def predict(model_path, im_path):
'''
Test procedure
---------------
:param model_path: path of the saved model
:param im_path: path of an image
'''
# TODO 3: load configurations from saved model, initialize the model.
# Note: you can complete this section by referring to Part 4: test.
# step 1: load configurations from saved model using torch.load(model_path)
# and get the configs dictionary, configs = checkpoint['configs'],
# then get each config from configs, eg., norm_size = configs['norm_size']
checkpoint = torch.load(model_path)
configs = checkpoint['configs']
norm_size = configs['norm_size']
output_size = configs['output_size']
hidden_size = configs['hidden_size']
n_layers = configs['n_layers']
act_type = configs['act_type']
# step 2: initialize the model by MLP()
model = MLP(norm_size[0] * norm_size[1], output_size, hidden_size,
n_layers, act_type)
# step 3: load model parameters we saved in model_path
# hint: similar to what we do in Part 4: test.
model.load_state_dict(checkpoint['state_dict'])
# End TODO 3
# enter the evaluation mode
model.eval()
# image pre-processing, similar to what we do in ListDataset()
transform = transforms.Compose([
transforms.ToPILImage(),
transforms.Resize(norm_size),
transforms.ToTensor()
])
# image pre-processing, similar to what we do in ListDataset()
im = cv2.imread(im_path)
im = cv2.cvtColor(im, cv2.COLOR_BGR2GRAY)
im = transform(im)
im.sub_(0.5).div_(0.5)
# input im into the model
with torch.no_grad():
input = im.view(1, -1)
out = model(input)
prediction = out.argmax(1)[0].item()
# convert index of prediction to the corresponding character
letters = string.ascii_letters[-26:] # ABCD...XYZ
prediction = letters[prediction]
print('Prediction: {}'.format(prediction))
# epoch, step, len(train_loader), batch_time=batch_time, loss=losses, top1=top1))
train_loss = losses.avg
train_acc = top1.avg
train_losses.append(train_loss)
train_accs.append(train_acc)
# vis.plot_many({'train loss': train_loss, 'train acc': train_acc})
# validation phase
print('validation phase')
# vis.log('validation phase')
batch_time = AverageMeter()
losses = AverageMeter()
top1 = AverageMeter()
net.eval()
end = time.time()
for step, (images, labels) in enumerate(val_loader):
images = images.to(device)
labels = labels.to(device)
with torch.no_grad():
logits = images
# logits = clf(clf_norm(images))
# logits = logits - logits.max(dim=1, keepdim=True)[0] # Normalizing by reducing the maximum
# logits = logits / logits.norm(p=2, dim=1, keepdim=True) # L2 normalization
output = net(logits)
loss = criterion(output, labels)
_, pred = output.max(dim=1)
def test(model_path,
im_dir='data/character_classification/images',
test_file_path='data/character_classification/test.json',
batch_size=8,
device='cpu'):
'''
Test procedure
---------------
:param model_path: path of the saved model
:param im_dir: path to directory with images
:param test_file_path: file with test image paths and labels
:param batch_size: test batch size
:param device: 'cpu' or 'cuda'
'''
# load configurations from saved model, initialize and test the model
checkpoint = torch.load(model_path)
configs = checkpoint['configs']
norm_size = configs['norm_size']
output_size = configs['output_size']
hidden_size = configs['hidden_size']
n_layers = configs['n_layers']
act_type = configs['act_type']
# initialize the model by MLP()
model = MLP(norm_size[0] * norm_size[1], output_size, hidden_size,
n_layers, act_type)
# load model parameters we saved in model_path
model.load_state_dict(checkpoint['state_dict'])
model = model.to(device)
print('[Info] Load model from {}'.format(model_path))
# enter the evaluation mode
model.eval()
# test loader
testloader = dataLoader(im_dir, test_file_path, norm_size, batch_size)
# run the test process
n_correct = 0.
n_ims = 0.
logits = []
all_labels = []
with torch.no_grad(
): # we do not need to compute gradients during test stage
for ims, labels in testloader:
ims, labels = ims.to(device), labels.type(torch.float).to(device)
input = ims.view(ims.size(0), -1)
out = model(input)
predictions = out.argmax(1)
n_correct += torch.sum(predictions == labels)
n_ims += ims.size(0)
logits.append(out)
all_labels.append(labels)
logits = torch.cat(logits, dim=0).detach().cpu().numpy()
all_labels = torch.cat(all_labels, dim=0).cpu().numpy()
tsne = TSNE(n_components=2, init='pca')
Y = tsne.fit_transform(logits)
letters = list(string.ascii_letters[-26:])
Y = (Y - Y.min(0)) / (Y.max(0) - Y.min(0))
for i in range(len(all_labels)):
if (all_labels[i] < 26):
c = plt.cm.rainbow(float(all_labels[i]) / 26)
plt.text(Y[i, 0],
Y[i, 1],
s=letters[int(all_labels[i])],
color=c)
plt.show()
print('[Info] Test accuracy = {:.1f}%'.format(100 * n_correct / n_ims))
def train_val(im_dir,
train_file_path,
val_file_path,
hidden_size,
n_layers,
act_type,
norm_size,
n_epochs,
batch_size,
n_letters,
lr,
optim_type,
momentum,
weight_decay,
valInterval,
device='cpu'):
'''
The main training procedure
----------------------------
:param im_dir: path to directory with images
:param train_file_path: file list of training image paths and labels
:param val_file_path: file list of validation image paths and labels
:param hidden_size: a list of hidden size for each hidden layer
:param n_layers: number of layers in the MLP
:param act_type: type of activation function, can be none, sigmoid, tanh, or relu
:param norm_size: image normalization size, (height, width)
:param n_epochs: number of training epochs
:param batch_size: batch size of training and validation
:param n_letters: number of classes, in this task it is 26 English letters
:param lr: learning rate
:param optim_type: optimizer, can be 'sgd', 'adagrad', 'rmsprop', 'adam', or 'adadelta'
:param momentum: only used if optim_type == 'sgd'
:param weight_decay: the factor of L2 penalty on network weights
:param valInterval: the frequency of validation, e.g., if valInterval = 5, then do validation after each 5 training epochs
:param device: 'cpu' or 'cuda', we can use 'cpu' for our homework if GPU with cuda support is not available
'''
# training and validation data loader
trainloader = dataLoader(im_dir, train_file_path, norm_size, batch_size)
valloader = dataLoader(im_dir, val_file_path, norm_size, batch_size)
# TODO 1: initialize the MLP model and loss function
# what is the input size of the MLP?
# hint 1: we convert an image to a vector as the input of the MLP,
# each image has shape [norm_size[0], norm_size[1]]
# hint 2: Input parameters for MLP: input_size, output_size, hidden_size, n_layers, act_type
model = MLP(norm_size[0] * norm_size[1], n_letters, hidden_size, n_layers,
act_type)
# loss function
cal_loss = CrossEntropyLoss.apply
# End TODO 1
# put the model on CPU or GPU
model = model.to(device)
# optimizer
if optim_type == 'sgd':
optimizer = optim.SGD(model.parameters(),
lr,
momentum=momentum,
weight_decay=weight_decay)
elif optim_type == 'adagrad':
optimizer = optim.Adagrad(model.parameters(),
lr,
weight_decay=weight_decay)
elif optim_type == 'rmsprop':
optimizer = optim.RMSprop(model.parameters(),
lr,
weight_decay=weight_decay)
elif optim_type == 'adam':
optimizer = optim.Adam(model.parameters(),
lr,
weight_decay=weight_decay)
elif optim_type == 'adadelta':
optimizer = optim.Adadelta(model.parameters(),
lr,
weight_decay=weight_decay)
else:
print(
'[Error] optim_type should be one of sgd, adagrad, rmsprop, adam, or adadelta'
)
raise NotImplementedError
# training
# to save loss of each training epoch in a python "list" data structure
losses = []
for epoch in range(n_epochs):
# set the model in training mode
model.train()
# to save total loss in one epoch
total_loss = 0.
#TODO 2: calculate losses and train the network using the optimizer
for step, (ims,
labels) in enumerate(trainloader): # get a batch of data
# step 1: set data type and device
ims = ims.to(device)
labels = labels.to(device)
# step 2: convert an image to a vector as the input of the MLP
ims = ims.view(batch_size, norm_size[0] * norm_size[1])
# hint: clear gradients in the optimizer
optimizer.zero_grad()
# step 3: run the model which is the forward process
pred = model(ims)
# step 4: compute the loss, and call backward propagation function
loss = cal_loss(pred, labels)
loss.backward()
# step 5: sum up of total loss, loss.item() return the value of the tensor as a standard python number
# this operation is not differentiable
total_loss += loss.item()
# step 6: call a function, optimizer.step(), to update the parameters of the model
optimizer.step()
# End TODO 2
# average of the total loss for iterations
avg_loss = total_loss / len(trainloader)
losses.append(avg_loss)
print('Epoch {:02d}: loss = {:.3f}'.format(epoch + 1, avg_loss))
# validation
if (epoch + 1) % valInterval == 0:
# set the model in evaluation mode
model.eval()
n_correct = 0. # number of images that are correctly classified
n_ims = 0. # number of total images
with torch.no_grad(
): # we do not need to compute gradients during validation
# calculate losses for validation data and do not need train the network
for ims, labels in valloader:
# set data type and device
ims, labels = ims.to(device), labels.type(
torch.float).to(device)
# convert an image to a vector as the input of the MLP
input = ims.view(ims.size(0), -1)
# run the model which is the forward process
out = model(input)
# get the predicted value by the output using out.argmax(1)
predictions = out.argmax(1)
# sum up the number of images correctly recognized and the total image number
n_correct += torch.sum(predictions == labels)
n_ims += ims.size(0)
# show prediction accuracy
print('Epoch {:02d}: validation accuracy = {:.1f}%'.format(
epoch + 1, 100 * n_correct / n_ims))
# save model parameters in a file
model_save_path = 'saved_models/recognition.pth'.format(epoch + 1)
torch.save(
{
'state_dict': model.state_dict(),
'configs': {
'norm_size': norm_size,
'output_size': n_letters,
'hidden_size': hidden_size,
'n_layers': n_layers,
'act_type': act_type
}
}, model_save_path)
print('Model saved in {}\n'.format(model_save_path))
# draw the loss curve
plot_loss(losses)
