def prep_newshapes_train(init_with = "last", FCN_layers= False, batch_sz =5, epoch_steps = 4, folder_name= "mrcnn_newshape_training_logs"):
import mrcnn.new_shapes as new_shapes
MODEL_DIR = os.path.join(MODEL_PATH, folder_name)
# Build configuration object -----------------------------------------------
config = new_shapes.NewShapesConfig()
config.BATCH_SIZE = batch_sz # Batch size is 2 (# GPUs * images/GPU).
config.IMAGES_PER_GPU = batch_sz # Must match BATCH_SIZE
config.STEPS_PER_EPOCH = epoch_steps
config.FCN_INPUT_SHAPE = config.IMAGE_SHAPE[0:2]
# Build shape dataset -----------------------------------------------
# Training dataset
dataset_train = new_shapes.NewShapesDataset()
dataset_train.load_shapes(3000, config.IMAGE_SHAPE[0], config.IMAGE_SHAPE[1])
dataset_train.prepare()
# Validation dataset
dataset_val = new_shapes.NewShapesDataset()
dataset_val.load_shapes(500, config.IMAGE_SHAPE[0], config.IMAGE_SHAPE[1])
dataset_val.prepare()
try :
del model
print('delete model is successful')
gc.collect()
except:
pass
KB.clear_session()
model = modellib.MaskRCNN(mode="training", config=config, model_dir=MODEL_DIR,FCN_layers = FCN_layers)
print('MODEL_PATH : ', MODEL_PATH)
print('COCO_MODEL_PATH : ', COCO_MODEL_PATH)
print('RESNET_MODEL_PATH : ', RESNET_MODEL_PATH)
print('MODEL_DIR : ', MODEL_DIR)
print('Last Saved Model : ', model.find_last())
load_model(model, init_with = 'last')
train_generator = data_generator(dataset_train, model.config, shuffle=True,
batch_size=model.config.BATCH_SIZE,
augment = False)
config.display()
return [model, dataset_train, train_generator, config]
def prep_newshapes_test(init_with = 'last', FCN_layers = False, batch_sz = 5, epoch_steps = 4,folder_name= "mrcnn_newshape_test_logs"):
import mrcnn.new_shapes as new_shapes
MODEL_DIR = os.path.join(MODEL_PATH, folder_name)
# Build configuration object -----------------------------------------------
config = new_shapes.NewShapesConfig()
config.BATCH_SIZE = batch_sz # Batch size is 2 (# GPUs * images/GPU).
config.IMAGES_PER_GPU = batch_sz # Must match BATCH_SIZE
config.STEPS_PER_EPOCH = epoch_steps
config.FCN_INPUT_SHAPE = config.IMAGE_SHAPE[0:2]
# Build shape dataset -----------------------------------------------
# Training dataset
dataset_test = new_shapes.NewShapesDataset()
dataset_test.load_shapes(3000, config.IMAGE_SHAPE[0], config.IMAGE_SHAPE[1])
dataset_test.prepare()
# Recreate the model in inference mode
try :
del model
print('delete model is successful')
gc.collect()
except:
pass
KB.clear_session()
model = modellib.MaskRCNN(mode="inference",
config=config,
model_dir=MODEL_DIR,
FCN_layers = FCN_layers )
print(' COCO Model Path : ', COCO_MODEL_PATH)
print(' Checkpoint folder Path: ', MODEL_DIR)
print(' Model Parent Path : ', MODEL_PATH)
print(' Resent Model Path : ', RESNET_MODEL_PATH)
load_model(model, init_with = init_with)
test_generator = data_generator(dataset_test, model.config, shuffle=True,
batch_size=model.config.BATCH_SIZE,
augment = False)
model.config.display()
return [model, dataset_test, test_generator, config]
args.batch_size) # Batch size is 2 (# GPUs * images/GPU). config.IMAGES_PER_GPU = int(args.batch_size) # Must match BATCH_SIZE config.STEPS_PER_EPOCH = int(args.steps_in_epoch) config.LEARNING_RATE = float(args.lr) config.EPOCHS_TO_RUN = int(args.epochs) config.FCN_INPUT_SHAPE = config.IMAGE_SHAPE[0:2] config.LAST_EPOCH_RAN = int(args.last_epoch) config.display() ##------------------------------------------------------------------------------------ ## Build shape dataset ##------------------------------------------------------------------------------------ # Training dataset # generate 500 shapes dataset_train = shapes.NewShapesDataset() dataset_train.load_shapes(10000, config.IMAGE_SHAPE[0], config.IMAGE_SHAPE[1]) dataset_train.prepare() # Validation dataset dataset_val = shapes.NewShapesDataset() dataset_val.load_shapes(2500, config.IMAGE_SHAPE[0], config.IMAGE_SHAPE[1]) dataset_val.prepare() ##------------------------------------------------------------------------------------ ## Load and display random samples ##------------------------------------------------------------------------------------ # image_ids = np.random.choice(dataset_train.image_ids, 3) # for image_id in [3]: # image = dataset_train.load_image(image_id) # mask, class_ids = dataset_train.load_mask(image_id)
def prep_newshapes_train(init_with = "last", FCN_layers= False, batch_sz =5, epoch_steps = 4, training_folder= None):
MODEL_DIR = os.path.join(TRAINING_DIR, training_folder)
# Build configuration object -----------------------------------------------
config = new_shapes.NewShapesConfig()
config.BATCH_SIZE = batch_sz # Batch size is 2 (# GPUs * images/GPU).
config.IMAGES_PER_GPU = batch_sz # Must match BATCH_SIZE
config.STEPS_PER_EPOCH = epoch_steps
config.FCN_INPUT_SHAPE = config.IMAGE_SHAPE[0:2]
# Build shape dataset -----------------------------------------------
# Training dataset
dataset_train = new_shapes.NewShapesDataset(config)
dataset_train.load_shapes(10000)
dataset_train.prepare()
# Validation dataset
dataset_val = new_shapes.NewShapesDataset(config)
dataset_val.load_shapes(2500)
dataset_val.prepare()
try :
del model
print('delete model is successful')
gc.collect()
except:
pass
KB.clear_session()
model = modellib.MaskRCNN(mode="training", config=config, model_dir=MODEL_DIR,FCN_layers = FCN_layers)
print('TRAINING_DIR : ', TRAINING_DIR)
print('COCO_TRAINING_DIR : ', COCO_TRAINING_DIR)
print('RESNET_TRAINING_DIR : ', RESNET_TRAINING_DIR)
print('MODEL_DIR : ', MODEL_DIR)
print('Last Saved Model : ', model.find_last())
# exclude_layers = \
# ['fcn_block1_conv1'
# ,'fcn_block1_conv2'
# ,'fcn_block1_pool'
# ,'fcn_block2_conv1'
# ,'fcn_block2_conv2'
# ,'fcn_block2_pool'
# ,'fcn_block3_conv1'
# ,'fcn_block3_conv2'
# ,'fcn_block3_conv3'
# ,'fcn_block3_pool'
# ,'fcn_block4_conv1'
# ,'fcn_block4_conv2'
# ,'fcn_block4_conv3'
# ,'fcn_block4_pool'
# ,'fcn_block5_conv1'
# ,'fcn_block5_conv2'
# ,'fcn_block5_conv3'
# ,'fcn_block5_pool'
# ,'fcn_fc1'
# ,'dropout_1'
# ,'fcn_fc2'
# ,'dropout_2'
# ,'fcn_classify'
# ,'fcn_bilinear'
# ,'fcn_heatmap_norm'
# ,'fcn_scoring'
# ,'fcn_heatmap'
# ,'fcn_norm_loss']
# load_model(model, init_with = 'last', exclude = exclude_layers)
model.load_model_weights(init_with = init_with)
# print('=====================================')
# print(" Load second weight file ?? ")
# model.keras_model.load_weights('E:/Models/vgg16_weights_tf_dim_ordering_tf_kernels.h5', by_name= True)
train_generator = data_generator(dataset_train, model.config, shuffle=True,
batch_size=model.config.BATCH_SIZE,
augment = False)
val_generator = data_generator(dataset_val, model.config, shuffle=True,
batch_size=model.config.BATCH_SIZE,
augment=False)
config.display()
return [model, dataset_train, dataset_val, train_generator, val_generator, config]
def prep_newshapes_train2(init_with = "last", config=None):
import mrcnn.new_shapes as new_shapes
config.CHECKPOINT_FOLDER = os.path.join(TRAINING_DIR, config.CHECKPOINT_FOLDER)
# Build shape dataset -----------------------------------------------
# Training dataset
dataset_train = new_shapes.NewShapesDataset(config)
dataset_train.load_shapes(config.TRAINING_IMAGES)
dataset_train.prepare()
# Validation dataset
dataset_val = new_shapes.NewShapesDataset(config)
dataset_val.load_shapes(config.VALIDATION_IMAGES)
dataset_val.prepare()
try :
del model
print('delete model is successful')
gc.collect()
except:
pass
KB.clear_session()
model = modellib.MaskRCNN(mode="training", config=config, model_dir=config.CHECKPOINT_FOLDER, FCN_layers = config.FCN_LAYERS)
print('TRAINING_DIR : ', TRAINING_DIR)
print('COCO_TRAINING_DIR : ', COCO_TRAINING_DIR)
print('RESNET_TRAINING_DIR : ', RESNET_TRAINING_DIR)
print('CHECKPOINT_DIR : ', config.CHECKPOINT_FOLDER)
print('Last Saved Model : ', model.find_last())
# exclude_layers = \
# ['fcn_block1_conv1'
# ,'fcn_block1_conv2'
# ,'fcn_block1_pool'
# ,'fcn_block2_conv1'
# ,'fcn_block2_conv2'
# ,'fcn_block2_pool'
# ,'fcn_block3_conv1'
# ,'fcn_block3_conv2'
# ,'fcn_block3_conv3'
# ,'fcn_block3_pool'
# ,'fcn_block4_conv1'
# ,'fcn_block4_conv2'
# ,'fcn_block4_conv3'
# ,'fcn_block4_pool'
# ,'fcn_block5_conv1'
# ,'fcn_block5_conv2'
# ,'fcn_block5_conv3'
# ,'fcn_block5_pool'
# ,'fcn_fc1'
# ,'dropout_1'
# ,'fcn_fc2'
# ,'dropout_2'
# ,'fcn_classify'
# ,'fcn_bilinear'
# ,'fcn_heatmap_norm'
# ,'fcn_scoring'
# ,'fcn_heatmap'
# ,'fcn_norm_loss']
# load_model(model, init_with = 'last', exclude = exclude_layers)
model.load_model_weights(init_with = init_with)
# print('=====================================')
# print(" Load second weight file ?? ")
# model.keras_model.load_weights('E:/Models/vgg16_weights_tf_dim_ordering_tf_kernels.h5', by_name= True)
train_generator = data_generator(dataset_train, model.config, shuffle=True,
batch_size=model.config.BATCH_SIZE,
augment = False)
val_generator = data_generator(dataset_val, model.config, shuffle=True,
batch_size=model.config.BATCH_SIZE,
augment=False)
config.display()
return [model, dataset_train, dataset_val, train_generator, val_generator, config]
def prep_newshapes_test(init_with = 'last', FCN_layers = False, batch_sz = 5, epoch_steps = 4,training_folder= "mrcnn_newshape_test_logs"):
MODEL_DIR = os.path.join(TRAINING_DIR, training_folder)
# Build configuration object -----------------------------------------------
config = new_shapes.NewShapesConfig()
config.BATCH_SIZE = batch_sz # Batch size is 2 (# GPUs * images/GPU).
config.IMAGES_PER_GPU = batch_sz # Must match BATCH_SIZE
config.STEPS_PER_EPOCH = epoch_steps
config.FCN_INPUT_SHAPE = config.IMAGE_SHAPE[0:2]
config.DETECTION_MIN_CONFIDENCE = 0.1
# Build shape dataset -----------------------------------------------
# Training dataset
dataset_test = new_shapes.NewShapesDataset(config)
dataset_test.load_shapes(3000)
dataset_test.prepare()
# Recreate the model in inference mode
try :
del model
print('delete model is successful')
gc.collect()
except:
pass
KB.clear_session()
model = modellib.MaskRCNN(mode="inference",
config=config,
model_dir=MODEL_DIR,
FCN_layers = FCN_layers )
print(' COCO Model Path : ', COCO_TRAINING_DIR)
print(' Checkpoint folder Path: ', MODEL_DIR)
print(' Model Parent Path : ', TRAINING_DIR)
print(' Resent Model Path : ', RESNET_TRAINING_DIR)
# exclude_layers = \
# ['fcn_block1_conv1'
# ,'fcn_block1_conv2'
# ,'fcn_block1_pool'
# ,'fcn_block2_conv1'
# ,'fcn_block2_conv2'
# ,'fcn_block2_pool'
# ,'fcn_block3_conv1'
# ,'fcn_block3_conv2'
# ,'fcn_block3_conv3'
# ,'fcn_block3_pool'
# ,'fcn_block4_conv1'
# ,'fcn_block4_conv2'
# ,'fcn_block4_conv3'
# ,'fcn_block4_pool'
# ,'fcn_block5_conv1'
# ,'fcn_block5_conv2'
# ,'fcn_block5_conv3'
# ,'fcn_block5_pool'
# ,'fcn_fc1'
# ,'dropout_1'
# ,'fcn_fc2'
# ,'dropout_2'
# ,'fcn_classify'
# ,'fcn_bilinear'
# ,'fcn_heatmap_norm'
# ,'fcn_scoring'
# ,'fcn_heatmap'
# ,'fcn_norm_loss']
# load_model(model, init_with = init_with, exclude = exclude_layers)
model.load_model_weights(init_with = init_with)
# print('=====================================')
# print(" Load second weight file ?? ")
# model.keras_model.load_weights('E:/Models/vgg16_weights_tf_dim_ordering_tf_kernels.h5', by_name= True )
test_generator = data_generator(dataset_test, model.config, shuffle=True,
batch_size=model.config.BATCH_SIZE,
augment = False)
model.config.display()
return [model, dataset_test, test_generator, config]
comb_config.FCN_INPUT_SHAPE = comb_config.IMAGE_SHAPE[0:2]
comb_config.LAST_EPOCH_RAN = int(args.last_epoch)
comb_config.WEIGHT_DECAY = 2.0e-4
comb_config.VALIDATION_STEPS = 100
comb_config.REDUCE_LR_FACTOR = 0.5
comb_config.REDUCE_LR_COOLDOWN = 30
comb_config.REDUCE_LR_PATIENCE = 40
comb_config.EARLY_STOP_PATIENCE = 80
comb_config.NEW_LOG_FOLDER = True
comb_config.EARLY_STOP_MIN_DELTA = 1.0e-4
comb_config.MIN_LR = 1.0e-10
##------------------------------------------------------------------------------------
## Build shape dataset for Training and Validation
##------------------------------------------------------------------------------------
dataset_train = shapes.NewShapesDataset(comb_config)
dataset_train.load_shapes(10000)
dataset_train.prepare()
dataset_val = shapes.NewShapesDataset(comb_config)
dataset_val.load_shapes(2500)
dataset_val.prepare()
##------------------------------------------------------------------------------------
## Build Model
##------------------------------------------------------------------------------------
try:
del model
gc.collect()
except:
pass