fixed loss

Author: minerva-ml

common_blocks/models.py

@@ -1,6 +1,7 @@
 import numpy as np
 import torch
 import torch.nn as nn
+from torch.nn import functional as F
 import torch.optim as optim
 from toolkit.pytorch_transformers.models import Model
 from torch.autograd import Variable
@@ -167,10 +168,11 @@ def set_loss(self):
         if self.activation_func == 'softmax':
             raise NotImplementedError('No softmax loss defined')
         elif self.activation_func == 'sigmoid':
-            loss_function = lovasz_loss
-            # loss_function = DiceLoss()
-            # loss_function = FocalWithLogitsLoss()
+            loss_function = weighted_sum_loss
             # loss_function = nn.BCEWithLogitsLoss()
+            # loss_function = DiceWithLogitsLoss()
+            # loss_function = lovasz_loss
+            # loss_function = FocalWithLogitsLoss()
         else:
             raise Exception('Only softmax and sigmoid activations are allowed')
         self.loss_function = [('mask', loss_function, 1.0)]
@@ -191,34 +193,49 @@ def load(self, filepath):
 
 
 class FocalWithLogitsLoss(nn.Module):
-    def __init__(self, alpha=1.0, gamma=1.0):
+    def __init__(self, alpha=1.0, gamma=1.0, reduction='elementwise_mean'):
         super().__init__()
         self.alpha = alpha
         self.gamma = gamma
+        self.reduction = reduction
 
-    def forward(self, input, target):
-        if not (target.size() == input.size()):
-            raise ValueError("Target size ({}) must be the same as input size ({})".format(target.size(), input.size()))
+    def forward(self, output, target):
+        if not (target.size() == output.size()):
+            raise ValueError(
+                "Target size ({}) must be the same as input size ({})".format(target.size(), output.size()))
 
-        max_val = (-input).clamp(min=0)
-        logpt = input - input * target + max_val + ((-max_val).exp() + (-input - max_val).exp()).log()
+        max_val = (-output).clamp(min=0)
+        logpt = output - output * target + max_val + ((-max_val).exp() + (-output - max_val).exp()).log()
         pt = torch.exp(-logpt)
         at = self.alpha * target + (1 - target)
         loss = at * ((1 - pt).pow(self.gamma)) * logpt
-        return loss
 
+        if self.reduction == 'none':
+            return loss
+        elif self.reduction == 'elementwise_mean':
+            return loss.mean()
+        else:
+            return loss.sum()
 
-class DiceLoss(nn.Module):
+
+class DiceWithLogitsLoss(nn.Module):
     def __init__(self, smooth=0, eps=1e-7):
         super().__init__()
         self.smooth = smooth
         self.eps = eps
 
     def forward(self, output, target):
+        output = F.sigmoid(output)
         return 1 - (2 * torch.sum(output * target) + self.smooth) / (
                 torch.sum(output) + torch.sum(target) + self.smooth + self.eps)
 
 
+def weighted_sum_loss(output, target):
+    bce = nn.BCEWithLogitsLoss()(output, target)
+    dice = DiceWithLogitsLoss()(output, target)
+    return bce + 0.25 * dice
+
+
 def lovasz_loss(output, target):
     target = target.long()
     return lovasz_hinge(output, target)
@@ -246,6 +263,6 @@ def callbacks_network(callbacks_config):
     init_lr_finder = cbk.InitialLearningRateFinder()
     return cbk.CallbackList(
         callbacks=[experiment_timing, training_monitor, validation_monitor,
-                   model_checkpoints, lr_scheduler, neptune_monitor, early_stopping,
-                   # init_lr_finder
+                   model_checkpoints, neptune_monitor, early_stopping,
+                   lr_scheduler,  #init_lr_finder,
                    ])

main.py

@@ -30,7 +30,7 @@
 CLONE_EXPERIMENT_DIR_FROM = ''  # When running eval in the cloud specify this as for example /input/SAL-14/output/experiment
 OVERWRITE_EXPERIMENT_DIR = False
 DEV_MODE = False
-USE_TTA = True
+USE_TTA = False
 
 if OVERWRITE_EXPERIMENT_DIR and os.path.isdir(EXPERIMENT_DIR):
     shutil.rmtree(EXPERIMENT_DIR)

neptune.yaml

@@ -1,7 +1,7 @@
 #project: USERNAME/PROJECT
 
 name: airbus_ships_challenge
-tags: [solution-1]
+tags: [solution-2]
 
 metric:
   channel: 'f2'
@@ -25,7 +25,7 @@ parameters:
   # Data Paths
   train_images_dir:     /public/challenges/kaggle-ship-detection/train
   test_images_dir:      /public/challenges/kaggle-ship-detection/test
-  metadata_filepath:    /outputs/ships_metadata.csv
+  metadata_filepath:    /output/metadata.csv
   annotation_file:      /public/challenges/kaggle-ship-detection/train_ship_segmentations.csv
   masks_overlayed_dir:  /output/masks_overlayed
 
@@ -38,17 +38,17 @@ parameters:
   resize_target_size: 256
   pad_method: symmetric
   target_format: 'joblib'
-  dev_mode_size: 500
+  dev_mode_size: 50
 
   # General parameters
   image_h: 256
   image_w: 256
   image_channels: 3
-  training_sampler_size: 2000
+  training_sampler_size: 2500
   training_sampler_empty_fraction: 0.0
   evaluation_size: 10000
   evaluation_empty_fraction: 0.52
-  in_train_evaluation_size: 1000
+  in_train_evaluation_size: 500
   fine_tuning: 1
 
   # Network parameters
@@ -57,13 +57,13 @@ parameters:
   architecture: UNetSeResNetXt
 
   # Training schedule
-  epochs_nr: 10
+  epochs_nr: 1000
   batch_size_train: 4
   batch_size_inference: 4
-  lr: 0.0007
+  lr: 0.0006
   momentum: 0.9
   gamma: 0.95
-  patience: 10
+  patience: 20
   validation_metric_name: 'f2'
   minimize_validation_metric: 0
   reduce_factor: 0.5