def main():
tf.random.set_seed(42)
print('[INFO] loading data...')
(train_x, train_y), (val_x, val_y) = datasets.cifar10.load_data()
train_x, val_x = normalize(train_x, val_x)
train_loader = tf.data.Dataset.from_tensor_slices((train_x, train_y))
train_loader = train_loader.map(preprocess).shuffle(50000).batch(
BATCH_SIZE)
val_loader = tf.data.Dataset.from_tensor_slices((val_x, val_y))
val_loader = train_loader.map(preprocess).batch(BATCH_SIZE)
model = VGG16([32, 32, 3])
criterion = K.losses.CategoricalCrossentropy(from_logits=True)
criterion_save = K.metrics.CategoricalCrossentropy(from_logits=True)
metric = K.metrics.CategoricalAccuracy()
opt = optimizers.Adam(lr=LR)
for e in range(EPOCHS):
for batch_x, batch_y in train_loader:
# [b, 1] => [b]
batch_y = tf.squeeze(batch_y, axis=1)
# [b, 10]
batch_y = tf.one_hot(batch_y, depth=10)
with tf.GradientTape() as tape:
logits = model(batch_x)
loss = criterion(batch_y, logits)
criterion_save.update_state(batch_y, logits)
metric.update_state(batch_y, logits)
grads = tape.gradient(loss, model.trainable_variables)
# MUST clip gradient here or it will disverge
grads = [tf.clip_by_norm(g, 15) for g in grads]
opt.apply_gradients(zip(grads, model.trainable_variables))
train_loss, acc = criterion_save.result().numpy(), metric.result(
).numpy()
criterion_save.reset_states()
metric.reset_states()
for batch_x, batch_y in val_loader:
# [b, 1] => [b]
batch_y = tf.squeeze(batch_y, axis=1)
# [b, 10]
batch_y = tf.one_hot(batch_y, depth=10)
logits = model.predict(batch_x)
criterion_save.update_state(batch_y, logits)
metric.update_state(batch_y, logits)
print(
f'[INFO] Epoch {e+1}/{EPOCHS} - loss: {train_loss:.4f} - accuracy: {acc:.4f} - val loss: {criterion_save.result().numpy():.4f} - val acc: {metric.result().numpy():.4f}'
)
criterion_save.reset_states()
metric.reset_states()
def main():
tf.random.set_seed(22)
print('loading data...')
(x, y), (x_test, y_test) = datasets.cifar10.load_data()
x, x_test = normalize(x, x_test)
print(x.shape, y.shape, x_test.shape, y_test.shape)
train_loader = tf.data.Dataset.from_tensor_slices((x, y))
train_loader = train_loader.map(prepare_cifar).shuffle(50000).batch(256)
test_loader = tf.data.Dataset.from_tensor_slices((x_test, y_test))
test_loader = test_loader.map(prepare_cifar).shuffle(10000).batch(256)
print('done.')
model = VGG16([32, 32, 3])
model.compile(optimizer='Adam',
loss='categorical_crossentropy',
metrics=['accuracy'])
model.build(input_shape=(None, 32, 32, 3))
print("Number of variables in the model : ", len(model.variables))
model.summary()
model.fit(train_loader, epochs=40, validation_data=test_loader, verbose=1)
def main(_):
parser = argparse.ArgumentParser()
parser.add_argument('--action', dest='action', type=str, default='train',
help='actions: train, test')
parser.add_argument('--model_id',type=str)
parser.add_argument('--test_step',type=int,default=1000)
args = parser.parse_args()
if args.action not in ['train', 'test']:
print('invalid action: ', args.action)
print("Please input a action: train, test")
else:
model = VGG16(configure(args))
getattr(model, args.action)()
img_size),
batch_size=batch_size,
class_mode='categorical')
validation_generator = test_datagen.flow_from_directory(
args["val_dir"],
target_size=(img_size, img_size),
batch_size=batch_size,
class_mode='categorical')
##### Step-3:
############ Create VGG-16 network graph without the last layers and load imagenet pretrained weights
############ Default image size is 160
print('loading the model and the pre-trained weights...')
base_model = VGG16.VGG16(include_top=False, weights='imagenet')
## Here we will print the layers in the network
i = 0
for layer in base_model.layers:
layer.trainable = False
i = i + 1
print(i, layer.name)
#sys.exit()
##### Step-4:
############ Add the top as per number of classes in our dataset
############ Note that we are using Dropout layer with value of 0.2, i.e. we are discarding 20% weights
############
x = base_model.output
x = Dense(128)(x)
def main():
tf.random.set_seed(22)
print('loading data...')
(x, y), (x_test, y_test) = datasets.cifar10.load_data()
x, x_test = normalize(x, x_test)
print(x.shape, y.shape, x_test.shape, y_test.shape)
# x = tf.convert_to_tensor(x)
# y = tf.convert_to_tensor(y)
train_loader = tf.data.Dataset.from_tensor_slices((x, y))
train_loader = train_loader.map(prepare_cifar).shuffle(50000).batch(256)
test_loader = tf.data.Dataset.from_tensor_slices((x_test, y_test))
test_loader = test_loader.map(prepare_cifar).shuffle(10000).batch(256)
print('done.')
model = VGG16([32, 32, 3])
# must specify from_logits=True!
criteon = keras.losses.CategoricalCrossentropy(from_logits=True)
metric = keras.metrics.CategoricalAccuracy()
optimizer = optimizers.Adam(learning_rate=0.0001)
for epoch in range(250):
for step, (x, y) in enumerate(train_loader):
# [b, 1] => [b]
y = tf.squeeze(y, axis=1)
# [b, 10]
y = tf.one_hot(y, depth=10)
with tf.GradientTape() as tape:
logits = model(x)
loss = criteon(y, logits)
# loss2 = compute_loss(logits, tf.argmax(y, axis=1))
# mse_loss = tf.reduce_sum(tf.square(y-logits))
# print(y.shape, logits.shape)
metric.update_state(y, logits)
grads = tape.gradient(loss, model.trainable_variables)
# MUST clip gradient here or it will disconverge!
grads = [tf.clip_by_norm(g, 15) for g in grads]
optimizer.apply_gradients(zip(grads, model.trainable_variables))
if step % 40 == 0:
# for g in grads:
# print(tf.norm(g).numpy())
print(epoch, step, 'loss:', float(loss), 'acc:',
metric.result().numpy())
metric.reset_states()
if epoch % 1 == 0:
metric = keras.metrics.CategoricalAccuracy()
for x, y in test_loader:
# [b, 1] => [b]
y = tf.squeeze(y, axis=1)
# [b, 10]
y = tf.one_hot(y, depth=10)
logits = model.predict(x)
# be careful, these functions can accept y as [b] without warnning.
metric.update_state(y, logits)
print('test acc:', metric.result().numpy())
metric.reset_states()
import argparse
ap = argparse.ArgumentParser()
ap.add_argument("-image",
"--image",
type=str,
default='test.jpg',
help="Path of test image")
ap.add_argument("-num_class",
"--class",
type=int,
default=2,
help="(required) number of classes to be trained")
args = vars(ap.parse_args())
base_model = VGG16.VGG16(include_top=False, weights=None)
x = base_model.output
x = Dense(128)(x)
x = GlobalAveragePooling2D()(x)
predictions = Dense(args["class"], activation='softmax')(x)
model = Model(inputs=base_model.input, outputs=predictions)
model.load_weights("cv-tricks_fine_tuned_model.h5")
inputShape = (224, 224) # Assumes 3 channel image
image = load_img(args["image"], target_size=inputShape)
image = img_to_array(image) # shape is (224,224,3)
image = np.expand_dims(image, axis=0) # Now shape is (1,224,224,3)
image = image / 255.0
def view_results():
base_model = VGG16.VGG16(include_top=False, weights=None)
x = base_model.output
x = Dense(128)(x)
x = GlobalAveragePooling2D()(x)
predictions = Dense(101, activation='softmax')(x)
model = Model(inputs=base_model.input, outputs=predictions)
model.load_weights("cv-tricks_fine_tuned_model.h5")
input_shape = (224, 224) # Assumes 3 channel image
image = load_img(os.path.join(app.config['FOOD_FOLDER'], 'test_image.jpg'),
target_size=input_shape)
image = img_to_array(image) # shape is (224,224,3)
image = np.expand_dims(image, axis=0) # Now shape is (1,224,224,3)
image = image / 255.0
preds = model.predict(image)
food_prob_array = preds[0]
food_prob_list = food_prob_array.tolist()
k.clear_session()
owd = os.getcwd()
os.chdir('D:\\Projects\\Python\\FineTunedImageRecognition\\training_data')
all_subdir = [d for d in os.listdir('.') if os.path.isdir(d)]
os.chdir(owd)
food_dict = dict(zip(all_subdir, food_prob_list))
sorted_d = sorted(food_dict.items(),
key=lambda food_dict: food_dict[1],
reverse=True)
print(sorted_d)
dish_name = string.capwords(str(sorted_d[0][0]).replace("_", " "))
prob_dish = sorted_d[0][1]
if prob_dish < app.config['FOOD_PROBABILITY_THRESHOLD']:
error_message = "The image was not identified as any Dish. Please capture it again by clicking 'Take A Snapshot Again' button"
return render_template('noResults.html', error_message=error_message)
else:
print("The dish was identified to be {} with {} probability".format(
dish_name, prob_dish))
nutrients, error = get_nutrients_data_from_usda(dish_name)
if nutrients is not None:
weight = request.args.get('weight')
print("*********** " + weight + " ****************")
if not weight:
weight = 100
weight_multiplier = float(weight) / 100
print(nutrients)
for nutrient in nutrients:
nutrient['value'] = str(
round(float(nutrient['value']) * weight_multiplier, 2))
if error is not None:
print(error)
could_also_be_dish_name = sorted_d[1][0]
could_also_be_dish_name_display_name = string.capwords(
could_also_be_dish_name.replace("_", " "))
# print("The dish could also be {} or {} ".format(sorted_d[1][0], sorted_d[2][0]))
return render_template(
'viewResults.html',
dish_name=dish_name,
nutrients=nutrients,
display_second=could_also_be_dish_name_display_name,
second_dish=could_also_be_dish_name,
error=error,
weight=weight)
def main():
tf.random.set_seed(22)
print('loading data...')
(x, y), (x_test, y_test) = datasets.cifar10.load_data()
x, x_test = normalize(x, x_test)
print(x.shape, y.shape, x_test.shape, y_test.shape)
# x = tf.convert_to_tensor(x)
# y = tf.convert_to_tensor(y)
train_loader = tf.data.Dataset.from_tensor_slices((x, y))
train_loader = train_loader.map(prepare_cifar).shuffle(50000).batch(1024)
test_loader = tf.data.Dataset.from_tensor_slices((x_test, y_test))
test_loader = test_loader.map(prepare_cifar).shuffle(10000).batch(256)
print('done.')
model = VGG16([32, 32, 3])
criteon = keras.losses.CategoricalCrossentropy(from_logits=True)
metric = keras.metrics.CategoricalAccuracy()
optimizer = optimizers.Adam(learning_rate=0.0001)
ckpt = tf.train.Checkpoint(step=tf.Variable(1),
optimizer=optimizer,
net=model)
manager = tf.train.CheckpointManager(ckpt, './tf_ckpts', max_to_keep=3)
ckpt.restore(manager.latest_checkpoint)
if manager.latest_checkpoint:
print("Restored from {}".format(manager.latest_checkpoint))
else:
print("Initializing from scratch.")
for epoch in range(2500):
for step, (x, y) in enumerate(train_loader):
y = tf.squeeze(y, axis=1)
y = tf.one_hot(y, depth=10)
with tf.GradientTape() as tape:
logits = model(x)
loss = criteon(y, logits)
metric.update_state(y, logits)
grads = tape.gradient(loss, model.trainable_variables)
# MUST clip gradient here or it will disconverge!
grads = [tf.clip_by_norm(g, 15) for g in grads]
optimizer.apply_gradients(zip(grads, model.trainable_variables))
if step % 40 == 0:
print(epoch, step, 'loss:', float(loss), 'acc:',
metric.result().numpy())
metric.reset_states()
if epoch % 1 == 0:
model.save_weights('easy_checkpoint')
save_path = manager.save()
print("Saved checkpoint for step {}: {}".format(
int(ckpt.step), save_path))
metric = keras.metrics.CategoricalAccuracy()
for x, y in test_loader:
# [b, 1] => [b]
y = tf.squeeze(y, axis=1)
# [b, 10]
y = tf.one_hot(y, depth=10)
logits = model.predict(x)
# be careful, these functions can accept y as [b] without warnning.
metric.update_state(y, logits)
print('test acc:', metric.result().numpy())
metric.reset_states()
def main(args):
parser = argparse.ArgumentParser()
parser.add_argument('run_name', metavar='N', type=str, help='name of run')
parser.add_argument('network_type',
metavar='N',
type=str,
help='name of run')
parser.add_argument('gpu_id',
metavar='G',
type=str,
help='which gpu to use')
parser.add_argument('--print_every',
metavar='N',
type=int,
help='number of iterations before printing',
default=-1)
parser.add_argument('--print_network',
action='store_true',
help='print_network for debugging')
parser.add_argument('--data_parallel',
type=int,
nargs='+',
default=None,
help='paralellize across multiple gpus')
parser.add_argument('--test', action='store_true', help='test')
parser.add_argument('--test_print', action='store_true', help='test')
parser.add_argument('--valid_iters',
metavar='I',
type=int,
default=100,
help='number of validation iters to run every epoch')
parser.add_argument('--csv_file',
metavar='CSV',
type=str,
default=None,
help='name of csv file to write to')
parser.add_argument(
'--sweep_lambda',
action='store_true',
help=
'preform a sweep over lambda values keeping other settings as in args')
parser.add_argument(
'--sweep_c',
action='store_true',
help=
'preform a sweep over lambda values keeping other settings as in args')
parser.add_argument('--sweep_start',
metavar='S',
type=float,
default=0.0,
help='lambda value to start sweep at')
parser.add_argument('--sweep_stop',
metavar='E',
type=float,
default=0.1,
help='lambda value to stop sweep at')
parser.add_argument('--sweep_step',
metavar='E',
type=float,
default=0.01,
help='step_size_between_sweep_points')
parser.add_argument('--sweep_exp',
action='store_true',
help='step_size_between_sweep_points')
parser.add_argument('--sweep_resume',
action='store_true',
help='resume sweep checkpts')
parser.add_argument(
'--sweep_con_runs',
metavar='C',
type=int,
default=1,
help='number of runs to run with same parameters to validate constiancy'
)
#checkpoints
parser.add_argument('--checkpoint_every',
type=int,
default=10,
help='checkpoint every n epochs')
parser.add_argument('--load_checkpoint',
action='store_true',
help='load checkpoint with same name')
parser.add_argument('--resume',
action='store_true',
help='resume from epoch we left off of when loading')
parser.add_argument('--checkpoint',
type=str,
default=None,
help='checkpoint to load')
#params
parser.add_argument('--epochs',
metavar='N',
type=int,
help='number of epochs to run for',
default=50)
parser.add_argument('--batch_size',
metavar='bs',
type=int,
default=1024,
help='batch size')
parser.add_argument('--lr',
metavar='lr',
type=float,
help='learning rate',
default=1e-3)
parser.add_argument('--rmsprop',
action='store_true',
help='use rmsprop optimizer')
parser.add_argument('--sgd', action='store_true', help='use sgd optimizer')
parser.add_argument('--lr_reduce_on_plateau',
action='store_true',
help='update optimizer on plateau')
parser.add_argument('--lr_exp',
action='store_true',
help='update optimizer on plateau')
parser.add_argument('--lr_step',
type=int,
nargs='+',
default=None,
help='decrease lr by gamma = 0.1 on these epochs')
parser.add_argument('--lr_list',
type=float,
nargs='+',
default=None,
help='decrease lr by gamma = 0.1 on these epochs')
parser.add_argument('--l2_reg', type=float, default=0.0)
DataLoader.add_args(parser)
#added so default argument come from the network that is loaded
network_type = args[1]
if network_type == 'auto_fc':
AutoFCNetwork.add_args(parser)
network_class = AutoFCNetwork
elif network_type == 'auto_conv':
AutoConvNetwork.add_args(parser)
network_class = AutoConvNetwork
elif network_type == 'class_conv':
ClassifyConvNetwork.add_args(parser)
network_class = ClassifyConvNetwork
elif network_type == 'vgg':
VGG16.add_args(parser)
network_class = VGG16
elif network_type == 'res':
ResidualConvNetwork.add_args(parser)
network_class = ResidualConvNetwork
elif network_type == 'res152':
ResNet152.add_args(parser)
network_class = ResNet152
elif network_type == 'dense':
DenseNet.add_args(parser)
network_class = DenseNet
else:
raise ValueError('unknown network type' + str(network_type))
args = parser.parse_args(args)
#***************
# GPU
#***************
if args.data_parallel is not None:
try:
del os.environ['CUDA_VISIBLE_DEVICES']
except KeyError:
pass
device = torch.device('cuda:%d' % args.data_parallel[0])
elif args.gpu_id == '-1':
os.environ['CUDA_VISIBLE_DEVICES'] = ''
device = torch.device('cpu')
else:
print(bcolors.OKBLUE + 'Using GPU' + str(args.gpu_id) + bcolors.ENDC)
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_id
device = torch.device('cuda')
#***************
# Run
#***************
if args.sweep_lambda or args.sweep_c:
run_dir = args.run_name
run_ckpt_dir = 'ckpts/%s' % run_dir
if not os.path.isdir(run_ckpt_dir):
os.mkdir(run_ckpt_dir)
val_l = []
loss_l = []
op_loss_l = []
reg_loss_l = []
nodes_l = []
images_l = []
val = args.sweep_start
while val <= args.sweep_stop:
print(bcolors.OKBLUE + 'Val: %0.1E' % val + bcolors.ENDC)
args.run_name = run_name = "%s/l%0.1E" % (run_dir, val)
for i in range(args.sweep_con_runs):
if args.sweep_con_runs > 0:
args.run_name = run_name + '_' + str(i)
if not (args.sweep_resume
and os.path.isdir('ckpts/' + args.run_name)):
print(bcolors.OKBLUE + 'Run: %s' % args.run_name +
bcolors.ENDC)
if args.sweep_lambda:
args.reg_lambda = float(val)
elif args.sweep_c:
args.reg_c = float(val)
run_wrapper = RunWrapper(args, network_class, device)
if not args.test:
run_wrapper.train()
loss, op_loss, reg_loss, rem_nodes, acc = run_wrapper.test(
load=args.test)
if isinstance(rem_nodes, list):
rem_nodes = sum(rem_nodes)
x, y_hat = run_wrapper.test_print(plot=False, load=False)
val_l.append(val)
loss_l.append(loss)
op_loss_l.append(op_loss)
reg_loss_l.append(reg_loss)
nodes_l.append(rem_nodes)
if images_l == []:
images_l.append(x[:17])
images_l.append(y_hat[:17])
#hopefully this cleans up the gpu memory
del run_wrapper
if args.sweep_exp:
if val == 0.0:
val = args.sweep_start
else:
val = val * args.sweep_step
else:
val = val + args.sweep_step
loss_l = np.array(loss_l)
op_loss_l = np.array(op_loss_l)
reg_loss_l = np.array(reg_loss_l)
nodes_l = np.array(nodes_l)
if args.sweep_lambda:
lc_l = val_l #[l * args.reg_C for l in val_l]
elif args.sweep_c:
lc_l = [c * args.reg_lambda for c in val_l]
print(lc_l)
print(nodes_l)
misc.plot_sweep(run_ckpt_dir, lc_l, op_loss_l, nodes_l)
#misc.sweep_to_image(images_l, run_ckpt_dir)
else:
#default single run behavior
run_wrapper = RunWrapper(args, network_class, device)
if args.test_print:
run_wrapper.test_print()
elif args.test:
run_wrapper.test()
else:
run_wrapper.train()
def main():
args = parser.parse_args()
tf.random.set_seed(22)
print('loading data...')
(x, y), (x_test_ori, y_test) = datasets.cifar10.load_data()
X_train = x / 255.
X_test_normal = x_test_ori / 255.
mean = np.mean(X_train, axis=(0, 1, 2, 3))
std = np.std(X_train, axis=(0, 1, 2, 3))
print('mean:', mean, 'std:', std)
model = VGG16([32, 32, 3])
model.load_weights('easy_checkpoint')
test_imge_dir = args.imagedir
print(test_imge_dir)
for root, dirs, files in os.walk(test_imge_dir, topdown=False):
for name in files:
file_name = os.path.join(root, name)
print(file_name)
test_imge = cv2.imread(file_name)
test_imge = test_imge / 255.
test_imge_one = (test_imge - mean) / (std + 1e-7)
input_x = np.expand_dims(test_imge_one, 0)
logits = model.predict(input_x)
result_index = np.argmax(logits)
print('predict value: ' + str(cifar10_labels[result_index]) + '\n')
exit(0)
img_index_all = 0
img_index_err = 0
x_test_dataset = (X_test_normal - mean) / (std + 1e-7)
test_loader = tf.data.Dataset.from_tensor_slices((x_test_dataset, y_test))
test_loader = test_loader.map(prepare_cifar).shuffle(10000).batch(1)
for x_test_input, y in test_loader:
# [b, 1] => [b]
y = tf.squeeze(y, axis=1)
print('true label: ' + str(cifar10_labels[y[0]]))
# # [b, 10]
# y = tf.one_hot(y, depth=10)
logits = model.predict(x_test_input)
result_index = np.argmax(logits)
print('predict label: ' + str(cifar10_labels[result_index]) + '\n')
if result_index != y[0]:
x_test_ori = (x_test_input * (std + 1e-7) + mean) * 255
cv2.imwrite(
'error_result/{}_true_label_{}_predict_{}.jpg'.format(
img_index_err, cifar10_labels[y[0]],
cifar10_labels[result_index]), x_test_ori[0].numpy())
img_index_err = img_index_err + 1
else:
x_test_ori = (x_test_input * (std + 1e-7) + mean) * 255
cv2.imwrite(
'good_result/{}_true_label_{}_predict_{}.jpg'.format(
img_index_err, cifar10_labels[y[0]],
cifar10_labels[result_index]), x_test_ori[0].numpy())
img_index_all = img_index_all + 1
print('所有图片数量 : {} 预测错误数量: {}'.format(img_index_all, img_index_err))
exit(0)
