okraus / deeploc Goto Github PK

When I run : "python DeepLoc_visualize_classes.py"

I am getting the following error:
"2017-10-22 17:35:22.753508: W tensorflow/core/framework/op_kernel.cc:1192] Not found: Key final_layer_batch_norm/final_layer/batch_norm/final_layer_batch_norm/moments/Squeeze_1/ExponentialMovingAverage not found in checkpoint"

It is happening for each layer - can you please let me know how to fix it?

Thanks!

The link in the file download_datasets.sh is broken - here is the link:
http://spidey.ccbr.utoronto.ca/~okraus/DeepLoc_full_datasets.zip

I am trying to run DeepLoc_eval.py script on pretrained models such that --logdir parameter is set to pretrained_DeepLoc/pretrained_models where pretrained models are saved. I ran this script on 2 batches (implementation is not changed). I get total accuracy of 0.09375 and total loss 0.291712125391, what looks like model would not be trained. Is that normal or am I doing anything wrong?

Hi,
I am using a quasi-heavily modified of your network for classification of µCT image data. Could you please tell me how did you visualize the maximally activated convolutional filter for sample images in your paper (Fig2. C). I would really appreciate it if you share the code.
Cheers,
Pouyan

Hi,

I think there's an issue with the stretchVector function defined in preprocess_images.py:

def stretchVector(vec, lowerPercentile, upperPercentile):
    minVal = np.percentile(vec, lowerPercentile)
    maxVal = np.percentile(vec, upperPercentile)
    vec[vec > maxVal] = maxVal
    vec = vec - minVal
    if (maxVal-minVal)>1.:
        vec = vec / (maxVal - minVal)

    return vec

The function as it's currently implemented doesn't cap the color channel value to be >= 0.0. And, in cases where vec contains values that are less than minVal, we'll have negative numbers in there. I encountered this issue with some of the images in the dataset.

This should be easy to fix by adding the following line after vec = vec - minVal:

vec[vec < 0.0] = 0.0

Thanks!

Hi, I'm trying to use your transfer learning scripts, but it seems that the explicit model construction using your nn_layers module creates a slightly different graph than the one that is necessary to load the pre-trained model.

NotFoundError (see above for traceback): Key conv_1_batch_norm/conv_1/batch_norm/conv_1_batch_norm/moments/Squeeze/ExponentialMovingAverage not found in checkpoint

The closest namespace in the pre-trained model that seems to be an analog to the node called out in the error message above is: conv_1_batch_norm/conv_1/batch_norm/conv_1_batch_norm/moments/moments_1/mean/ExponentialMovingAverage (DT_FLOAT) [64]

I think I narrowed the issue down to the batch normalization step in nn_layers, but I'm not able to figure out why the new graph construction doesn't match the pre-trained model. Any help would be greatly appreciated.

Thanks!