def main():
apollocaffe.set_device(0)
#apollocaffe.set_cpp_loglevel(0)
apollocaffe.set_random_seed(0)
np.random.seed(0)
job = sys.argv[1]
corpus_name = sys.argv[2]
config = util.Struct(**yaml.load(CONFIG))
if corpus_name == "abstract":
train_scenes, dev_scenes, test_scenes = corpus.load_abstract()
else:
assert corpus_name == "birds"
train_scenes, dev_scenes, test_scenes = corpus.load_birds()
apollo_net = ApolloNet()
print "loaded data"
print "%d training examples" % len(train_scenes)
listener0_model = Listener0Model(apollo_net, config.model)
speaker0_model = Speaker0Model(apollo_net, config.model)
sampling_speaker1_model = SamplingSpeaker1Model(apollo_net, config.model)
compiled_speaker1_model = CompiledSpeaker1Model(apollo_net, config.model)
if job == "train.base":
train(train_scenes, dev_scenes, listener0_model, apollo_net, config.opt)
train(train_scenes, dev_scenes, speaker0_model, apollo_net, config.opt)
apollo_net.save("models/%s.base.caffemodel" % corpus_name)
exit()
if job == "train.compiled":
apollo_net.load("models/%s.base.caffemodel" % corpus_name)
print "loaded model"
train(train_scenes, dev_scenes, compiled_speaker1_model, apollo_net,
config.opt)
apollo_net.save("models/%s.compiled.caffemodel" % corpus_name)
exit()
if job in ("sample.base", "sample.compiled"):
if job == "sample.base":
apollo_net.load("models/%s.base.caffemodel" % corpus_name)
else:
apollo_net.load("models/%s.compiled.caffemodel" % corpus_name)
print "loaded model"
if job == "sample.base":
models = {
"sampling_speaker1": sampling_speaker1_model,
}
elif job == "sample.compiled":
models = {
"compiled_speaker1": compiled_speaker1_model,
}
name = job.split(".")[1]
run_experiment("one_different", corpus_name, name, models, dev_scenes)
run_experiment("by_similarity", corpus_name, name, models, dev_scenes)
run_experiment("all_same", corpus_name, name, models, dev_scenes)
def make_abstract():
train_data, val_data, test_data = corpus.load_abstract()
by_prop = defaultdict(list)
for datum in val_data:
props = {(p.type_index, p.object_index) for p in datum.props}
props = {(i,j) if i not in (2,3) else (i,0) for i,j in props}
props = tuple(sorted(props))
by_prop[props].append(datum)
with open("experiments/all_same/abstract.ids.txt", "w") as id_f:
counter = 0
for key in by_prop:
if counter >= N_PAIRS:
break
images = set(d.image_id for d in by_prop[key])
if len(images) <= 3:
continue
images = list(images)
np.random.shuffle(images)
images = images[:2]
print >>id_f, ",".join(images) + ",0"
counter += 1
with open("experiments/one_different/abstract.ids.txt", "w") as id_f:
counter = 0
for key1 in by_prop:
if counter >= N_PAIRS:
break
keys = list(by_prop.keys())
np.random.shuffle(keys)
for key2 in by_prop:
if len(key1) != len(key2):
continue
if len(set(key1) ^ set(key2)) != 2:
continue
if key1 > key2:
continue
images = [by_prop[key1][0].image_id, by_prop[key2][0].image_id]
print >>id_f, ",".join(images) + ",2"
counter += 1
break
with open("experiments/by_similarity/abstract.ids.txt", "w") as id_f:
keys = list(by_prop.keys())
counter = 0
for key1 in by_prop:
if counter >= N_PAIRS:
break
attempts = 0
while attempts < 100:
attempts += 1
key2 = keys[np.random.randint(len(keys))]
similarity = len(set(key1) ^ set(key2))
if similarity > MAX_DIFFERENCES:
continue
if np.random.random() > 0.45 ** similarity:
continue
img1 = by_prop[key1][0].image_id
img2 = by_prop[key2][0].image_id
if img1 == img2:
continue
print >>id_f, "%s,%s,%s" % (img1, img2, similarity)
counter += 1
#print counter
break
def make_abstract():
train_data, val_data, test_data = corpus.load_abstract()
by_prop = defaultdict(list)
for datum in val_data:
props = {(p.type_index, p.object_index) for p in datum.props}
props = {(i, j) if i not in (2, 3) else (i, 0) for i, j in props}
props = tuple(sorted(props))
by_prop[props].append(datum)
with open("experiments/all_same/abstract.ids.txt", "w") as id_f:
counter = 0
for key in by_prop:
if counter >= N_PAIRS:
break
images = set(d.image_id for d in by_prop[key])
if len(images) <= 3:
continue
images = list(images)
np.random.shuffle(images)
images = images[:2]
print >> id_f, ",".join(images) + ",0"
counter += 1
with open("experiments/one_different/abstract.ids.txt", "w") as id_f:
counter = 0
for key1 in by_prop:
if counter >= N_PAIRS:
break
keys = list(by_prop.keys())
np.random.shuffle(keys)
for key2 in by_prop:
if len(key1) != len(key2):
continue
if len(set(key1) ^ set(key2)) != 2:
continue
if key1 > key2:
continue
images = [by_prop[key1][0].image_id, by_prop[key2][0].image_id]
print >> id_f, ",".join(images) + ",2"
counter += 1
break
with open("experiments/by_similarity/abstract.ids.txt", "w") as id_f:
keys = list(by_prop.keys())
counter = 0
for key1 in by_prop:
if counter >= N_PAIRS:
break
attempts = 0
while attempts < 100:
attempts += 1
key2 = keys[np.random.randint(len(keys))]
similarity = len(set(key1) ^ set(key2))
if similarity > MAX_DIFFERENCES:
continue
if np.random.random() > 0.45**similarity:
continue
img1 = by_prop[key1][0].image_id
img2 = by_prop[key2][0].image_id
if img1 == img2:
continue
print >> id_f, "%s,%s,%s" % (img1, img2, similarity)
counter += 1
#print counter
break
RHO = 0.95
EPS = 0.000001
LR = args.LR
CLIP = 10
N_TEST_IMAGES = 100
N_TEST = N_TEST_IMAGES * 10
N_EXPERIMENT_PAIRS = 100
# Not sure about num_scenes??
NUM_SCENES = 10
util.setup_logging(args)
train_scenes, dev_scenes, test_scenes = corpus.load_abstract()
VOCAB_SIZE = len(corpus.WORD_INDEX)
VOCAB_SIZE = len(corpus.WORD_INDEX)
output_size = 1
listener0_model = model.Listener0Model(VOCAB_SIZE, NUM_SCENES, args.hidden_sz, output_size, args.dropout) # need to pass in some parameters
speaker0_model = model.Speaker0Model(VOCAB_SIZE, 100, args.dropout, args.dec)
if torch.cuda.is_available():
listener0_model.cuda()
speaker0_model.cuda()
optimizer_l0 = optim.Adam(listener0_model.parameters(), lr=LR)