Metrics PR Error

[bwhwang]

Dear authors,

Thank you for your wonderful contribution!!!

When I tried to get precision and recall values during training by adding option, --metric pr, I got the following error

---

\precision_recall.py", line 179, in _evaluate feats = self._gen_feats(Gs, inception, minibatch_size, num_gpus, Gs_kwargs) NameError: name 'inception' is not defined

So, I have changed the lines in precision_recall.py. After the modification, It seems to work. I would greatly appreciate it if you kindly review my modification.

---

def _evaluate(self, Gs, Gs_kwargs, num_gpus, num_imgs, paths = None, **kwargs):

        if paths is not None: 
            # Extract features for local sample image files (paths)
----->  eval_features = self._paths_to_feats(paths, feat_func, minibatch_size, num_gpus, num_imgs)
        else:
            # Extract features for newly generated fake imgs
----->  eval_features = self._gen_feats(Gs, feature_net, minibatch_size, num_imgs, num_gpus, Gs_kwargs)

        # Compute precision and recall
        state = knn_precision_recall_features(ref_features = ref_features, eval_features = eval_features,
            feature_net = feature_net, nhood_sizes = [self.nhood_size], row_batch_size = self.row_batch_size,
----->  col_batch_size = self.row_batch_size, num_gpus = num_gpus, num_imgs = num_imgs)
        self._report_result(state.knn_precision[0], suffix = "_precision")
        self._report_result(state.knn_recall[0], suffix = "_recall")

-------------------------------------------------------------------------

[dorarad]

Hi, thanks so much for the kind words! Sorry about it that definitely looks like a bug, I will resolve today and update you!

[bwhwang]

Thank you for your prompt reply. ^^ I have a small question about the training options. For the duplex transformer mode(GANsformerd) in Table 1 , Do we add a single option, --g-img2ltnt to your gansformer_default options?

Thanks in advance!

[dorarad]

Sure! :) For the duplex originally yes but after I refactored all the code I had some training issue with the flag that I'm still exploring so for now I released only (an improved version of) the simplex as the default and plan to follow up on that

[dorarad]

The differences between original simplex fro paper and improved simplex here:

• Using --mapping_ltnt2ltnt so that the latents communicate with each other directly through self-attention inside the mapping network before starting to generate the image.
• Adding the --kmeans from the duplex mode.
• Adding an additional global latent (--style) to the k standard latent components, such that first the one global latent modulates all the image features uniformly, and then the k latents modulate different regions based on the attention.
  The global latent is useful for coordinating global aspects of the image such as global lighting conditions, global style properties of e.g. a face, etc.
• Turns outs that with these changes no need to add the transformer to the discriminator, so for simplicity disabled that.

[bwhwang]

Wow! Thanks for the detail explanation about the improved simplex. We would appreciate to your step-by-step revision of the code. Look forward to releasing the complete code fully supporting the duplex.:)

[dorarad]

Sure happy to help! :) I'll add these explanations also to the readme. I now also looked into your changes they're all almost correct (there was one more change needed in case paths is not None but it doesn't happen by default so won't affect the results you get). I made these changes to the repo: https://github.com/dorarad/gansformer/commit/175157caa8ec3547dc22a6ee6c10bc85f29e1958 Let me know if you have any further questions!