def test_embeddings():
model = Siamese('dummy', input_shape=(6, 8, 3), embedding_size=3)
model.make_embeddings('data_tiny/train',
'data_tiny/train.csv',
batch_size=1)
emb = pd.read_pickle(os.path.join(model.cache_dir, 'embeddings.pkl'))
print(emb)
model.predict('data_tiny/train')
pred = pd.read_pickle(os.path.join(model.cache_dir, 'predictions.pkl'))
print(pred)
import sys
sys.path.insert(0, '../')
from core.siamese import Siamese
model = Siamese('mobilenet_like', input_shape=(672, 896, 3), embedding_size=128, train_hidden_layers=True)
#model.load_weights('../trained/final_weights.h5')
model.load_weights('trained/final_weights.h5')
model.train('../data/train', 'train.csv', meta_dir='../data/meta', epochs=40, batch_size=20, learning_rate=0.001, margin=1.0)
from core.siamese import Siamese
model = Siamese('mobilenet_like',
input_shape=(672, 896, 3),
embedding_size=128,
strategy='batch_all')
# model.load_weights('cache/cache-190208-093118/training/checkpoint-03.h5')
# TODO mark pretrained mobilenet
model.load_weights('cache/cache-190208-093118/training/checkpoint-03.h5')
model.train('data/train.csv',
'data/train',
epochs=20,
batch_size=32,
learning_rate=0.0001,
margin=2.0)
from core.siamese import Siamese
# model = Siamese('shallow_mnist', input_shape=(28, 28, 3), embedding_size=64, strategy='batch_all')
# model.load_weights('cache/cache-190213-131256/training/checkpoint-07.h5')
model = Siamese.restore_from_config('cache/cache-190220-203237/config.json')
model.make_embeddings('data_mnist/train', 'data_mnist/train.csv', batch_size=200)
model.predict('data_mnist/train_subset')
model.make_csv('cache/cache-190213-131256/idx_to_whales_mapping.npy')
# model.load_embeddings('cache/cache-190205-070856/embeddings.pkl')
# model.load_predictions('cache/cache-190205-072026/predictions.pkl')
# model.make_kaggle_csv('cache/cache-190205-065005/idx_to_whales_mapping.npy')
# model.draw_tsne(model.predictions.values[:, 1:])
from core.siamese import Siamese
model = Siamese(input_shape=(224, 224, 3), n_classes=5004)
model.load_weights('data/new_whalizer.h5')
model.predict_new_whales('data/test', 'data/submission_no_new_whales.csv')
from core.siamese import Siamese
model = Siamese('mobilenet_like',
input_shape=(672, 896, 3),
embedding_size=128,
strategy='batch_hard')
model.load_weights('cache/cache-190213-095205/training/checkpoint-07.h5')
model.load_embeddings('cache/cache-190213-113426/embeddings.pkl')
# model.make_embeddings('data/train.csv', 'data/train', batch_size=32)
# model.predict('data/test')
model.predict('data/train_subset')
model.make_csv('cache/cache-190208-093118/idx_to_whales_mapping.npy')
from core.siamese import Siamese
model = Siamese('shallow_mnist',
input_shape=(28, 28, 3),
embedding_size=64,
strategy='batch_all')
#model.train('data_mnist/train.csv', 'data_mnist/train', 'C:/Users/Sergei/Desktop/humpback-whale-identification/data_mnist/meta',
# epochs=10, batch_size=100, learning_rate=0.0001, margin=1.0)
model.train(
'data_mnist/train',
'data_mnist/train.csv',
'C:/Users/Sergei/Desktop/humpback-whale-identification/data_mnist/meta',
epochs=10,
batch_size=100,
learning_rate=0.0001,
margin=0.5)
from core.siamese import Siamese
model = Siamese('mobilenet_like',
input_shape=(672, 896, 3),
embedding_size=128)
model.load_weights('trained/final_weights.h5')
#model.make_embeddings('data/train', 'data/train.csv', batch_size=5, meta_dir='data/meta')
model.load_embeddings('trained/embeddings.pkl')
model.predict('data/test', meta_dir='data/meta')
#model.load_predictions('trained/predictions.pkl')
model.make_kaggle_csv('data/meta/idx_to_whales_mapping.npy')
from core.siamese import Siamese
model = Siamese('mobilenet_like',
input_shape=(672, 896, 3),
embedding_size=128)
model.load_weights('model/mobilenet_imagenet.h5')
#model.load_weights('trained/final_weights.h5')
#model.train('data/train', 'data/split_train.csv', meta_dir='D:\\IdeaProjects\\whales\\data\\meta', epochs=20, batch_size=24, learning_rate=0.001)
model.train('data/train',
'data/train.csv',
meta_dir='data/meta',
epochs=500,
batch_size=25,
learning_rate=0.0005,
margin=1.0)
from utils.sequence import WhalesSequence
from core.siamese import Siamese
csv = 'val.csv'
#csv = 'train.csv'
mode = 'cos_angular'
#mode = 'classification'
input_shape = (224, 224, 3)
img_dir = '../data/train'
val = pd.read_csv(csv)
true_labels = val["Id"].values
if mode == 'classification':
model = Siamese(input_shape=(224, 224, 3),
train_hidden_layers=True,
n_classes=8)
model.load_weights('final_weights.h5')
img_names = val['Image'].values
bboxes = pd.read_pickle('../data/meta/bboxes.pkl').set_index('filename')
whales_seq = WhalesSequence(img_dir,
bboxes=bboxes,
input_shape=input_shape,
x_set=img_names,
batch_size=1)
pred = model.model.predict_generator(whales_seq, verbose=1)
pred_labels = np.argmax(pred, axis=1).reshape(-1)
for i, whale in enumerate(np.sort(np.unique(true_labels))):
true_labels[np.where(true_labels == whale)] = i
import sys
sys.path.insert(0, '../')
from core.siamese import Siamese
model = Siamese('mobilenet_like', input_shape=(672, 896, 3), embedding_size=128)
model.load_weights('trained/final_weights.h5')
model.make_embeddings('../data/train', 'train.csv', mappings_filename='../data/meta/whales_to_idx_mapping.npy', batch_size=25)
#model.load_embeddings('trained/embeddings.pkl')
model.predict('../data/train', 'val.csv')
#model.predict('../data/train', 'train.csv')
from core.siamese import Siamese
model = Siamese('resnet_like_33', input_shape=(384, 512, 3), embedding_size=128, strategy='batch_all')
model.load_weights('trained/checkpoint-05.h5')
model.load_embeddings('trained/embeddings.pkl')
model.load_predictions('trained/predictions.pkl')
model.make_kaggle_csv('trained/idx_to_whales_mapping.npy')
from core.siamese import Siamese
model = Siamese('shallow_mnist',
input_shape=(28, 28, 3),
embedding_size=64,
strategy='batch_all')
model.train('data_mnist/train.csv',
'data_mnist/train',
epochs=10,
batch_size=100,
learning_rate=0.0001,
margin=1.0)
from core.siamese import Siamese
# using softmax loss classification
# model = Siamese(input_shape=(224, 224, 3), n_classes=5004)
# model.load_weights('trained/final_weights.h5')
# model.predict('data/test')
# model.make_kaggle_csv('data/meta/idx_to_whales_mapping.npy')
# using cos_angular embeddings
model = Siamese(input_shape=(224, 224, 3),
n_classes=5004,
mode='cosface',
train=False)
model.load_weights('trained/final_weights.h5')
model.make_embeddings('data/train',
'data/train.csv',
batch_size=25,
meta_dir='data/meta')
#model.load_embeddings('trained/embeddings.pkl')
model.predict_using_embeddings('data/test', meta_dir='data/meta')
#model.load_predictions('trained/predictions.pkl')
model.make_kaggle_csv('data/meta/idx_to_whales_mapping.npy')
from core.siamese import Siamese
model = Siamese.restore_from_config('cache/cache-190226-111833/config.json')
model.make_embeddings('data/train', 'data/train.csv', batch_size=24)
# model.predict('data/train')
# model.make_csv('cache/cache-190225-233850/idx_to_whales_mapping.npy')
model.predict('data/test')
model.make_kaggle_csv('data/meta/idx_to_whales_mapping.npy')
import sys
sys.path.insert(0, '../')
from core.siamese import Siamese
#model = Siamese(input_shape=(224, 224, 3), train_hidden_layers=False, n_classes=8, mode='classification')
model = Siamese(input_shape=(224, 224, 3),
train_hidden_layers=True,
n_classes=8,
mode='arcface')
#model.load_weights('../trained/final_weights.h5')
model.load_weights('trained/final_weights.h5')
model.train('../data/train',
'train.csv',
meta_dir='../data/meta',
epochs=70,
batch_size=5,
learning_rate=0.0001)
# inference to get cos angle embeddings
import sys
sys.path.insert(0, '../')
from core.siamese import Siamese
model = Siamese(input_shape=(224, 224, 3),
n_classes=8,
mode='arcface',
train=False)
model.load_weights('trained/final_weights.h5')
model.make_embeddings(
'../data/train',
'train.csv',
mappings_filename='../data/meta/whales_to_idx_mapping.npy',
batch_size=25)
#model.load_embeddings('trained/embeddings.pkl')
model.predict_using_embeddings('../data/train', 'val.csv')
#model.predict_using_embeddings('../data/train', 'train.csv')
from core.siamese import Siamese
model = Siamese('shallow_mnist', input_shape=(28, 28, 3), embedding_size=64, strategy='batch_all')
model.load_weights('cache/cache-190213-131256/training/checkpoint-07.h5')
model.make_embeddings('data_mnist/train.csv', 'data_mnist/train', batch_size=200)
model.predict('data_mnist/train_subset')
model.make_csv('cache/cache-190213-131256/idx_to_whales_mapping.npy')
# model.load_embeddings('cache/cache-190205-070856/embeddings.pkl')
# model.load_predictions('cache/cache-190205-072026/predictions.pkl')
# model.make_kaggle_csv('cache/cache-190205-065005/idx_to_whales_mapping.npy')
# model.draw_tsne(model.predictions.values[:, 1:])
from core.siamese import Siamese
model = Siamese('resnet_like_33', input_shape=(384, 512, 3), embedding_size=128, strategy='batch_all')
model.load_weights('trained/checkpoint-05.h5')
model.load_embeddings('trained/embeddings.pkl')
model.load_predictions('trained/predictions.pkl')
model.draw_tsne(model.predictions.values[:, 1:])