#print("Current TestLoss: {}".format(loss))
return classification
if __name__ == '__main__':
tf.compat.v1.enable_eager_execution()
gpu = tf.config.experimental.list_physical_devices('GPU')
print("Num GPUs Available: ", len(gpu))
if len(gpu) > 0:
tf.config.experimental.set_memory_growth(gpu[0], True)
tf.config.experimental.set_memory_growth(gpu[1], True)
# read dataset
path_root = os.path.abspath(os.path.dirname(__file__))
bird_data = DataSet("/Volumes/Watermelon") # DataSet(path_root)
phi_train = bird_data.get_phi(set=0)
w = bird_data.get_w(alpha=1) # (50*150)
train_class_list, test_class_list = bird_data.get_class_split(mode="easy")
train_ds, test_ds = bird_data.load_gpu(batch_size=BATCH_SIZE)
#path_root = os.path.abspath(os.path.dirname(__file__))
#database = DataSet("/Volumes/Watermelon") # path_root)
#PHI = database.get_phi()
#DS, DS_test = database.load_gpu(batch_size=5) # image_batch, label_batch
modelaki = FinalModel()
# define loss and opt functions
loss_fun = Loss().final_loss
step = tf.Variable(0, trainable=False)
boundaries = [187 * 5, 187 * 10]
values = [0.05, 0.005, 0.0005]
from Losses import Loss
import sys
sys.path.append("../src")
from jointmodel import JFL
CHANNELS = 512
N_CLASSES = 200
SEMANTIC_SIZE = 28
BATCH_SIZE = 5
IMG_SIZE = 448
IMG_SHAPE = (IMG_SIZE, IMG_SIZE, 3)
# read dataset
database = DataSet("/Volumes/Watermelon")
PHI = database.get_phi()
DS, DS_test = database.load_gpu(batch_size=BATCH_SIZE)
tf.compat.v1.enable_eager_execution()
strategy = tf.distribute.MirroredStrategy()
print('Number of devices: {}'.format(strategy.num_replicas_in_sync))
BUFFER_SIZE = 5
BATCH_SIZE_PER_REPLICA = 32
GLOBAL_BATCH_SIZE = BATCH_SIZE_PER_REPLICA * strategy.num_replicas_in_sync
EPOCHS = 30
train_dataset = DS
train_dist_dataset = strategy.experimental_distribute_dataset(train_dataset)
# sum (and normalize?) the scores and mapped features
score = tf.add(
tf.add(global_theta, local_scores0),
local_scores1) #tf.math.reduce_sum(, axis=1, keepdims=True)
# avg_score = tf.multiply(sum_gll, 1.0 / 3.0)
# compute ||~phi_i - ~Ci|| and ||~phi_i - ~Cj||, '~' is normalization
# l2loss = self.l2loss(
# [tf.math.l2_normalize(tf.squeeze(out)), tf.math.l2_normalize(tf.transpose(center, perm=[0, 2, 1]))])
phi_mapped = tf.add(tf.add(global_phi, local0_phi), local1_phi)
#avg_phi = tf.multiply(sum_gll, 1.0 / 3.0)
y_pred = self.classifier(score)
return m0, m1, mask0, mask1, score, phi_mapped, y_pred, self.C
if __name__ == '__main__':
# just for testing
path_root = os.path.abspath(
os.path.dirname(__file__)) # '/content/gdrive/My Drive/data'
bird_data = DataSet("D:/MY2/ADDL/DD2412_project/basemodel")
PHI = bird_data.get_phi(set=0)
#w = bird_data.get_w(alpha=1) # (50*150)
#train_class_list, test_class_list = bird_data.get_class_split(mode="easy")
# only take 1000 images for local test
train_ds = bird_data.load(GPU=False, train=True, batch_size=4)
# test_ds = bird_data.load(GPU=False, train=False, batch_size=32)
image_batch, label_batch = next(iter(train_ds))
test_model = FinalModel()
m0, m1, mask0, mask1, scores, phi, y_pred, C = test_model(image_batch, PHI)
momentum=0.9)
# MODEL
#net = FinalModel()
#new_root = tf.train.Checkpoint(net=net)
#status = new_root.restore(tf.train.latest_checkpoint('./tf_ckpts/'))
net = FinalModel()
ckpt = tf.train.Checkpoint(step=tf.Variable(1, dtype=tf.int32),
optimizer=opt,
net=net)
ckpt.restore(tf.train.latest_checkpoint('./tf_ckpts/'))
#DATA
path_root = os.path.abspath(os.path.dirname(__file__))
bird_data = DataSet("/Volumes/Watermelon") # DataSet(path_root)
phi_train = bird_data.get_phi(set=0)
w = bird_data.get_w(alpha=1) # (50*150)
train_class_list, test_class_list = bird_data.get_class_split(mode="easy")
train_ds = bird_data.load(GPU=False, train=True, batch_size=32)
#test_ds = bird_data.load(GPU=False, train=False, batch_size=4) #.load_gpu(batch_size=4)
PHI = bird_data.get_phi(set=0)
for im, label in train_ds:
#im_path = "/Volumes/Watermelon/CUB_200_2011/CUB_200_2011/images/059.California_Gull/"
#img = tf.io.read_file(im_path)
#im = database.decode_img(img)
m0, m1, mask0, mask1, scores, phi, y_pred, C = net(
im, PHI) #tf.expand_dims(im,0)
nu = 50
ns = 150
W = tf.ones((nu, ns))
return tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(labels=labels, logits=score))
if __name__ == '__main__':
#os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
#import tensorflow as tf
tf.compat.v1.enable_eager_execution()
gpu = tf.config.experimental.list_physical_devices('GPU')
print("Num GPUs Available: ", len(gpu))
if gpu:
tf.config.experimental.set_memory_growth(gpu[0], True)
path_root = os.path.abspath(os.path.dirname(__file__)) # '/content/gdrive/My Drive/data'
bird_data = DataSet(path_root)
# load all imgs
phi = bird_data.get_phi()
train_ds, test_ds = bird_data.load_gpu(batch_size=BATCH_SIZE)
# only take 1000 images for local test
#train_ds = bird_data.load(GPU=False, train=True, batch_size=32)
#test_ds = bird_data.load(GPU=False, train=False, batch_size=32)
#image_batch, label_batch = next(iter(train_ds))
model = BaseModel(150, 312)
#opt = tf.keras.optimizers.SGD(learning_rate=0.001, momentum=0.9) # or SGDW with weight decay
opt = tfa.optimizers.SGDW(
learning_rate=0.0001, weight_decay=5 * 1e-4, momentum=0.9)
ckpt = tf.train.Checkpoint(step=tf.Variable(1), optimizer=opt, net=model)
manager = tf.train.CheckpointManager(ckpt, path_root + '/tf_ckpts',
max_to_keep=3) # keep only the three most recent checkpoints