diff --git a/answer_classifier_cached_features/train_no_pretraining.py b/answer_classifier_cached_features/train_no_pretraining.py
index d6a5b8a6dc5d951ee1a6736974d3e49a46d58a5c..3b3899b9a69b366fbbf08766f750423858fc0ba5 100644
--- a/answer_classifier_cached_features/train_no_pretraining.py
+++ b/answer_classifier_cached_features/train_no_pretraining.py
@@ -3,6 +3,7 @@ import object_attribute_classifier_cached_features.inference as feature_graph
import region_relevance_network.inference as relevance_graph
import answer_classifier_cached_features.inference as answer_graph
from tftools import var_collect, placeholder_management
+import tftools.train as multi_rate_train
import tftools.data
import losses
import constants
@@ -455,6 +456,7 @@ def create_vqa_batch_generator():
data_mgr = vqa_data.data(
constants.vqa_train_resnet_feat_dir,
constants.vqa_train_anno,
+ constants.vqa_train_subset_qids,
constants.vocab_json,
constants.vqa_answer_vocab_json,
constants.image_size,
@@ -462,11 +464,13 @@ def create_vqa_batch_generator():
constants.num_negative_answers,
resnet_feat_dim=constants.resnet_feat_dim)
+ num_train_subset_questions = len(data_mgr.qids)
+
index_generator = tftools.data.random(
constants.answer_batch_size,
- constants.num_train_questions,
+ num_train_subset_questions,
constants.answer_num_epochs,
- constants.answer_offset)
+ 0)
batch_generator = tftools.data.async_batch_generator(
data_mgr,
@@ -483,16 +487,20 @@ def create_vgenome_batch_generator():
constants.object_labels_json,
constants.attribute_labels_json,
constants.regions_json,
+ constants.genome_train_subset_region_ids,
constants.image_size,
channels=3,
resnet_feat_dim=constants.resnet_feat_dim,
mean_image_filename=None)
+ num_train_subset_regions = len(data_mgr.region_ids)
+ print num_train_subset_regions
+
index_generator = tftools.data.random(
constants.num_regions_with_labels,
- constants.region_num_samples,
+ num_train_subset_regions,
constants.region_num_epochs,
- constants.region_offset)
+ 0)
batch_generator = tftools.data.async_batch_generator(
data_mgr,
@@ -515,7 +523,9 @@ class attach_optimizer():
self.lr = lr
with graph.tf_graph.as_default():
all_trainable_vars = tf.trainable_variables()
- self.not_to_train = []#+ graph.object_attribute_vars
+
+ self.not_to_train = graph.object_attribute_vars + graph.word_vec_vars
+
vars_to_train = [
var for var in all_trainable_vars
if var not in self.not_to_train]
@@ -527,13 +537,28 @@ class attach_optimizer():
all_vars = tf.all_variables()
self.ops = dict()
+
+ self.optimizer = multi_rate_train.MultiRateOptimizer(
+ tf.train.AdamOptimizer)
+
+ self.optimizer.add_variables(
+ self.graph.object_attribute_vars + self.graph.word_vec_vars,
+ learning_rate = 1.0*self.lr)
- self.add_adam_optimizer(
- graph.total_loss,
+
+ self.optimizer.add_variables(
vars_to_train,
- 'optimizer')
+ learning_rate = self.lr)
+
+ self.train_op = self.optimizer.minimize(graph.total_loss)
+
+ # self.add_adam_optimizer(
+ # graph.total_loss,
+ # vars_to_train,
+ # 'optimizer')
- self.train_op = self.group_all_train_ops()
+ # self.train_op = self.group_all_train_ops()
+
all_vars_with_opt_vars = tf.all_variables()
self.opt_vars = [var for var in all_vars_with_opt_vars if var not in all_vars]
diff --git a/constants_vision_gpu_2.py b/constants_vision_gpu_2.py
index aa195ba13d2ec704bef230a185a42b23a028c1f6..16ae816e0bf74a8782baea6a193fef115581582c 100644
--- a/constants_vision_gpu_2.py
+++ b/constants_vision_gpu_2.py
@@ -5,7 +5,7 @@ def mkdir_if_not_exists(dir_name):
if not os.path.exists(dir_name):
os.mkdir(dir_name)
-experiment_name = 'QA_explicit_dot_joint_training_pretrained_fix_pretrained'
+experiment_name = 'QA_genome_split'
#experiment_name = 'object_attribute_classifier_large_images'
# Global output directory (all subexperiments will be saved here)
global_output_dir = '/data/tanmay/GenVQA_Exp_Results'
@@ -56,6 +56,18 @@ vocab_json = os.path.join(
data_absolute_path,
'restructured/vocab_subset.json')
+genome_train_subset_region_ids = os.path.join(
+ data_absolute_path,
+ 'restructured/train_subset_region_ids.json')
+
+genome_train_held_out_region_ids = os.path.join(
+ data_absolute_path,
+ 'restructured/train_held_out_region_ids.json')
+
+genome_test_region_ids = os.path.join(
+ data_absolute_path,
+ 'restructured/test_region_ids.json')
+
num_object_labels = 1000
num_attribute_labels = 1000
@@ -63,12 +75,12 @@ num_attribute_labels = 1000
# First 80% meant to be used for training
# Next 10% is set aside for validation
# Last 10% is to be used for testing
-num_total_regions = 1951768
-num_train_regions = 1561416 # First 80%
-num_val_regions = 195176 # Next 10%
-num_test_regions = num_total_regions \
- - num_train_regions \
- - num_val_regions
+# num_total_regions = 1951768
+# num_train_regions = 1561416 # First 80%
+# num_val_regions = 195176 # Next 10%
+# num_test_regions = num_total_regions \
+# - num_train_regions \
+# - num_val_regions
# Pretrained resnet ckpt
resnet_ckpt = '/home/nfs/tgupta6/data/Resnet/' + \
@@ -88,7 +100,7 @@ pretrained_vocab_word_vectors_npy = os.path.join(
# Object Attribute Classifier Training Params
region_batch_size = 200
-region_num_samples = num_train_regions
+# region_num_samples = num_train_regions
region_num_epochs = 4
region_offset = 0
region_queue_size = 400
@@ -171,11 +183,11 @@ vqa_answer_vocab_json = os.path.join(
answer_batch_size = 50
answer_num_epochs = 10
answer_offset = 0
-answer_obj_atr_loss_wt = 0.1
+answer_obj_atr_loss_wt = 0.0
answer_regularization_coeff = 1e-5
answer_queue_size = 500
answer_embedding_dim = 600
-answer_lr = 1e-4
+answer_lr = 1e-3
answer_log_every_n_iter = 500
answer_output_dir = os.path.join(
global_experiment_dir,
@@ -184,8 +196,8 @@ answer_output_dir = os.path.join(
mkdir_if_not_exists(answer_output_dir)
pretrained_model = '/home/nfs/tgupta6/projects/GenVQA/Exp_Results/' +\
- 'pretrained_object_attribute_classifier/' +\
- 'obj_atr_model_77500'
+ 'object_attribute_classifier_large_images_vqa_split/' +\
+ 'object_attribute_classifiers/model-80000'
answer_model = os.path.join(
answer_output_dir,
@@ -195,12 +207,12 @@ answer_model = os.path.join(
num_regions_with_labels = 100
# Answer fine tune params
-answer_fine_tune_from_iter = 18500
+answer_fine_tune_from_iter = 22500
answer_fine_tune_from = answer_model + '-' + str(answer_fine_tune_from_iter)
# Answer eval params
answer_eval_on = 'val'
-answer_model_to_eval = answer_model + '-18500'
+answer_model_to_eval = answer_model + '-69500'
answer_eval_data_json = os.path.join(
answer_output_dir,