Merge branch 'master' into label-scale-models

vanvalenlab · Aug 29, 2019 · 2d039b0 · 2d039b0
2 parents 76f2d81 + a1aad5b
commit 2d039b0
Show file tree

Hide file tree

Showing 13 changed files with 118 additions and 86 deletions.
diff --git a/deepcell/training.py b/deepcell/training.py
@@ -53,7 +53,7 @@
 def train_model_sample(model,
                        dataset,
                        expt='',
-                       test_size=.1,
+                       test_size=.2,
                        n_epoch=10,
                        batch_size=32,
                        num_gpus=None,
@@ -72,7 +72,7 @@ def train_model_sample(model,
                        flip=False,
                        shear=0,
                        zoom_range=0,
-                       seed=None,
+                       seed=0,
                        **kwargs):
     """Train a model using sample mode.
 
@@ -217,7 +217,7 @@ def loss_function(y_true, y_pred):
 def train_model_conv(model,
                      dataset,
                      expt='',
-                     test_size=.1,
+                     test_size=.2,
                      n_epoch=10,
                      batch_size=1,
                      num_gpus=None,
@@ -234,7 +234,7 @@ def train_model_conv(model,
                      flip=True,
                      shear=0,
                      zoom_range=0,
-                     seed=None,
+                     seed=0,
                      **kwargs):
     """Train a model using fully convolutional mode.
 
@@ -409,7 +409,7 @@ def loss_function(y_true, y_pred):
 def train_model_siamese_daughter(model,
                                  dataset,
                                  expt='',
-                                 test_size=.1,
+                                 test_size=.2,
                                  n_epoch=100,
                                  batch_size=1,
                                  num_gpus=None,
@@ -428,7 +428,7 @@ def train_model_siamese_daughter(model,
                                  flip=True,
                                  shear=0,
                                  zoom_range=0,
-                                 seed=None,
+                                 seed=0,
                                  **kwargs):
     is_channels_first = K.image_data_format() == 'channels_first'
 
@@ -545,7 +545,7 @@ def loss_function(y_true, y_pred):
 def train_model_retinanet(model,
                           dataset,
                           expt='',
-                          test_size=.1,
+                          test_size=.2,
                           n_epoch=10,
                           batch_size=1,
                           num_gpus=None,
@@ -575,7 +575,7 @@ def train_model_retinanet(model,
                           shear=0,
                           zoom_range=0,
                           compute_map=True,
-                          seed=None,
+                          seed=0,
                           **kwargs):
     """Train a RetinaNet model from the given backbone.
 

diff --git a/deepcell/utils/data_utils.py b/deepcell/utils/data_utils.py
@@ -52,7 +52,7 @@
 from deepcell.utils.tracking_utils import load_trks
 
 
-def get_data(file_name, mode='sample', test_size=.1, seed=None):
+def get_data(file_name, mode='sample', test_size=.2, seed=0):
     """Load data from NPZ file and split into train and test sets
 
     Args:

diff --git a/scripts/feature_pyramids/PanOpticFPN.ipynb b/scripts/feature_pyramids/PanOpticFPN.ipynb
@@ -38,7 +38,10 @@
    "source": [
     "# Download the data (saves to ~/.keras/datasets)\n",
     "filename = 'HeLa_S3.npz'\n",
-    "(X_train, y_train), (X_test, y_test) = deepcell.datasets.hela_s3.load_data(filename)\n",
+    "test_size = 0.2 # % of data saved as test\n",
+    "seed = 0 # seed for random train-test split\n",
+    "\n",
+    "(X_train, y_train), (X_test, y_test) = deepcell.datasets.hela_s3.load_data(filename, test_size=test_size, seed=seed)\n",
     "\n",
     "print('X.shape: {}\\ny.shape: {}'.format(X_train.shape, y_train.shape))"
    ]
@@ -118,7 +121,6 @@
     "model_name = '{}_retinanet'.format(backbone)\n",
     "\n",
     "n_epoch = 5  # Number of training epochs\n",
-    "test_size = .20  # % of data saved as test\n",
     "lr = 1e-5\n",
     "\n",
     "optimizer = Adam(lr=lr, clipnorm=0.001)\n",
@@ -285,6 +287,7 @@
     "    iou_threshold=0.5,\n",
     "    max_detections=100,\n",
     "    test_size=test_size,\n",
+    "    seed=seed,\n",
     "    optimizer=optimizer,\n",
     "    batch_size=batch_size,\n",
     "    n_epoch=n_epoch,\n",

diff --git a/scripts/feature_pyramids/RetinaMask.ipynb b/scripts/feature_pyramids/RetinaMask.ipynb
@@ -38,7 +38,10 @@
    "source": [
     "# Download the data (saves to ~/.keras/datasets)\n",
     "filename = 'HeLa_S3.npz'\n",
-    "(X_train, y_train), (X_test, y_test) = deepcell.datasets.hela_s3.load_data(filename)\n",
+    "test_size = 0.1 # % of data saved as test\n",
+    "seed = 0 # seed for random train-test split\n",
+    "\n",
+    "(X_train, y_train), (X_test, y_test) = deepcell.datasets.hela_s3.load_data(filename, test_size=test_size, seed=seed)\n",
     "\n",
     "print('X.shape: {}\\ny.shape: {}'.format(X_train.shape, y_train.shape))"
    ]
@@ -111,7 +114,6 @@
     "backbone = 'resnet50'  # vgg16, vgg19, resnet50, densenet121, densenet169, densenet201\n",
     "\n",
     "n_epoch = 3  # Number of training epochs\n",
-    "test_size = .10  # % of data saved as test\n",
     "lr = 1e-5\n",
     "\n",
     "optimizer = Adam(lr=lr, clipnorm=0.001)\n",
@@ -251,6 +253,7 @@
     "    iou_threshold=0.5,\n",
     "    max_detections=100,\n",
     "    test_size=test_size,\n",
+    "    seed=seed,\n",
     "    optimizer=optimizer,\n",
     "    batch_size=batch_size,\n",
     "    n_epoch=n_epoch,\n",

diff --git a/scripts/feature_pyramids/RetinaNet.ipynb b/scripts/feature_pyramids/RetinaNet.ipynb
@@ -38,7 +38,10 @@
    "source": [
     "# Download the data (saves to ~/.keras/datasets)\n",
     "filename = 'HeLa_S3.npz'\n",
-    "(X_train, y_train), (X_test, y_test) = deepcell.datasets.hela_s3.load_data(filename)\n",
+    "test_size = 0.2 # % of data saved as test\n",
+    "seed = 0 # seed for random train-test split\n",
+    "\n",
+    "(X_train, y_train), (X_test, y_test) = deepcell.datasets.hela_s3.load_data(filename, test_size=test_size, seed=seed)\n",
     "\n",
     "print('X.shape: {}\\ny.shape: {}'.format(X_train.shape, y_train.shape))"
    ]
@@ -118,7 +121,6 @@
     "model_name = '{}_retinanet'.format(backbone)\n",
     "\n",
     "n_epoch = 20  # Number of training epochs\n",
-    "test_size = .20  # % of data saved as test\n",
     "lr = 1e-5\n",
     "\n",
     "optimizer = Adam(lr=lr, clipnorm=0.001)\n",
@@ -290,13 +292,7 @@
       "9978 instances of class 0 with average precision: 0.8226\n",
       "mAP: 0.8226\n",
       "160/160 [==============================] - 125s 781ms/step - loss: 0.7435 - regression_loss: 0.6493 - classification_loss: 0.0942 - val_loss: 0.5919 - val_regression_loss: 0.5273 - val_classification_loss: 0.0646\n",
-      "Epoch 16/20\n"
-     ]
-    },
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
+      "Epoch 16/20\n",
       "159/160 [============================>.] - ETA: 0s - loss: 0.7279 - regression_loss: 0.6376 - classification_loss: 0.0902\n",
       "Epoch 00016: val_loss did not improve from 0.59190\n",
       "9978 instances of class 0 with average precision: 0.8234\n",
@@ -351,6 +347,7 @@
     "    iou_threshold=0.5,\n",
     "    max_detections=100,\n",
     "    test_size=test_size,\n",
+    "    seed=seed,\n",
     "    optimizer=optimizer,\n",
     "    batch_size=batch_size,\n",
     "    n_epoch=n_epoch,\n",