TAO does not provide an API to dump the numpy array for the feature map tensors. So currently it is not feasible to visualize layer’s feature map with NVIDIA FME.
Not surprisingly, in order to use Feature Map Explorer, you will need to generate some feature maps from your trained model. This is a straightforward task with most training frameworks, since FME directly consumes feature map tensors written out as standard numpy array files (i.e., the .npy files written by the np.save() method). These files are assumed to contain 32-bit single precision floating point values in a 4D array with order NCHW (where N is the batch size, C is the channel count, and H and W are height and width respectively). If your training data is in NHWC (channels last) layout, you will need to transpose the axes (e.g., permute the axes with a ( 0 , 3 , 1 , 2 ) permutation) before calling np.save() .