From d80c835a54f30bd1e8805217636d7f992e32b6fd Mon Sep 17 00:00:00 2001
From: tgupta6 <tgupta6@illinois.edu>
Date: Sat, 20 Aug 2016 15:01:22 -0500
Subject: [PATCH] Circular batch and train held out
---
answer_classifier_cached_features/eval.py | 23 ++++++++++---------
.../select_best_model.py | 21 +++++++++--------
answer_classifier_cached_features/train.py | 2 +-
constants_crunchy.py | 5 ++--
tftools/data.py | 10 ++++----
5 files changed, 34 insertions(+), 27 deletions(-)
diff --git a/answer_classifier_cached_features/eval.py b/answer_classifier_cached_features/eval.py
index 09d5e6d..5e32665 100644
--- a/answer_classifier_cached_features/eval.py
+++ b/answer_classifier_cached_features/eval.py
@@ -38,23 +38,18 @@ def create_initializer(graph, sess, model):
return initializer()
def create_batch_generator(mode):
- if mode=='val_subset':
+ if mode=='val':
vqa_resnet_feat_dir = constants.vqa_val_resnet_feat_dir
vqa_anno = constants.vqa_val_anno
- num_questions = constants.num_val_subset_questions
+ num_questions = constants.num_val_questions
offset = 0
- elif mode=='val_rest':
+ elif mode=='val_subset':
vqa_resnet_feat_dir = constants.vqa_val_resnet_feat_dir
vqa_anno = constants.vqa_val_anno
- num_questions = constants.num_val_rest_questions
- offset = constants.num_val_subset_questions
- elif mode=='train':
- vqa_resnet_feat_dir = constants.vqa_train_resnet_feat_dir
- vqa_anno = constants.vqa_train_anno
- num_questions = constants.num_train_questions
+ num_questions = constants.num_val_subset_questions
offset = 0
else:
- print "mode needs to be one of {'train','val_subset','val_rest'}, found " + mode
+ print "mode needs to be one of {'val','val_subset'}, found " + mode
data_mgr = vqa_data.data(
vqa_resnet_feat_dir,
@@ -143,6 +138,7 @@ class eval_mgr():
self.correct = 0
self.total = 0
self.results = []
+ self.seen_qids = set()
def eval(self, iter, eval_vars_dict, batch):
batch_size = len(batch['question_unencoded'])
@@ -173,7 +169,12 @@ class eval_mgr():
'question_id': int(question_id),
'answer': pred_answer
}
- self.results.append(result_entry)
+
+ if question_id not in self.seen_qids:
+ self.seen_qids.add(question_id)
+ self.results.append(result_entry)
+ else:
+ print 'Already evaluated on this sample'
self.eval_data[str(question_id)] = dict_entry
diff --git a/answer_classifier_cached_features/select_best_model.py b/answer_classifier_cached_features/select_best_model.py
index ee6b270..f483337 100644
--- a/answer_classifier_cached_features/select_best_model.py
+++ b/answer_classifier_cached_features/select_best_model.py
@@ -39,16 +39,13 @@ def create_initializer(graph, sess, model):
return initializer()
def create_batch_generator(mode):
- if mode=='val':
- vqa_resnet_feat_dir = constants.vqa_val_resnet_feat_dir
- vqa_anno = constants.vqa_val_anno
- num_questions = constants.num_val_questions
- elif mode=='train':
+ if mode=='train_subset':
vqa_resnet_feat_dir = constants.vqa_train_resnet_feat_dir
vqa_anno = constants.vqa_train_anno
- num_questions = constants.num_train_questions
+ num_questions = constants.num_train_held_out_questions
+ offset = constants.num_train_subset_questions
else:
- print "mode needs to be one of {'train','test','val'}, found " + mode
+ print "mode needs to be one of {'train_subset'}, found " + mode
data_mgr = vqa_data.data(
vqa_resnet_feat_dir,
@@ -64,7 +61,7 @@ def create_batch_generator(mode):
constants.answer_batch_size,
num_questions,
1,
- 0)
+ offset)
batch_generator = tftools.data.async_batch_generator(
data_mgr,
@@ -135,6 +132,7 @@ class eval_mgr():
self.correct = 0
self.total = 0
self.results = []
+ self.seen_qids = set()
def eval(self, iter, eval_vars_dict, batch):
batch_size = len(batch['question_unencoded'])
@@ -151,7 +149,12 @@ class eval_mgr():
'question_id': int(question_id),
'answer': pred_answer
}
- self.results.append(result_entry)
+
+ if question_id not in self.seen_qids:
+ self.seen_qids.add(question_id)
+ self.results.append(result_entry)
+ else:
+ print 'Already evaluated on this sample'
self.total += batch_size
diff --git a/answer_classifier_cached_features/train.py b/answer_classifier_cached_features/train.py
index 608516a..63dfdf3 100644
--- a/answer_classifier_cached_features/train.py
+++ b/answer_classifier_cached_features/train.py
@@ -475,7 +475,7 @@ def create_vqa_batch_generator():
index_generator = tftools.data.random(
constants.answer_batch_size,
- constants.num_train_questions,
+ constants.num_train_subset_questions,
constants.answer_num_epochs,
constants.answer_offset)
diff --git a/constants_crunchy.py b/constants_crunchy.py
index ac7d315..df19281 100644
--- a/constants_crunchy.py
+++ b/constants_crunchy.py
@@ -154,9 +154,10 @@ vqa_answer_vocab_json = os.path.join(
# VQA dataset params
num_train_questions = 248349
-num_val_subset_questions = 10000
+num_train_held_out_questions = 24835
+num_train_subset_questions = num_train_question - num_train_held_out_questions
num_val_questions = 121512
-num_val_rest_questions = num_val_questions - num_val_subset_questions
+num_val_subset_questions = 10000
num_test_questions = 0
diff --git a/tftools/data.py b/tftools/data.py
index 1a60f44..a872a06 100644
--- a/tftools/data.py
+++ b/tftools/data.py
@@ -6,21 +6,23 @@ import time
def sequential(batch_size, num_samples, num_epochs=1, offset=0):
"""Generate sequence indices.
"""
+ num_samples_ = int(batch_size*np.ceil(num_samples/float(batch_size)))
for epoch in range(num_epochs):
- indices = np.arange(num_samples) + offset
+ indices = np.arange(num_samples_)%num_samples + offset
indices = indices.tolist()
- for i in range(0, num_samples - batch_size + 1, batch_size):
+ for i in range(0, num_samples_ - batch_size + 1, batch_size):
yield indices[i:i+batch_size]
def random(batch_size, num_samples, num_epochs, offset=0):
"""Generate random indices.
"""
+ num_samples_ = int(batch_size*np.ceil(num_samples/float(batch_size)))
for epoch in range(num_epochs):
# np.random.seed(epoch)
- indices = np.random.permutation(num_samples) + offset
+ indices = np.random.permutation(num_samples_)%num_samples + offset
indices = indices.tolist()
- for i in range(0, num_samples - batch_size + 1, batch_size):
+ for i in range(0, num_samples_ - batch_size + 1, batch_size):
yield indices[i:i+batch_size]
--
GitLab