# dataset_csv_file=test_csv,
# output_size=cnn_image_size)
# dataloader_test = DataLoader(dataset_test, batch_size=args.batch_size,
# shuffle=True, num_workers=4)
training_set = StrongLossDataset(file=args.training_file,
image_path=args.image_path,
transforms=transformer)
validation_set = StrongLossDataset(file=args.validation_file,
image_path=args.image_path,
transforms=transformer)
# build dataloader
training_loader = DataLoader(training_set,
batch_size=args.batch_size,
num_workers=4,
shuffle=True)
validation_loader = DataLoader(validation_set,
batch_size=args.batch_size,
num_workers=4,
shuffle=True)
# Define checkpoint name
checkpoint_name = os.path.join(
args.result_model_dir,
datetime.datetime.now().strftime("%Y-%m-%d_%H:%M") + '_' +
args.result_model_fn + '.pth.tar')
print('Checkpoint name: ' + checkpoint_name)
# Train
args = parser.parse_args()
return args.spec, args.test_dataset
if __name__ == "__main__":
spec, dataset = parse_spec()
if not os.path.exists(spec):
raise FileNotFoundError("Path to spec file not found")
with open(spec, "r") as file:
name = os.path.basename(spec)[:-5]
spec_data = json.load(file)
model_path = os.path.join(SAVED_MODEL_DIR, f'{name}.hp5')
if not os.path.exists(model_path):
raise UnsavedModelError("Run the experiment before evaluating")
datasets = spec_data['data']['datasets']
data_params = spec_data['data'].get('data_params', {})
loader = DataLoader(datasets, **data_params)
test_data = loader.generate_test_dataset(dataset)
# TODO: This could easily be done to exhaustion. Alter DataLoader API to handle test sets better
model = tf.keras.models.load_model(model_path)
results = model.evaluate(test_data)
print(f"Model Results: {results}")
normalization_tnf = normalize_image_dict_caffe
else:
normalization_tnf = NormalizeImageDict(['source_image', 'target_image'])
batch_preprocessing_fn = BatchTensorToVars(use_cuda=use_cuda)
# Dataset and dataloader
dataset = Dataset(transform=normalization_tnf,
dataset_image_path=args.dataset_image_path,
dataset_csv_path=args.dataset_csv_path,
dataset_csv_file=train_csv,
output_size=cnn_image_size,
random_affine=bool(args.random_affine))
dataloader = DataLoader(dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=0)
dataset_test = Dataset(transform=normalization_tnf,
dataset_image_path=args.dataset_image_path,
dataset_csv_path=args.dataset_csv_path,
dataset_csv_file=test_csv,
output_size=cnn_image_size)
dataloader_test = DataLoader(dataset_test,
batch_size=args.batch_size,
shuffle=True,
num_workers=4)
# Define checkpoint name
checkpoint_name = os.path.join(
spec_data = json.load(file)
# Validate the json file
validate_spec(spec_data)
if spec_data["architecture"] == "simple":
builder = partial(build_simple_cnn1, **spec_data["builder_args"])
trainer = partial(train_from_scratch, **spec_data["trainer_args"])
elif spec_data["architecture"] == "baseline":
builder = partial(build_baseline)
trainer = partial(train_from_scratch, **spec_data["trainer_args"])
elif spec_data["architecture"] == "vgg16":
builder = partial(build_vgg_model, **spec_data["builder_args"])
if spec_data["builder_args"].get("weights") == "imagenet":
trainer = partial(train_from_pretrained,
**spec_data["trainer_args"])
else:
trainer = partial(train_from_scratch, **spec_data["trainer_args"])
save_dir = "/home/matt/Projects/CV/final/saved_models/"
results_dir = "/home/matt/Projects/CV/final/experiments/results/"
datasets = spec_data['data']['datasets']
data_params = spec_data['data'].get('data_params', {})
loader = DataLoader(datasets, **data_params)
save_path = os.path.join(save_dir, f'{name}.hp5')
results_path = os.path.join(results_dir, name)
main(builder, trainer, loader, save_path, results_path)