
	End: image.Point{X: 1.0, Y: 1.0}},
	image.Size{Height: 112, Width: 112})

model := tg.LoadModel("./model_pb", []string{"serve"}, nil)

qmf_results := model.Exec([]tf.Output{
	model.Op("PartitionedCall", 0),
		}, map[tf.Output]*tf.Tensor{
	qmf_model.Op("serving_default_input.1", 0): r_img.Value(),
})

Any idea on how to resolve this?

from tfgo.

Hey @galeone :)
Any idea about the above error?

from tfgo.

The Cannot use 'r_img.Value()' (type tf.Output) as the type *tf.Tensor has already been discussed in #27, so closing this one.

from tfgo.