Image label work

codes/data/image_label_parser.py

@@ -67,5 +67,63 @@ class VsNetImageLabeler:
         self.categories[binding]['labeledImages'].append(lbl)
 
     def save(self):
-        with open(self.label_file[0], "wb") as file:
+        with open(self.label_file[-1], "wb") as file:
+            file.write(json.dumps(self.categories))
+
+
+# A more compact format that is simpler to parse and understand.
+class CompactJsonLabeler:
+    def __init__(self, lbl_files):
+        if not isinstance(lbl_files, list):
+            lbl_files = [lbl_files]
+        self.label_files = lbl_files
+        self.config, self.labels, self.label_map, self.images = None, None, None, None
+        for lfil in lbl_files:
+            with open(lfil, "r") as read_file:
+                # Format:
+                # {
+                #   'config': { 'dim' }
+                #   'labels': [{ 'label', 'key'}] <- ordered by label index.
+                #   'images': {'file': [{ 'lid', 'top', 'left' }}
+                # 'labelMap' {<mapping of string labels to ids>}
+                # }
+                parsed = json.loads(read_file.read())
+                if self.config is None:
+                    self.config = parsed['config']
+                    self.labels = parsed['labels']
+                    self.images = parsed['images']
+                    self.label_map = parsed['label_map']
+                    self.binding_map = {}
+                    for i, lbl in enumerate(self.labels):
+                        self.binding_map[lbl['key']] = i
+                else:
+                    assert self.config == parsed['config']
+                    assert self.labels == parsed['labels']
+                    assert self.label_map == parsed['label_map']
+                    self.images.update(parsed['images'])  # This will overwrite existing images, which is acceptable.
+
+    def get_labeled_paths(self, base_path):
+        return [os.path.join(base_path, pth) for pth in self.images.keys()]
+
+    def get_labels_as_tensor(self, hq, img_key, resize_factor):
+        _, h, w = hq.shape
+        labels = torch.zeros((1,h,w), dtype=torch.long)
+        mask = torch.zeros((1,h,w), dtype=torch.float)
+        lbl_list = self.images[img_key]
+        for patch_lbl in lbl_list:
+            t, l, h, w = patch_lbl['top'] // resize_factor, patch_lbl['left'] // resize_factor, \
+                         self.config['dim'] // resize_factor, self.config['dim'] // resize_factor
+            val = patch_lbl['labelValue']
+            labels[:,t:t+h,l:l+w] = val
+            mask[:,t:t+h,l:l+w] = 1.0
+        return labels, mask, self.str_labels
+
+    def add_label(self, binding, img_name, top, left, dim):
+        lbl = {'lid': self.binding_map[binding], 'top': top, 'left': left}
+        if img_name not in self.images.keys():
+            self.images[img_name] = []
+        self.images[img_name].append(lbl)
+
+    def save(self):
+        with open(self.label_file[-1], "wb") as file:
             file.write(json.dumps(self.categories))

codes/scripts/ui/image_labeler/image_labeler_ui.py

@@ -37,7 +37,7 @@ def update_mode_label():
 # Handles the "change mode" hotkey. Changes the classification label being targeted.
 def change_mode(event):
     global mode, pending_labels
-    mode += 1
+    mode = (mode + 1) % len(labeler.str_labels)
     update_mode_label()
 
 
@@ -88,7 +88,7 @@ def next_batch():
     scale = hq.shape[-1] // res.shape[-1]
 
     # These are the confidence bounds. They apply to a post-softmax output. They are currently fixed.
-    conf_lower = .8
+    conf_lower = .4
     conf_upper = 1
     valid_res = ((res > conf_lower) * (res < conf_upper)) * 1.0  # This results in a tensor of 0's where tensors are outside of the confidence bound, 1's where inside.
 
@@ -96,8 +96,10 @@ def next_batch():
     batch_sz = hq.shape[0]
     while cur_img < batch_sz:  # Note: cur_img can (intentionally) be changed outside of this loop.
         # Build a random permutation for every image patch in the image. We will search for patches that fall within the confidence bound and yield them.
-        #permutation = torch.randperm(res.shape[-1] * res.shape[-2])
-        for p in range(res.shape[-1]*res.shape[-2]):
+        permutation = torch.randperm(res.shape[-1] * res.shape[-2])
+        for p in permutation:
+            p = p.item()
+        #for p in range(res.shape[-1]*res.shape[-2]):
             # Reconstruct a top & left coordinate.
             t = p // res.shape[-1]
             l = p % res.shape[-1]

codes/scripts/ui/image_labeler/label_editor.py

@@ -0,0 +1,33 @@
+import orjson
+
+from data.image_label_parser import VsNetImageLabeler
+
+
+# Translates from the label JSON output of the VS.NET UI to something more compact and usable.
+def convert_from_vsnet_labels():
+    labeler = VsNetImageLabeler(['F:\\4k6k\datasets\\ns_images\\512_unsupervised\\categories.json',
+                                 'F:\\4k6k\datasets\\ns_images\\512_unsupervised\\categories_new.json',
+                                 'F:\\4k6k\datasets\\ns_images\\512_unsupervised\\categories_new_new.json'])
+    # Proposed format:
+    # 'config': { 'dim' }
+    # 'labels': [{ 'label', 'key'}] <- ordered by label index.
+    # 'images': {'file': [{ 'lid', 'top', 'left' }}
+    # 'labelMap' {<mapping of string labels to ids>}
+    out_dict = {
+        'config': {
+            'dim': next(iter(labeler.labeled_images.values()))[0]['patch_width']
+        },
+        'labels': [{'label': cat['label'], 'key': cat['keyBinding']} for cat in labeler.categories.values()],
+    }
+    out_dict['labelMap'] = {}
+    for i, lbl in enumerate(out_dict['labels']):
+        out_dict['labelMap'][lbl['label']] = i
+    out_dict['images'] = {}
+    for fname, ilbls in labeler.labeled_images.items():
+        out_dict['images'][fname] = [{'lid': out_dict['labelMap'][il['label']], 'top': il['patch_top'], 'left': il['patch_left']} for il in ilbls]
+    with open("label_editor.json", 'wb') as fout:
+        fout.write(orjson.dumps(out_dict))
+
+
+if __name__ == '__main__':
+    convert_from_vsnet_labels()

codes/train.py

@@ -293,7 +293,7 @@ class Trainer:
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('-opt', type=str, help='Path to option YAML file.', default='../options/train_imgset_structural_classifier.yml')
+    parser.add_argument('-opt', type=str, help='Path to option YAML file.', default='../options/train_exd_mi1_rrdb_bigboi_pretrain.yml')
     parser.add_argument('--launcher', choices=['none', 'pytorch'], default='none', help='job launcher')
     parser.add_argument('--local_rank', type=int, default=0)
     args = parser.parse_args()