def eval_fidelity(tree: ProtoTree,
test_loader: DataLoader,
device,
log: Log = None,
progress_prefix: str = 'Fidelity') -> dict:
tree = tree.to(device)
# Keep an info dict about the procedure
info = dict()
# Make sure the model is in evaluation mode
tree.eval()
# Show progress on progress bar
test_iter = tqdm(enumerate(test_loader),
total=len(test_loader),
desc=progress_prefix,
ncols=0)
distr_samplemax_fidelity = 0
distr_greedy_fidelity = 0
# Iterate through the test set
for i, (xs, ys) in test_iter:
xs, ys = xs.to(device), ys.to(device)
# Use the model to classify this batch of input data, with 3 types of routing
out_distr, _ = tree.forward(xs, 'distributed')
ys_pred_distr = torch.argmax(out_distr, dim=1)
out_samplemax, _ = tree.forward(xs, 'sample_max')
ys_pred_samplemax = torch.argmax(out_samplemax, dim=1)
out_greedy, _ = tree.forward(xs, 'greedy')
ys_pred_greedy = torch.argmax(out_greedy, dim=1)
# Calculate fidelity
distr_samplemax_fidelity += torch.sum(
torch.eq(ys_pred_samplemax, ys_pred_distr)).item()
distr_greedy_fidelity += torch.sum(
torch.eq(ys_pred_greedy, ys_pred_distr)).item()
# Update the progress bar
test_iter.set_postfix_str(f'Batch [{i + 1}/{len(test_iter)}]')
del out_distr
del out_samplemax
del out_greedy
distr_samplemax_fidelity = distr_samplemax_fidelity / float(
len(test_loader.dataset))
distr_greedy_fidelity = distr_greedy_fidelity / float(
len(test_loader.dataset))
info['distr_samplemax_fidelity'] = distr_samplemax_fidelity
info['distr_greedy_fidelity'] = distr_greedy_fidelity
log.log_message(
"Fidelity between standard distributed routing and sample_max routing: "
+ str(distr_samplemax_fidelity))
log.log_message(
"Fidelity between standard distributed routing and greedy routing: " +
str(distr_greedy_fidelity))
return info
def save_tree_description(tree: ProtoTree, optimizer, scheduler,
description: str, log: Log):
tree.eval()
# Save model with description
tree.save(f'{log.checkpoint_dir}/' + description)
tree.save_state(f'{log.checkpoint_dir}/' + description)
torch.save(optimizer.state_dict(),
f'{log.checkpoint_dir}/' + description + '/optimizer_state.pth')
torch.save(scheduler.state_dict(),
f'{log.checkpoint_dir}/' + description + '/scheduler_state.pth')
def eval(tree: ProtoTree,
test_loader: DataLoader,
epoch,
device,
log: Log = None,
sampling_strategy: str = 'distributed',
log_prefix: str = 'log_eval_epochs',
progress_prefix: str = 'Eval Epoch') -> dict:
tree = tree.to(device)
# Keep an info dict about the procedure
info = dict()
if sampling_strategy != 'distributed':
info['out_leaf_ix'] = []
# Build a confusion matrix
cm = np.zeros((tree._num_classes, tree._num_classes), dtype=int)
# Make sure the model is in evaluation mode
tree.eval()
# Show progress on progress bar
test_iter = tqdm(enumerate(test_loader),
total=len(test_loader),
desc=progress_prefix + ' %s' % epoch,
ncols=0)
# Iterate through the test set
for i, (xs, ys) in test_iter:
xs, ys = xs.to(device), ys.to(device)
# Use the model to classify this batch of input data
out, test_info = tree.forward(xs, sampling_strategy)
ys_pred = torch.argmax(out, dim=1)
# Update the confusion matrix
cm_batch = np.zeros((tree._num_classes, tree._num_classes), dtype=int)
for y_pred, y_true in zip(ys_pred, ys):
cm[y_true][y_pred] += 1
cm_batch[y_true][y_pred] += 1
acc = acc_from_cm(cm_batch)
test_iter.set_postfix_str(
f'Batch [{i + 1}/{len(test_iter)}], Acc: {acc:.3f}')
# keep list of leaf indices where test sample ends up when deterministic routing is used.
if sampling_strategy != 'distributed':
info['out_leaf_ix'] += test_info['out_leaf_ix']
del out
del ys_pred
del test_info
info['confusion_matrix'] = cm
info['test_accuracy'] = acc_from_cm(cm)
log.log_message("\nEpoch %s - Test accuracy with %s routing: " %
(epoch, sampling_strategy) + str(info['test_accuracy']))
return info
def save_best_test_tree(tree: ProtoTree, optimizer, scheduler,
best_test_acc: float, test_acc: float, log: Log):
tree.eval()
if test_acc > best_test_acc:
best_test_acc = test_acc
tree.save(f'{log.checkpoint_dir}/best_test_model')
tree.save_state(f'{log.checkpoint_dir}/best_test_model')
torch.save(
optimizer.state_dict(),
f'{log.checkpoint_dir}/best_test_model/optimizer_state.pth')
torch.save(
scheduler.state_dict(),
f'{log.checkpoint_dir}/best_test_model/scheduler_state.pth')
return best_test_acc
def train_leaves_epoch(tree: ProtoTree,
train_loader: DataLoader,
epoch: int,
device,
progress_prefix: str = 'Train Leafs Epoch'
) -> dict:
#Make sure the tree is in eval mode for updating leafs
tree.eval()
with torch.no_grad():
_old_dist_params = dict()
for leaf in tree.leaves:
_old_dist_params[leaf] = leaf._dist_params.detach().clone()
# Optimize class distributions in leafs
eye = torch.eye(tree._num_classes).to(device)
# Show progress on progress bar
train_iter = tqdm(enumerate(train_loader),
total=len(train_loader),
desc=progress_prefix+' %s'%epoch,
ncols=0)
# Iterate through the data set
update_sum = dict()
# Create empty tensor for each leaf that will be filled with new values
for leaf in tree.leaves:
update_sum[leaf] = torch.zeros_like(leaf._dist_params)
for i, (xs, ys) in train_iter:
xs, ys = xs.to(device), ys.to(device)
#Train leafs without gradient descent
out, info = tree.forward(xs)
target = eye[ys] #shape (batchsize, num_classes)
for leaf in tree.leaves:
if tree._log_probabilities:
# log version
update = torch.exp(torch.logsumexp(info['pa_tensor'][leaf.index] + leaf.distribution() + torch.log(target) - out, dim=0))
else:
update = torch.sum((info['pa_tensor'][leaf.index] * leaf.distribution() * target)/out, dim=0)
update_sum[leaf] += update
for leaf in tree.leaves:
leaf._dist_params -= leaf._dist_params #set current dist params to zero
leaf._dist_params += update_sum[leaf] #give dist params new value
def save_tree(tree: ProtoTree, optimizer, scheduler, epoch: int, log: Log,
args: argparse.Namespace):
tree.eval()
# Save latest model
tree.save(f'{log.checkpoint_dir}/latest')
tree.save_state(f'{log.checkpoint_dir}/latest')
torch.save(optimizer.state_dict(),
f'{log.checkpoint_dir}/latest/optimizer_state.pth')
torch.save(scheduler.state_dict(),
f'{log.checkpoint_dir}/latest/scheduler_state.pth')
# Save model every 10 epochs
if epoch == args.epochs or epoch % 10 == 0:
tree.save(f'{log.checkpoint_dir}/epoch_{epoch}')
tree.save_state(f'{log.checkpoint_dir}/epoch_{epoch}')
torch.save(optimizer.state_dict(),
f'{log.checkpoint_dir}/epoch_{epoch}/optimizer_state.pth')
torch.save(scheduler.state_dict(),
f'{log.checkpoint_dir}/epoch_{epoch}/scheduler_state.pth')
def project_with_class_constraints(
tree: ProtoTree,
project_loader: DataLoader,
device,
args: argparse.Namespace,
log: Log,
log_prefix: str = 'log_projection_with_constraints', # TODO
progress_prefix: str = 'Projection') -> dict:
log.log_message(
"\nProjecting prototypes to nearest training patch (with class restrictions)..."
)
# Set the model to evaluation mode
tree.eval()
torch.cuda.empty_cache()
# The goal is to find the latent patch that minimizes the L2 distance to each prototype
# To do this we iterate through the train dataset and store for each prototype the closest latent patch seen so far
# Also store info about the image that was used for projection
global_min_proto_dist = {j: np.inf for j in range(tree.num_prototypes)}
global_min_patches = {j: None for j in range(tree.num_prototypes)}
global_min_info = {j: None for j in range(tree.num_prototypes)}
# Get the shape of the prototypes
W1, H1, D = tree.prototype_shape
# Build a progress bar for showing the status
projection_iter = tqdm(enumerate(project_loader),
total=len(project_loader),
desc=progress_prefix,
ncols=0)
with torch.no_grad():
# Get a batch of data
xs, ys = next(iter(project_loader))
batch_size = xs.shape[0]
# For each internal node, collect the leaf labels in the subtree with this node as root.
# Only images from these classes can be used for projection.
leaf_labels_subtree = dict()
for branch, j in tree._out_map.items():
leaf_labels_subtree[branch.index] = set()
for leaf in branch.leaves:
leaf_labels_subtree[branch.index].add(
torch.argmax(leaf.distribution()).item())
for i, (xs, ys) in projection_iter:
xs, ys = xs.to(device), ys.to(device)
# Get the features and distances
# - features_batch: features tensor (shared by all prototypes)
# shape: (batch_size, D, W, H)
# - distances_batch: distances tensor (for all prototypes)
# shape: (batch_size, num_prototypes, W, H)
# - out_map: a dict mapping decision nodes to distances (indices)
features_batch, distances_batch, out_map = tree.forward_partial(xs)
# Get the features dimensions
bs, D, W, H = features_batch.shape
# Get a tensor containing the individual latent patches
# Create the patches by unfolding over both the W and H dimensions
# TODO -- support for strides in the prototype layer? (corresponds to step size here)
patches_batch = features_batch.unfold(2, W1, 1).unfold(
3, H1, 1) # Shape: (batch_size, D, W, H, W1, H1)
# Iterate over all decision nodes/prototypes
for node, j in out_map.items():
leaf_labels = leaf_labels_subtree[node.index]
# Iterate over all items in the batch
# Select the features/distances that are relevant to this prototype
# - distances: distances of the prototype to the latent patches
# shape: (W, H)
# - patches: latent patches
# shape: (D, W, H, W1, H1)
for batch_i, (distances, patches) in enumerate(
zip(distances_batch[:, j, :, :], patches_batch)):
#Check if label of this image is in one of the leaves of the subtree
if ys[batch_i].item() in leaf_labels:
# Find the index of the latent patch that is closest to the prototype
min_distance = distances.min()
min_distance_ix = distances.argmin()
# Use the index to get the closest latent patch
closest_patch = patches.view(D, W * H, W1,
H1)[:,
min_distance_ix, :, :]
# Check if the latent patch is closest for all data samples seen so far
if min_distance < global_min_proto_dist[j]:
global_min_proto_dist[j] = min_distance
global_min_patches[j] = closest_patch
global_min_info[j] = {
'input_image_ix': i * batch_size + batch_i,
'patch_ix': min_distance_ix.item(
), # Index in a flattened array of the feature map
'W': W,
'H': H,
'W1': W1,
'H1': H1,
'distance': min_distance.item(),
'nearest_input':
torch.unsqueeze(xs[batch_i], 0),
'node_ix': node.index,
}
# Update the progress bar if required
projection_iter.set_postfix_str(
f'Batch: {i + 1}/{len(project_loader)}')
del features_batch
del distances_batch
del out_map
# Copy the patches to the prototype layer weights
projection = torch.cat(tuple(global_min_patches[j].unsqueeze(0)
for j in range(tree.num_prototypes)),
dim=0,
out=tree.prototype_layer.prototype_vectors)
del projection
return global_min_info, tree
def train_epoch(tree: ProtoTree,
train_loader: DataLoader,
optimizer: torch.optim.Optimizer,
epoch: int,
disable_derivative_free_leaf_optim: bool,
device,
log: Log = None,
log_prefix: str = 'log_train_epochs',
progress_prefix: str = 'Train Epoch'
) -> dict:
tree = tree.to(device)
# Make sure the model is in eval mode
tree.eval()
# Store info about the procedure
train_info = dict()
total_loss = 0.
total_acc = 0.
# Create a log if required
log_loss = f'{log_prefix}_losses'
nr_batches = float(len(train_loader))
with torch.no_grad():
_old_dist_params = dict()
for leaf in tree.leaves:
_old_dist_params[leaf] = leaf._dist_params.detach().clone()
# Optimize class distributions in leafs
eye = torch.eye(tree._num_classes).to(device)
# Show progress on progress bar
train_iter = tqdm(enumerate(train_loader),
total=len(train_loader),
desc=progress_prefix+' %s'%epoch,
ncols=0)
# Iterate through the data set to update leaves, prototypes and network
for i, (xs, ys) in train_iter:
# Make sure the model is in train mode
tree.train()
# Reset the gradients
optimizer.zero_grad()
xs, ys = xs.to(device), ys.to(device)
# Perform a forward pass through the network
ys_pred, info = tree.forward(xs)
# Learn prototypes and network with gradient descent.
# If disable_derivative_free_leaf_optim, leaves are optimized with gradient descent as well.
# Compute the loss
if tree._log_probabilities:
loss = F.nll_loss(ys_pred, ys)
else:
loss = F.nll_loss(torch.log(ys_pred), ys)
# Compute the gradient
loss.backward()
# Update model parameters
optimizer.step()
if not disable_derivative_free_leaf_optim:
#Update leaves with derivate-free algorithm
#Make sure the tree is in eval mode
tree.eval()
with torch.no_grad():
target = eye[ys] #shape (batchsize, num_classes)
for leaf in tree.leaves:
if tree._log_probabilities:
# log version
update = torch.exp(torch.logsumexp(info['pa_tensor'][leaf.index] + leaf.distribution() + torch.log(target) - ys_pred, dim=0))
else:
update = torch.sum((info['pa_tensor'][leaf.index] * leaf.distribution() * target)/ys_pred, dim=0)
leaf._dist_params -= (_old_dist_params[leaf]/nr_batches)
F.relu_(leaf._dist_params) #dist_params values can get slightly negative because of floating point issues. therefore, set to zero.
leaf._dist_params += update
# Count the number of correct classifications
ys_pred_max = torch.argmax(ys_pred, dim=1)
correct = torch.sum(torch.eq(ys_pred_max, ys))
acc = correct.item() / float(len(xs))
train_iter.set_postfix_str(
f'Batch [{i + 1}/{len(train_loader)}], Loss: {loss.item():.3f}, Acc: {acc:.3f}'
)
# Compute metrics over this batch
total_loss+=loss.item()
total_acc+=acc
if log is not None:
log.log_values(log_loss, epoch, i + 1, loss.item(), acc)
train_info['loss'] = total_loss/float(i+1)
train_info['train_accuracy'] = total_acc/float(i+1)
return train_info