Edit

How do I evaluate models in Python

  • Article

Evaluate a saved convolutional network

There are a few things to consider with models trained on images. At this point the transformations are not part of the model, so subtracting the mean has to be done manually. Another issue is that PIL loads images in a different order than what was used during training and a transposition is required.

Assuming that:

  • during training you subtracted 128 from all channels
  • the image you want to predict on is "foo.jpg"
  • you saved your model in Python using z.save("mycnn.dnn")

then you can do the following:

from cntk.ops.functions import load_model
from PIL import Image 
import numpy as np

z = load_model("mycnn.dnn")
rgb_image = np.asarray(Image.open("foo.jpg"), dtype=np.float32) - 128
bgr_image = rgb_image[..., [2, 1, 0]]
pic = np.ascontiguousarray(np.rollaxis(bgr_image, 2))

predictions = np.squeeze(z.eval({z.arguments[0]:[pic]}))
top_class = np.argmax(predictions)

If you are loading an old model trained by NDL or BrainScript, then you will need to find the model output node as follows:

for index in range(len(z.outputs)):
   print("Index {} for output: {}.".format(index, z.outputs[index].name))

...
Index 0 for output: CE_output.
Index 1 for output: Err_output.
Index 2 for output: OutputNodes.z_output.
...

We care only about 'z_output' which has index 2. So in order to get the real model output, do the following

import cntk

z_out = cntk.combine([z.outputs[2].owner])
predictions = np.squeeze(z_out.eval({z_out.arguments[0]:[pic]}))
top_class = np.argmax(predictions)

The reason for the above, is that in old model the training information is saved in addition to the actual model parameters.

Extract features from a specific layer using a trained model?

There is an example here.