Hi, thanks for the great code!

I wonder do you have plans to support resuming from checkpoints for classification? As we all know, in terms of training ImageNet, the training process is really long and it can be interrupted somehow, but I haven't notice any code related to "resume" in scripts/classification/train_imagenet.py.

Maybe @hetong007 ? Thanks in advance.

Thank you for the great effort, you are putting into this project :) There is, however, a feature I miss; rotated bounding boxes. Especially when objects are thin and diagonal, an ordinary bounding box fits poorly. Examples of such cases are shown here: rotated bounding boxes

A way annotation could be

import numpy as np
from sklearn.metrics import confusion_matrix

class runningScore(object):
    def __init__(self, n_classes):
        self.n_classes = n_classes
        self.confusion_matrix = np.zeros((n_classes, n_classes))

    def _fast_hist(self, label_true, label_pred, n_class):
        mask = (label_true >= 0) & (label_true < n_class)
        hist = np.bincount(
          

In experiments.rst, in the first diagram, I'm inferring that ovals are static resources and parallelograms are processes, and that dashed lines means optional. But if that's true, then I'm not sure what the meaning of purple vs. blue is, the scenes should be ovals and solid lines. It might be simpler and easier to parse if everything was a blue oval, although the dashed line makes sense. In the

I trained the model with 4 GPUs (GTX 2080) with batchsize 4, and the PSPNet achieves the mIOU of 76.42%. But I found the visualization results are not correct. Here is an example.
.
But the ground-truth label is like:
![image](https://user-images.githubusercontent.com/35193979/64757149-