def prepare_data():
data_transforms = {
'train':
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(256),
transforms.Resize(224),
transforms.RandomHorizontalFlip(),
transforms.RandomVerticalFlip(),
transforms.RandomRotation(90),
transforms.ToTensor(),
transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])
]),
'validation':
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(256),
transforms.Resize(224),
transforms.ToTensor(),
transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])
]),
}
# Train dataset
dataset_train = dataset(opt.data_train_folder, data_transforms["train"])
# Validation dataset
dataset_val = dataset(opt.data_val_folder, data_transforms["validation"])
dataloader_train = torch.utils.data.DataLoader(dataset_train,
batch_size=opt.bs,
shuffle=True,
num_workers=opt.num_workers)
dataloader_val = torch.utils.data.DataLoader(dataset_val,
batch_size=opt.bs,
shuffle=False,
num_workers=opt.num_workers)
train_size = len(dataset_train)
val_size = len(dataset_val)
print("train dataset size =", train_size)
print("validation dataset size=", val_size)
print("dataset train class order= ", dataset_train.class_to_idx)
print("dataset val class order= ", dataset_train.class_to_idx)
#exit() # just for testing
return {
"train": dataloader_train,
"val": dataloader_val,
"dataset_size": {
"train": train_size,
"val": val_size
}
}
def prepare_data():
data_transforms = {
'train': transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(256),
transforms.Resize(224),
transforms.RandomHorizontalFlip(),
transforms.RandomVerticalFlip(),
transforms.RandomRotation(90),
transforms.ToTensor(),
transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])
]),
'validation': transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(256),
transforms.Resize(224),
transforms.ToTensor(),
transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])
]),
}
# Use selected fold for validation
train_folds = list(set(opt.all_folds) - set([opt.val_fold]))
validation_fold = opt.val_fold
# Train datasets
image_datasets_train_all = {x: dataset(os.path.join(opt.data_root, x),
data_transforms["train"])
for x in train_folds}
# create one dataset from all datasets of training
dataset_train = torch.utils.data.ConcatDataset([image_datasets_train_all[i] for i in train_folds])
# Validation datasets
dataset_val = dataset(os.path.join(opt.data_root, validation_fold),
data_transforms["validation"])
dataloader_train = torch.utils.data.DataLoader(dataset_train, batch_size=opt.bs,
shuffle=True, num_workers=opt.num_workers)
dataloader_val = torch.utils.data.DataLoader(dataset_val, batch_size=opt.bs,
shuffle=False, num_workers=opt.num_workers)
train_size = len(dataset_train)
val_size = len(dataset_val)
print("train dataset size =", train_size)
print("validation dataset size=", val_size)
return {"train":dataloader_train, "val":dataloader_val, "dataset_size":{"train": train_size, "val":val_size} }
def inference():
#if opt.bs != 1:
# print("Please run with bs = 1")
# exit()
test_model_checkpoint = input("Please enter the path of test model:")
checkpoint = torch.load(test_model_checkpoint)
model = prepare_model()
model.load_state_dict(checkpoint["model_state_dict"])
model.eval()
trnsfm = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(256),
transforms.Resize(224),
transforms.ToTensor(),
transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])
])
dataset_new = dataset(opt.data_to_inference, trnsfm)
dataloader_new = torch.utils.data.DataLoader(dataset_new,
batch_size=opt.bs,
shuffle=False,
num_workers=opt.num_workers)
class_names = list(string.ascii_uppercase)[:23]
print(class_names)
print("lenth of dataloader:", len(dataloader_new))
df = pd.DataFrame(columns=["filename", "predicted-label"] + class_names)
with torch.no_grad():
for i, data in tqdm(enumerate(dataloader_new, 0)):
inputs, labels, paths = data
df_temp = pd.DataFrame(columns=["filename", "predicted-label"] +
class_names)
#print("paths:", paths)
filenames = []
for p in paths:
#print(p)
#print([list(p.split("/"))[-1]])
filenames = filenames + [list(p.split("/"))[-1]]
#print("filenames:", filename)
#df_temp["filename"] = filename
#print("file names:", filenames)
inputs = inputs.to(device)
labels = labels.to(device)
outputs = model(inputs)
outputs = F.softmax(outputs, 1)
predicted_probability, predicted = torch.max(outputs.data, 1)
predicted = predicted.data.cpu().numpy()
#print("predicted items=", predicted)
#print("paths:", paths)
df_temp["predicted-label"] = predicted
df_temp["filename"] = filenames
#df_temp[2:-1] =
#print(df_temp)
#df_temp["actual-label"] = class_names[labels.item()]
# print("actual label:", labels.item())
#print("predicted label:", predicted.item())
# print("probabilities :", outputs.cpu())
probabilities = outputs.cpu().squeeze()
probabilities = probabilities.tolist()
probabilities = np.around(probabilities, decimals=3)
#print(probabilities)
#print(probabilities)
df_temp[class_names] = probabilities
#print(df_temp)
#record = record + [class_names[labels.item()]] + [class_names[predicted.item()]]
#print(record)
#print(df_temp.head())
df = df.append(df_temp)
# break
print(df.head())
print("length of DF:", len(df))
prob_file_name = "%s/%s_inference.csv" % (opt.out_dir, py_file_name)
df.to_csv(prob_file_name, index=False)