File size: 5,230 Bytes
e9f9fd3 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 |
from ..torch_core import *
from ..basic_data import *
from ..basic_train import *
from .image import *
from ..train import Interpretation
from textwrap import wrap
__all__ = ['SegmentationInterpretation', 'ObjectDetectionInterpretation']
class SegmentationInterpretation(Interpretation):
"Interpretation methods for segmenatation models."
def __init__(self, learn:Learner, preds:Tensor, y_true:Tensor, losses:Tensor,
ds_type:DatasetType=DatasetType.Valid):
super(SegmentationInterpretation, self).__init__(learn,preds,y_true,losses,ds_type)
self.pred_class = self.preds.argmax(dim=1)
self.c2i = {c:i for i,c in enumerate(self.data.classes)}
self.i2c = {i:c for c,i in self.c2i.items()}
def top_losses(self, sizes:Tuple, k:int=None, largest=True):
"Reduce flatten loss to give a single loss value for each image"
losses = self.losses.view(-1, np.prod(sizes)).mean(-1)
return losses.topk(ifnone(k, len(losses)), largest=largest)
def _interp_show(self, ims:ImageSegment, classes:Collection=None, sz:int=20, cmap='tab20',
title_suffix:str=None):
"Show ImageSegment with color mapping labels"
fig,axes=plt.subplots(1,2,figsize=(sz,sz))
np_im = to_np(ims.data).copy()
# tab20 - qualitative colormaps support max of 20 distinc colors
# if len(classes) > 20 close idxs map to same color
# image
if classes is not None:
class_idxs = [self.c2i[c] for c in classes]
mask = np.max(np.stack([np_im==i for i in class_idxs]),axis=0)
np_im = (np_im*mask).astype(np.float)
np_im[np.where(mask==0)] = np.nan
im=axes[0].imshow(np_im[0], cmap=cmap)
# labels
np_im_labels = list(np.unique(np_im[~np.isnan(np_im)]))
c = len(np_im_labels); n = math.ceil(np.sqrt(c))
label_im = np.array(np_im_labels + [np.nan]*(n**2-c)).reshape(n,n)
axes[1].imshow(label_im, cmap=cmap)
for i,l in enumerate([self.i2c[l] for l in np_im_labels]):
div,mod=divmod(i,n)
l = "\n".join(wrap(l,10)) if len(l) > 10 else l
axes[1].text(mod, div, f"{l}", ha='center', color='white', fontdict={'size':sz})
if title_suffix:
axes[0].set_title(f"{title_suffix}_imsegment")
axes[1].set_title(f"{title_suffix}_labels")
def show_xyz(self, i, classes:list=None, sz=10):
'show (image, true and pred) from self.ds with color mappings, optionally only plot'
x,y = self.ds[i]
self.ds.show_xys([x],[y], figsize=(sz/2,sz/2))
self._interp_show(ImageSegment(self.y_true[i]), classes, sz=sz, title_suffix='true')
self._interp_show(ImageSegment(self.pred_class[i][None,:]), classes, sz=sz, title_suffix='pred')
def _generate_confusion(self):
"Average and Per Image Confusion: intersection of pixels given a true label, true label sums to 1"
single_img_confusion = []
mean_confusion = []
n = self.pred_class.shape[0]
for c_j in range(self.data.c):
true_binary = self.y_true.squeeze(1) == c_j
total_true = true_binary.view(n,-1).sum(dim=1).float()
for c_i in range(self.data.c):
pred_binary = self.pred_class == c_i
total_intersect = (true_binary*pred_binary).view(n,-1).sum(dim=1).float()
p_given_t = (total_intersect / (total_true))
p_given_t_mean = p_given_t[~torch.isnan(p_given_t)].mean()
single_img_confusion.append(p_given_t)
mean_confusion.append(p_given_t_mean)
self.single_img_cm = to_np(torch.stack(single_img_confusion).permute(1,0).view(-1, self.data.c, self.data.c))
self.mean_cm = to_np(torch.tensor(mean_confusion).view(self.data.c, self.data.c))
return self.mean_cm, self.single_img_cm
def _plot_intersect_cm(self, cm, title="Intersection with Predict given True"):
"Plot confusion matrices: self.mean_cm or self.single_img_cm generated by `_generate_confusion`"
from IPython.display import display, HTML
fig,ax=plt.subplots(1,1,figsize=(10,10))
im=ax.imshow(cm, cmap="Blues")
ax.set_xlabel("Predicted")
ax.set_ylabel("True")
ax.set_title(f"{title}")
ax.set_xticks(range(self.data.c))
ax.set_yticks(range(self.data.c))
ax.set_xticklabels(self.data.classes, rotation='vertical')
ax.set_yticklabels(self.data.classes)
fig.colorbar(im)
df = (pd.DataFrame([self.data.classes, cm.diagonal()], index=['label', 'score'])
.T.sort_values('score', ascending=False))
with pd.option_context('display.max_colwidth', -1):
display(HTML(df.to_html(index=False)))
return df
class ObjectDetectionInterpretation(Interpretation):
"Interpretation methods for classification models."
def __init__(self, learn:Learner, preds:Tensor, y_true:Tensor, losses:Tensor, ds_type:DatasetType=DatasetType.Valid):
raise NotImplementedError
super(ObjectDetectionInterpretation, self).__init__(learn,preds,y_true,losses,ds_type)
|