Tripletloss, TripletCNN and further Tensorflow 2 support

config_test/test_train_nodaug.yml

@@ -0,0 +1,12 @@
+test:
+    orig:
+      daug_params: nodaug.yml
+      metrics: 
+        - accuracy 
+        - top3
+train:
+    orig:
+      daug_params: nodaug.yml
+      metrics: 
+        - accuracy 
+        - top3

config_train/cifar/tripletcnn/noreg_bn.yml

@@ -0,0 +1,46 @@
+network: 
+  name: tripletcnn
+  reg:
+    dropout: 0.
+    weight_decay: !!null
+  batch_norm: True
+optimizer:
+  name: Adam
+  loss: categorical_crossentropy
+#   momentum: 0.9
+#   nesterov: True
+  daug_invariance_params_file: noinv.yml
+  class_invariance_params_file: noinv.yml
+train:
+  lr:
+    init_lr: 0.001 
+    decay_factor: 0.1
+    decay_epochs: [300]
+  batch_size:
+    tr: 128
+    val: 128
+  epochs: 40
+  simulate:
+    bs_lr: bs # bs or lr
+    rep_samples: !!null
+    norep_samples: False
+    true_epochs: False
+data:
+  data_file: ~/datasets/hdf5/cifar10.hdf5
+  group_tr: cifar10_train
+  group_val: cifar10_test
+  shuffle_train_val: False
+  chunk_size: !!null
+  queue_size: 50
+daug:
+  nodaug: nodaug.yml
+  daug_params_file: nodaug.yml
+  aug_per_img_tr: 1
+  aug_per_img_val: 1
+seeds:
+  tf: 33
+  np: 39
+  daug: 59
+  batch_shuffle: 47
+  train_val: 27
+metrics: [accuracy]

data_input.py

@@ -107,7 +107,7 @@ def get_generator(images, **dict_params):
 
 
 def generate_batches(image_gen, images, labels, batch_size, aug_per_im, shuffle, 
-                     seed=None, n_inv_layers=0):
+                     seed=None, n_inv_layers=0, uses_triplet_loss=False):
     for batch in image_gen.flow_dask(images, labels,
                                      batch_size=batch_size,
                                      aug_per_im=aug_per_im,
@@ -117,18 +117,31 @@ def generate_batches(image_gen, images, labels, batch_size, aug_per_im, shuffle,
             # The list of targets for the invariance outputs are:
             # [daug_invariance_target, class_invariance_target, mean: ones]
             # for each invariance layer
+            invariance_targets = []
+
             bs = batch[0].shape[0]
             class_inv_y = np.dot(batch[1][0], batch[1][0].T)
             ones = np.ones([bs, bs], dtype=np.uint8)
             tril = np.tril_indices(bs)
             class_inv_y[tril] = 0
             ones[tril] = 0
+
             daug_inv_y = np.stack([batch[1][1], class_inv_y], axis=2)
+            invariance_targets.append(daug_inv_y)
+
             class_inv_y = np.stack([class_inv_y, ones], axis=2)
-            invariance_targets = [daug_inv_y, 
-                                  class_inv_y,
-                                  np.ones(batch[0].shape[0], dtype=np.uint8)] \
-                                 * n_inv_layers
+            invariance_targets.append(class_inv_y)
+
+            if uses_triplet_loss:
+                # append the daug inputs another time since it contains
+                # both class and daug labels needed for the triplet loss
+                invariance_targets.append(daug_inv_y)
+
+            mean_inv_y = np.ones(batch[0].shape[0], dtype=np.uint8)
+            invariance_targets.append(mean_inv_y)
+
+            invariance_targets *= n_inv_layers
+
             yield (batch[0],
                    [batch[1][0]] + invariance_targets)
         else:

examples/score-invariance.sh

@@ -0,0 +1,16 @@
+for MODEL in r26 r29
+do
+    echo -e "\nCOMPUTING INVARIANCE FOR MODEL $MODEL\n"
+    for EPOCH in 010 020 030 040
+    do
+        echo -e "\nMODEL $MODEL, EPOCH $EPOCH\n"
+        python invariance_numpy.py \
+        --data_file ~/Storage/Datasets/cifar10/hdf5/cifar10.hdf5 \
+        --group test \
+        --daug_params heavier.yml \
+        --model log/cifar10/tripletcnn/noreg/bn/inv/$MODEL/model_*_$EPOCH \
+        --output_mse_matrix_hdf5 mse_matrix_ep$EPOCH \
+        --output_pickle invscores_ep$EPOCH \
+        --layer_regex flatten
+    done
+done

examples/score-test-prediction.sh

@@ -0,0 +1,15 @@
+for MODEL in r26 r29
+do
+    for EPOCH in 010 020 030 040
+    do
+        echo -e "\nMODEL $MODEL, EPOCH $EPOCH\n"
+        python test.py \
+        --data_file ~/Storage/Datasets/cifar10/hdf5/cifar10.hdf5 \
+        --group_tr train \
+        --group_tt test \
+        --test_config config_test/test_train_nodaug.yml \
+        --model log/cifar10/tripletcnn/noreg/bn/inv/$MODEL/model_*_$EPOCH \
+        --output_dir -1 \
+        --output_basename testscores/testscore_ep$EPOCH 
+    done
+done

examples/train-tripletcnn-cifar-cls.sh

@@ -0,0 +1,20 @@
+# This script trains the TripletCNN architecture on CIFAR-10 with data 
+# augmentation and class triplet loss
+#
+# It is meant to to be run from the root of the repository
+# Consider changing the paths to files and directories
+#
+# CIFAR-10
+#
+# All-CNN
+#
+# no-reg bn heavier inv
+python train.py \
+--data_file ~/Storage/Datasets/cifar10/hdf5/cifar10.hdf5 \
+--group_tr train \
+--group_val test \
+--train_dir ./log/cifar10/tripletcnn/noreg/bn/inv/r24/ \
+--daug_params heavier.yml \
+--train_config_file config_train/cifar/tripletcnn/noreg_bn.yml \
+--save_model_every 10 \
+--no_fit_generator

examples/train-tripletcnn-cifar-daug.sh

@@ -0,0 +1,25 @@
+# This script trains the TripletCNN architecture on CIFAR-10 with data 
+# augmentation and class triplet loss
+#
+# It is meant to to be run from the root of the repository
+# Consider changing the paths to files and directories
+#
+# CIFAR-10
+#
+# All-CNN
+#
+# no-reg bn heavier inv
+python train.py \
+--data_file ~/Storage/Datasets/cifar10/hdf5/cifar10.hdf5 \
+--group_tr train \
+--group_val test \
+--train_dir ./log/cifar10/tripletcnn/noreg/bn/inv/r31/ \
+--daug_params heavier.yml \
+--train_config_file config_train/cifar/tripletcnn/noreg_bn.yml \
+--aug_per_img_tr 16 \
+--daug_invariance_params loss1_exp.yml \
+--class_invariance_params noinv.yml \
+--save_model_every 10 \
+--no_fit_generator \
+--resume_training ./log/cifar10/tripletcnn/noreg/bn/inv/r31/model_*_030 \
+--triplet_loss 

invariance.py

@@ -26,15 +26,19 @@
 
 from data_input import hdf52dask, get_generator, batch_generator
 from utils import get_daug_scheme_path
-from utils import pairwise_loss, mean_loss, invariance_loss
+from utils import pairwise_loss, mean_loss, invariance_loss, triplet_loss
 from activations import get_activations
 
-import keras.backend as K
-from keras.models import load_model
-import keras.losses
-keras.losses.pairwise_loss = pairwise_loss
-keras.losses.invariance_loss = invariance_loss
-keras.losses.mean_loss = mean_loss
+import tensorflow.compat.v1 as tf
+from tensorflow.compat.v1 import keras
+import tensorflow.compat.v1.keras.backend as K
+from tensorflow.compat.v1.keras.models import load_model
+import tensorflow.compat.v1.keras.losses
+
+# Disable eager execution, otherwise K.function will not work when passing
+# K.learning_phase() as input (see issue
+# https://github.com/tensorflow/tensorflow/issues/34201)
+tf.disable_eager_execution()
 
 import os
 import argparse
@@ -50,6 +54,10 @@
 
 def main(argv=None):
 
+    config = tf.ConfigProto()
+    config.gpu_options.allow_growth = True
+    K.set_session(tf.Session(config=config))
+
 #     cluster = LocalCluster(dashboard_address=None)
 #     client = Client(cluster, memory_limit='{}GB'.format(FLAGS.memory_limit),
 #                     processes=False)
@@ -76,7 +84,14 @@ def main(argv=None):
 
     # Initialize the network model
     model_filename = FLAGS.model
-    model = load_model(model_filename)
+    model = load_model(
+        model_filename, 
+        custom_objects = {
+            'invariance_loss': invariance_loss,
+            'triplet_loss': triplet_loss,
+            'pairwise_loss': pairwise_loss,
+            'mean_loss': mean_loss
+            })
 
     # Print the model summary
     model.summary()
@@ -119,7 +134,14 @@ def main(argv=None):
         # Reload the model
         if layer_idx > 0:
             K.clear_session()
-            model = load_model(model_filename)
+            model = load_model(
+                model_filename, 
+                custom_objects = {
+                    'invariance_loss': invariance_loss,
+                    'triplet_loss': triplet_loss,
+                    'pairwise_loss': pairwise_loss,
+                    'mean_loss': mean_loss
+                    })
 
         layer = model.get_layer(layer_name)