Initial crowdai commit

.gitignore

@@ -1,3 +1,4 @@
+scratch
 saved_models
 wandb
 checkpoints

aicrowd.json

@@ -0,0 +1,8 @@
+{
+  "challenge_id": "aicrowd-neurips-2019-disentanglement-challenge",
+  "grader_id": "aicrowd-neurips-2019-disentanglement-challenge",
+  "authors": ["amirabdi"],
+  "description": "Pytorch implementation of disentanglement algorithms",
+  "license": "MIT",
+  "gpu": true
+}

aicrowd_helpers.py

@@ -0,0 +1,66 @@
+#!/usr/bin/env python
+import crowdai_api
+import os
+
+########################################################################
+# Instatiate Event Notifier
+########################################################################
+crowdai_events = crowdai_api.events.CrowdAIEvents()
+
+
+def execution_start():
+    ########################################################################
+    # Register Evaluation Start event
+    ########################################################################
+    print("Training Start...")
+    crowdai_events.register_event(
+                event_type=crowdai_events.CROWDAI_EVENT_INFO,
+                message="execution_started",
+                payload={ #Arbitrary Payload
+                    "event_type": "disentanglement_challenge:execution_started"
+                    }
+                )
+
+
+def register_progress(progress):
+    ########################################################################
+    # Register Evaluation Progress event
+    # progress : float [0, 1]
+    ########################################################################
+    print("Training Progress : {}".format(progress))
+    crowdai_events.register_event(
+                event_type=crowdai_events.CROWDAI_EVENT_INFO,
+                message="register_progress",
+                payload={ #Arbitrary Payload
+                    "event_type": "disentanglement_challenge:register_progress",
+                    "training_progress" : progress
+                    }
+                )
+
+def submit(payload={}):
+    ########################################################################
+    # Register Evaluation Complete event
+    ########################################################################
+    print("AIcrowd Submit")
+    crowdai_events.register_event(
+                event_type=crowdai_events.CROWDAI_EVENT_SUCCESS,
+                message="submit",
+                payload={ #Arbitrary Payload
+                    "event_type": "disentanglement_challenge:submit",
+                    },
+                blocking=True
+                )
+
+def execution_error(error):
+    ########################################################################
+    # Register Evaluation Complete event
+    ########################################################################
+    crowdai_events.register_event(
+                event_type=crowdai_events.CROWDAI_EVENT_ERROR,
+                message="execution_error",
+                payload={ #Arbitrary Payload
+                    "event_type": "disentanglement_challenge:execution_error",
+                    "error" : error
+                    },
+                blocking=True
+                )

apt.txt

@@ -0,0 +1,5 @@
+curl
+gcc
+git
+vim
+ssh

common/dataset.py

@@ -57,6 +57,7 @@ def __init__(self, root, transform, labels, label_weights, name, class_values, n
 
         self.label_handler = LabelHandler(labels, label_weights, class_values)
 
+    @property
     def name(self):
         return self._name
 
@@ -76,22 +77,15 @@ def num_channels(self):
         return self._num_channels
 
     def __getitem__(self, index1):
-        index2 = random.choice(self.indices)
-
         path1 = self.imgs[index1][0]
-        path2 = self.imgs[index2][0]
         img1 = self.loader(path1)
-        img2 = self.loader(path2)
         if self.transform is not None:
             img1 = self.transform(img1)
-            img2 = self.transform(img2)
 
         label1 = 0
-        label2 = 0
         if self.label_handler.has_labels():
             label1 = self.label_handler.get_label(index1)
-            label2 = self.label_handler.get_label(index2)
-        return img1, img2, label1, label2
+        return img1, label1
 
 
 class CustomNpzDataset(Dataset):
@@ -106,6 +100,7 @@ def __init__(self, data_images, transform, labels, label_weights, name, class_va
         self.transform = transform
         self.indices = range(len(self))
 
+    @property
     def name(self):
         return self._name
 
@@ -125,20 +120,14 @@ def num_channels(self):
         return self._num_channels
 
     def __getitem__(self, index1):
-        index2 = random.choice(self.indices)
-
         img1 = Image.fromarray(self.data_npz[index1] * 255)
-        img2 = Image.fromarray(self.data_npz[index2] * 255)
         if self.transform is not None:
             img1 = self.transform(img1)
-            img2 = self.transform(img2)
 
         label1 = 0
-        label2 = 0
         if self.label_handler.has_labels():
             label1 = self.label_handler.get_label(index1)
-            label2 = self.label_handler.get_label(index2)
-        return img1, img2, label1, label2
+        return img1, label1
 
     def __len__(self):
         return self.data_npz.shape[0]

common/utils.py

@@ -123,7 +123,7 @@ def random_idx():
         return random.randint(0, dataset.__len__() - 1)
 
     sample_idx = {}
-    dset_name = dataset.name()
+    dset_name = dataset.name
     if dset_name.lower() == 'dsprites':
         fixed_idx = [87040, 332800, 578560]  # square ellipse heart
         sample_idx = {'{}_{}'.format(c.FIXED, c.SQUARE): fixed_idx[0],
@@ -156,7 +156,7 @@ def random_idx():
 
         labels[key] = None
         if dataset.has_labels():
-            labels[key] = data[2].to(device, dtype=torch.long).unsqueeze(0)
+            labels[key] = data[1].to(device, dtype=torch.long).unsqueeze(0)
 
     return images, labels