Plot_top_losses : Intuition behing heatmap_thresh parameter?

Kalanit · March 2, 2019, 12:12pm

Hi,

plot_top_losses now plots by default the heatmat (great feature by the way - many thanks ;))

I read the definition of the function here:

fastai/fastai/blob/master/fastai/vision/learner.py#L105


    meta = cnn_config(base_arch)
    model = create_cnn_model(base_arch, data.c, cut, pretrained, lin_ftrs, ps=ps, custom_head=custom_head,
        split_on=split_on, bn_final=bn_final, concat_pool=concat_pool)
    learn = Learner(data, model, **kwargs)
    learn.split(split_on or meta['split'])
    if pretrained: learn.freeze()
    if init: apply_init(model[1], nn.init.kaiming_normal_)
    return learn


def create_cnn(data, base_arch, **kwargs):
    warn("`create_cnn` is deprecated and is now named `cnn_learner`.")
    return cnn_learner(data, base_arch, **kwargs)


def unet_learner(data:DataBunch, arch:Callable, pretrained:bool=True, blur_final:bool=True,
                 norm_type:Optional[NormType]=NormType, split_on:Optional[SplitFuncOrIdxList]=None, blur:bool=False,
                 self_attention:bool=False, y_range:Optional[Tuple[float,float]]=None, last_cross:bool=True,
                 bottle:bool=False, cut:Union[int,Callable]=None, **learn_kwargs:Any)->Learner:
    "Build Unet learner from `data` and `arch`."
    meta = cnn_config(arch)
    body = create_body(arch, pretrained, cut)
    model = to_device(models.unet.DynamicUnet(body, n_classes=data.c, blur=blur, blur_final=blur_final,

Did not really help to get a feeling of why this threshold is initialized to 16, what are the possible value to it (range from 0 to 100 - percentage of max intensity they mention here maybe ?).

I also not understand how this threshold is influencing the plotting.
I would expect that if I increase (or decrease) I would get less area (or more) colored on my heatmaps.

What I observe is same results if I put the threshold to 5 or 45 - and nothing colored at all if I put it to 50.

Can anyone help me get an intuition of what this 16 defaut value means, and what it changes to change that value?

Thank you
K.

Example on a dummy X-ray dataset (epaule=shoulder, hanches=hips)

sgugger · March 3, 2019, 3:03pm

No, the threshold doesn’t have anything to do with values of activations, it’s a threshold of how many activations are necessary to plot the heatmap (if you just have a 1x1 result from your CNN for instance, it won’t give a nice result).

Kalanit · March 3, 2019, 3:23pm

Thank you for your answer. It is still unclear to me why I have exactly same heatmap per pic with 5 and 45 as threshold, and no heatmap for 50. Does it make sense to you ?

(This particular example is a training of resnet34 for binary classification)

sgugger · March 3, 2019, 3:28pm

That means you have less than 50 features (probably 7 by 7 = 49).

Kalanit · March 3, 2019, 3:34pm

Ah ok! Indeed. Got it.
thanks a lot @sgugger