We Keep Coding

AttributeError: 'KMeans' object has no attribute 'labels_' pytorch

first of all I thank , I tried to train model with pytorch but I got the following error: AttributeError: 'KMeans' object has no attribute 'labels_'.I am trying to model a extract features point cloud using deep learning in pytorch. I get the following error . Could anyone help on this? ************** *************** Thanks!
# Training loop
def training_loop(gpu, training_dataloader, model, loss_fn, optimizer):
losses = []
correct = 0
batch_results = dict()
conf_mat = np.zeros((10,10))
for batch_n, batch in enumerate(training_dataloader): #batch[batch, pos, ptr, y]
batch_size = int(batch.batch.size()[0] / sample_points)
if dimensionality == 3:
# Input dim [:,3] for your geometry x,y,z
X = batch.pos.cuda(non_blocking=True).view(batch_size, sample_points, -1) + torch.normal(
torch.zeros(batch_size, sample_points, dimensionality), torch.full((batch_size, sample_points,
dimensionality), fill_value=0.1)).cuda(gpu)
else:
# Input dim [:,6] for your geometry x,y,z and normals nx,ny,nz
X = torch.cat((batch.pos.cuda(non_blocking=True), batch.normal.cuda(non_blocking=True)), 1).view(batch_size, sample_points, -1) + torch.normal(
torch.zeros(batch_size, sample_points, dimensionality), torch.full((batch_size, sample_points,
dimensionality), fill_value=0.1)).cuda(gpu)
y = batch.y.cuda(non_blocking=True).flatten() #size (batch_size) --> torch.Size([8])
# Compute predictions
pred = model(None, X) #size (batch_size,classes) --> torch.Size([8, 10])
if overall_classes_loss:
# weighted CE Loss over all classes
loss = loss_fn(pred, y)
else:
# weighted batchwise Loss
sample_count = np.array([[x, batch.y.tolist().count(x)] for x in batch.y])[:,1]
batch_weights = 1. / sample_count
batch_weights = torch.from_numpy(batch_weights)
batch_weights = batch_weights.double()
loss = element_weighted_loss(pred, batch.y, batch_weights, gpu)
correct += (pred.argmax(1) == y).type(torch.float).sum().item()
print(f"Loss: {loss}")
tensor_list_y = [torch.ones_like(y) for _ in range(dist.get_world_size())]
tensor_list_pred = [torch.ones_like(y) for _ in range(dist.get_world_size())]
torch.distributed.all_gather(tensor_list_y, y, group=None, async_op=False)
torch.distributed.all_gather(tensor_list_pred, pred.argmax(1), group=None, async_op=False)
tensor_list_y = torch.cat(tensor_list_y)
tensor_list_pred = torch.cat(tensor_list_pred)
# Confusion Matrix
conf_mat += confusion_matrix(tensor_list_y.cpu().detach().numpy(), tensor_list_pred.cpu().detach().numpy(), labels=np.arange(0,10))
# Backpropagation
optimizer.zero_grad()
loss.backward()
optimizer.step()
losses.append(loss.item())
# Save batch predictions
batch_results[batch_n] = {'true':tensor_list_y, 'pred':tensor_list_pred}
if verbosity == True:
print(f"\n\nTRAIN on GPU:{gpu}: True Label {y} - Prediction {pred.argmax(1)} - Loss {loss}")
truevalue = '\t\t'.join(classes[items] for items in y.tolist())
predvalues = '\t\t'.join(classes[items] for items in pred.argmax(1).tolist())
print(f"INFO on GPU:{gpu}: TRAIN - True Value\t {truevalue}")
print(f"INFO on GPU:{gpu}: TRAIN - Predictions\t {predvalues}")
if batch_n % 25 == 0:
torch.distributed.reduce(loss, 0)
return torch.tensor(losses, device=f"cuda:{gpu}"), torch.tensor(correct, device=f"cuda:{gpu}"), batch_results, conf_mat
# Test loop
def test_loop(gpu, test_dataloader, model, loss_fn):
test_losses = []
correct = 0
batch_results = dict()
conf_mat = np.zeros((10,10))
with torch.no_grad():
for batch_n, batch in enumerate(test_dataloader):
batch_size = int(batch.batch.size()[0] / sample_points)
if dimensionality == 3:
# Input dim [:,3] for your geometry x,y,z
X = batch.pos.cuda(non_blocking=True).view(batch_size, sample_points, -1)
else:
# Input dim [:,6] for your geometry x,y,z and normals nx,ny,nz
X = torch.cat((batch.pos.cuda(non_blocking=True), batch.normal.cuda(non_blocking=True)), 1).view(batch_size, sample_points, -1)
y = batch.y.cuda(non_blocking=True).flatten()
pred = model(None, X) #size (batch,classes) per batch_n
if overall_classes_loss:
# weighted CE Loss over all classes
loss = loss_fn(pred, y)
else:
# weighted batchwise Loss
sample_count = np.array([[x, batch.y.tolist().count(x)] for x in batch.y])[:,1]
batch_weights = 1. / sample_count
batch_weights = torch.from_numpy(batch_weights)
batch_weights = batch_weights.double()
loss = element_weighted_loss(pred, batch.y, batch_weights, gpu)
test_losses.append(loss.item())
correct += (pred.argmax(1) == y).type(torch.float).sum().item()
print(f"Loss: {loss}")
tensor_list_y = [torch.ones_like(y) for _ in range(dist.get_world_size())]
tensor_list_pred = [torch.ones_like(y) for _ in range(dist.get_world_size())]
torch.distributed.all_gather(tensor_list_y, y, group=None, async_op=False)
torch.distributed.all_gather(tensor_list_pred, pred.argmax(1), group=None, async_op=False)
tensor_list_y = torch.cat(tensor_list_y)
tensor_list_pred = torch.cat(tensor_list_pred)
# Confusion Matrix
conf_mat += confusion_matrix(tensor_list_y.cpu().detach().numpy(), tensor_list_pred.cpu().detach().numpy(), labels=np.arange(0,10))
# Save batch predictions
batch_results[batch_n] = {'true':tensor_list_y, 'pred':tensor_list_pred}
if verbosity == True:
print(f"\n\nTEST on GPU:{gpu}: True Label {y} - Prediction {pred.argmax(1)} - Loss {loss}")
truevalue = '\t\t'.join(classes[items] for items in y.tolist())
predvalues = '\t\t'.join(classes[items] for items in pred.argmax(1).tolist())
print(f"INFO on GPU:{gpu}: TEST - True Value\t {truevalue}")
print(f"INFO on GPU:{gpu}: TEST - Predictions\t {predvalues}")
test_loss = statistics.mean(test_losses)
return torch.tensor(correct, device=f"cuda:{gpu}"), torch.tensor(test_loss, device=f"cuda:{gpu}"), batch_results, conf_mat
def train_optimisation(gpu, gpus, training_dataloader, test_dataloader, model, loss_fn, optimizer, scheduler, dir_path, initial_epoch):
epoch_losses = []
training_accuracies = []
test_losses = []
test_accuracies = []
learning_rates = []
counter = 0 #early stopping counter
batchwise_results = dict()
# Learning Rate Scheduler
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', patience=20)
for i in range(initial_epoch, initial_epoch + epochs):
if gpu == 0:
if initial_epoch > 0:
print(f"\n\nEpoch {i}\n-------------------------------")
else:
print(f"\n\nEpoch {i + 1}\n-------------------------------")
# TRAIN
losses, training_accuracy, train_batch_result, train_conf_mat = training_loop(gpu, training_dataloader, model, loss_fn, optimizer)
average_loss = torch.mean(losses)
torch.distributed.reduce(average_loss, 0, torch.distributed.ReduceOp.SUM)
torch.distributed.reduce(training_accuracy, 0, torch.distributed.ReduceOp.SUM)
# TEST
test_accuracy, test_loss, test_batch_result, test_conf_mat = test_loop(gpu, test_dataloader, model, loss_fn)
torch.distributed.reduce(test_accuracy, 0, torch.distributed.ReduceOp.SUM)
torch.distributed.reduce(test_loss, 0, torch.distributed.ReduceOp.SUM)
# save results
batchwise_results[i] = {'train':train_batch_result, 'test':test_batch_result}
if gpu == 0: # the following operations are performed only by the process running in the first gpu
average_loss = average_loss / torch.tensor(gpus, dtype=torch.float) # average loss among all gpus
test_accuracy = test_accuracy / torch.tensor(len(test_dataloader.dataset),
dtype=torch.float) * torch.tensor(100.0)
training_accuracy = training_accuracy / torch.tensor(len(training_dataloader.dataset),
dtype=torch.float) * torch.tensor(100.0)
test_loss = test_loss / torch.tensor(gpus, dtype=torch.float)
epoch_losses.append(average_loss.item())
training_accuracies.append(training_accuracy.item())
test_losses.append(test_loss.item())
test_accuracies.append(test_accuracy.item())
learning_rates.append((optimizer.param_groups[0])["lr"])
print(f"\nBatch size: {batch_size * int(gpus)}")
print(f"average Training Loss: {average_loss.item():.6f}")
print(f"average Test Loss: {test_loss.item():.6f}")
print(f"\naverage Training Acc: {training_accuracy.item():.6f}")
print(f"average Test Acc: {test_accuracy.item():.6f}")
printLearningRate(optimizer)
scheduler.step(test_loss)
# saving model checkpoint
save_checkpoint(model, optimizer, scheduler, i, epoch_losses, training_accuracies, test_losses, test_accuracies, learning_rates,
os.path.join(dir_path, f"epoch{i}.pth"), {key: value for key, value in batchwise_results[i].items() if key == 'train'}, {key: value for key, value in batchwise_results[i].items() if key == 'test'}, train_conf_mat, test_conf_mat)
#TODO: implement ONNX Export
# early stopping scheduler
if early_stopping(test_losses) == True:
counter += 1
print(f"Early Stopping counter: {counter} of {patience}")
else:
counter += 0
if counter < patience:
pass
else:
print("\n\nEarly Stopping activated")
print(f"Training stopped at Epoch{i + 1}")
dist.destroy_process_group()
exit()
def train(gpu, gpus, world_size):
torch.manual_seed(0)
torch.cuda.set_device(gpu)
try:
dist.init_process_group(backend='nccl', world_size=world_size, rank=gpu) #for distributed GPU training
except RuntimeError:
print("\n\nINFO:RuntimeError is raised >> Used gloo backend instead of nccl!\n")
dist.init_process_group(backend='gloo', world_size=world_size, rank=gpu) #as a fallback option
dir_path = None
if gpu == 0:
dir_path = "stackgraphConvPool3DPnet"
createdir(dir_path)
training_number = next_training_number(dir_path)
dir_path = os.path.join(dir_path, f"train{training_number}")
createdir(dir_path)
#save hyper-parameters in txt protocol file
save_hyperparameters(dir_path, 'hyperparameters.txt')
print("\nINFO: Protocol File saved successfully . . .")
model = Classifier(shrinkingLayers, mlpClassifier)
torch.cuda.set_device(gpu)
model.cuda(gpu)
#setting up optimizer
if optimizer_str == "SGD":
optimizer = torch.optim.SGD(model.parameters(), learning_rate, momentum=momentum, weight_decay=weight_decay)
elif optimizer_str == "RMSprop":
optimizer = torch.optim.RMSprop(model.parameters(), learning_rate, weight_decay=weight_decay)
else:
optimizer = torch.optim.Adam(model.parameters(), learning_rate, weight_decay=weight_decay)
# single-program multiple-data training paradigm (Distributed Data-Parallel Training)
model = DDP(model, device_ids=[gpu])
if dimensionality == 3:
training_data = ModelNet("ModelNet10_train_data", transform=lambda x: NormalizeScale()(SamplePoints(num=sample_points)(x)))
else:
training_data = ModelNet("ModelNet10_train_data", transform=lambda x: NormalizeScale()(NormalizeRotation()(SamplePoints(num=sample_points, remove_faces=True, include_normals=True)(x))))
training_sampler = DistributedWeightedSampler(training_data, num_replicas=world_size) #weight unbalanced classes by 1/cls_count
training_dataloader = DataLoader(dataset=training_data, batch_size=batch_size, shuffle=data_shuffle, num_workers=0,
pin_memory=True, sampler=training_sampler)
if dimensionality == 3:
test_data = ModelNet("ModelNet10_test_data", train=False, transform=lambda x: NormalizeScale()(SamplePoints(num=sample_points)(x)))
else:
test_data = ModelNet("ModelNet10_test_data", train=False, transform=lambda x: NormalizeScale()(NormalizeRotation()(SamplePoints(num=sample_points, remove_faces=True, include_normals=True)(x))))
test_sampler = DistributedWeightedSampler(test_data, num_replicas=world_size) #weight unbalanced classes by 1/cls_count
test_dataloader = DataLoader(dataset=test_data, batch_size=batch_size, shuffle=data_shuffle, num_workers=0,
pin_memory=True, sampler=test_sampler)
# weighted CE Loss over all Classes C
class_sample_count = np.array([len(np.where(training_data.data.y == t)[0]) for t in np.unique(training_data.data.y)])
weight = 1. / class_sample_count
weight = torch.from_numpy(weight)
weight = weight.float()
loss_fn = nn.CrossEntropyLoss(weight=weight).cuda(gpu)
# continue training from certain checkpoint
continue_from_scratch = True if args.resume is None else False
if continue_from_scratch:
if gpu == 0:
print("\nINFO: Train from scratch has started . . .")
train_optimisation(gpu, gpus, training_dataloader, test_dataloader, model, loss_fn, optimizer, None, dir_path, 0)
else:
checkpoint_path = "stackgraphConvPool3DPnet/" + args.resume
if gpu == 0:
print(f"\nINFO: Train has started from certain checkpoint {checkpoint_path.split('/')[2].split('.')[0]} in {checkpoint_path.split('/')[1]} . . .")
model.load_state_dict(torch.load(checkpoint_path)['model_state_dict'], strict=False)
optimizer.load_state_dict(torch.load(checkpoint_path)['optimizer_state_dict'])
final_epoch = (torch.load("stackgraphConvPool3DPnet/" + args.resume)['epoch'])+1
train_optimisation(gpu, gpus, training_dataloader, test_dataloader, model, loss_fn, optimizer, None, dir_path, final_epoch)
code tools:
class KMeansInitMostDistantFromMean:
def __call__(self, *args, **kwargs):
X, k = args
mean = np.mean(X, axis=0)
arg_sorted = np.argsort(np.apply_along_axis(lambda y: euclidean(mean, y), 1, X))
output = X[np.flip(arg_sorted)[:k]]
return output
class KMeansInit:
def __call__(self, *args, **kwargs):
X, k = args
current_centroids = np.expand_dims(np.mean(X, axis=0), 0)
for i in range(k - 1):
X, current_centroids = self.next_centroid(X, current_centroids)
return current_centroids
def next_centroid(self, X, curr_centroids):
highest_dist = 0.0
next_centroid = None
next_centroid_index = None
for i, x in enumerate(X):
max_dist = np.amax(np.apply_along_axis(lambda y: euclidean(x, y), 1, curr_centroids))
if max_dist > highest_dist:
next_centroid = x
highest_dist = max_dist
next_centroid_index = i
return np.delete(X, next_centroid_index, 0), np.append(curr_centroids, np.expand_dims(next_centroid, 0), 0)
class Conv(gnn.MessagePassing):
def __init__(self, sigma: nn.Module, F: nn.Module, W: nn.Module, M: nn.Module, C: int, P: int):
super().__init__(aggr="mean")
self.sigma = sigma
self.F = F
self.W = W
self.M = M
self.C = C
self.P = P
self.B = torch.randn(C+P, requires_grad=True)
def forward(self, feature_matrix, edge_index):
return self.propagate(edge_index, feature_matrix=feature_matrix)
def message(self, feature_matrix_i, feature_matrix_j):
message = self.F(feature_matrix_j - feature_matrix_i)
message = message.view(-1, self.C + self.P, self.C)
feature_matrix_i_ = feature_matrix_i.unsqueeze(2)
output = torch.bmm(message, feature_matrix_i_).squeeze()
return output
def update(self, aggr_out, feature_matrix):
Weight = self.M(aggr_out)
aggr_out = aggr_out * Weight
transform = self.W(feature_matrix)
transform = transform.view(-1, self.C + self.P, self.C)
feature_matrix = feature_matrix.unsqueeze(2)
transformation = torch.bmm(transform, feature_matrix).squeeze()
aggr_out = aggr_out + transformation
output = aggr_out + self.B
output = self.sigma(output)
return output
class Aggregation(nn.Module):
def __init__(self, mlp1: nn.Module, mlp2: nn.Module):
super().__init__()
self.mlp1 = mlp1
self.mlp2 = mlp2
self.softmax = nn.Softmax(0)
def forward(self, feature_matrix_batch: torch.Tensor, conv_feature_matrix_batch: torch.Tensor):
N, I, D = feature_matrix_batch.size()
N_, I_, D_ = conv_feature_matrix_batch.size()
augmentation = D_ - D
if augmentation > 0:
feature_matrix_batch = F.pad(feature_matrix_batch, (0, augmentation))
S1 = torch.mean(feature_matrix_batch, 1)
S2 = torch.mean(conv_feature_matrix_batch, 1)
Z1 = self.mlp1(S1)
Z2 = self.mlp2(S2)
M = self.softmax(torch.stack((Z1, Z2), 0))
M1 = M[0]
M2 = M[1]
M1 = M1.unsqueeze(1).expand(-1, I, -1)
M2 = M2.unsqueeze(1).expand(-1, I, -1)
output = (M1 * feature_matrix_batch) + (M2 * conv_feature_matrix_batch)
return output
class MaxPool(nn.Module):
def __init__(self, k: int):
super().__init__()
self.k = k
def forward(self, feature_matrix_batch: torch.Tensor, cluster_index: torch.Tensor):
N, I, D = feature_matrix_batch.size()
feature_matrix_batch = feature_matrix_batch.view(-1, D)
output = scatter_max(feature_matrix_batch, cluster_index, dim=0)[0]
output = output.view(N, self.k, -1)
return output
class GraphConvPool3DPnet(nn.Module):
def __init__(self, shrinkingLayers: [ShrinkingUnit], mlp: nn.Module):
super().__init__()
self.neuralNet = nn.Sequential(*shrinkingLayers, mlp)
def forward(self, x: torch.Tensor, pos: torch.Tensor):
feature_matrix_batch = torch.cat((pos, x), 2) if x is not None else pos
return self.neuralNet(feature_matrix_batch)
class ShrinkingUnitStack(nn.Module):
def __init__(self, input_stack: int, stack_fork: int, mlp: nn.Module, learning_rate: int, k: int, kmeansInit, n_init, sigma: nn.Module, F: nn.Module, W: nn.Module,
M: nn.Module, C, P, mlp1: nn.Module, mlp2: nn.Module):
super().__init__()
self.stack_fork = stack_fork
stack_size = input_stack * stack_fork
self.selfCorrStack = SelfCorrelationStack(stack_size, mlp, learning_rate)
self.kmeansConvStack = KMeansConvStack(stack_size, k, kmeansInit, n_init, sigma, F, W, M, C, P)
self.localAdaptFeaAggreStack = AggregationStack(stack_size, mlp1, mlp2)
self.graphMaxPoolStack = MaxPoolStack(stack_size, k)
def forward(self, feature_matrix_batch):
feature_matrix_batch = torch.repeat_interleave(feature_matrix_batch, self.stack_fork, dim=0)
feature_matrix_batch = self.selfCorrStack(feature_matrix_batch)
feature_matrix_batch_, conv_feature_matrix_batch, cluster_index = self.kmeansConvStack(feature_matrix_batch)
feature_matrix_batch = self.localAdaptFeaAggreStack(feature_matrix_batch, conv_feature_matrix_batch)
output = self.graphMaxPoolStack(feature_matrix_batch, cluster_index)
return output
class SelfCorrelationStack(nn.Module):
def __init__(self, stack_size: int, mlp: nn.Module, learning_rate: int = 1.0):
super().__init__()
self.selfCorrelationStack = nn.ModuleList([SelfCorrelation(copy.deepcopy(mlp), learning_rate) for i in range(stack_size)])
self.apply(init_weights)
def forward(self, feature_matrix_batch: torch.Tensor):
# feature_matrix_batch size = (S,N,I,D) where S=stack_size, N=batch number, I=members, D=member dimensionality
output = selfCorrThreader(self.selfCorrelationStack, feature_matrix_batch)
# output size = (S,N,I,D) where where S=stack_size, N=batch number, I=members, D=member dimensionality
return output
class KMeansConvStack(nn.Module):
def __init__(self, stack_size: int, k: int, kmeansInit, n_init: int, sigma: nn.Module, F: nn.Module, W: nn.Module,
M: nn.Module, C: int, P: int):
super().__init__()
self.kmeansConvStack = nn.ModuleList([
KMeansConv(k, kmeansInit, n_init, copy.deepcopy(sigma), copy.deepcopy(F), copy.deepcopy(W),
copy.deepcopy(M), C, P) for i in range(stack_size)])
self.apply(init_weights)
def forward(self, feature_matrix_batch: torch.Tensor):
# feature_matrix_batch size = (S,N,I,D) where S=stack size, N=batch number, I=members, D=member dimensionality
feature_matrix_batch, conv_feature_matrix_batch, cluster_index = kmeansConvThreader(self.kmeansConvStack,
feature_matrix_batch)
return feature_matrix_batch, conv_feature_matrix_batch, cluster_index
class AggregationStack(nn.Module):
def __init__(self, stack_size: int, mlp1: nn.Module, mlp2: nn.Module):
super().__init__()
self.localAdaptFeatAggreStack = nn.ModuleList([Aggregation(copy.deepcopy(mlp1), copy.deepcopy(mlp2)) for i
in range(stack_size)])
self.apply(init_weights)
def forward(self, feature_matrix_batch: torch.Tensor, conv_feature_matrix_batch: torch.Tensor):
output = threader(self.localAdaptFeatAggreStack, feature_matrix_batch, conv_feature_matrix_batch)
return output
class MaxPoolStack(nn.Module):
def __init__(self, stack_size: int, k: int):
super().__init__()
self.graphMaxPoolStack = nn.ModuleList([MaxPool(k) for i in range(stack_size)])
self.apply(init_weights)
def forward(self, feature_matrix_batch: torch.Tensor, cluster_index: torch.Tensor):
output = threader(self.graphMaxPoolStack, feature_matrix_batch, cluster_index)
return output
def selfCorrThreader(modules, input_tensor):
list_append = []
threads = []
for i, t in enumerate(input_tensor):
threads.append(Thread(target=selfCorrAppender, args=(modules[i], t, list_append, i)))
[t.start() for t in threads]
[t.join() for t in threads]
list_append.sort()
list_append = list(map(lambda x: x[1], list_append))
return torch.stack(list_append)
def selfCorrAppender(module, tensor, list_append, index):
list_append.append((index, module(tensor)))
def kmeansConvThreader(modules, input_tensor):
list1_append = []
list2_append = []
list3_append = []
threads = []
for i, t in enumerate(input_tensor):
threads.append(
Thread(target=kmeansAppender, args=(modules[i], t, list1_append, list2_append, list3_append, i)))
[t.start() for t in threads]
[t.join() for t in threads]
list1_append.sort()
list2_append.sort()
list3_append.sort()
list1_append = list(map(lambda x: x[1], list1_append))
list2_append = list(map(lambda x: x[1], list2_append))
list3_append = list(map(lambda x: x[1], list3_append))
return torch.stack(list1_append), torch.stack(list2_append), torch.stack(list3_append)
def kmeansAppender(module, input, list1_append, list2_append, list3_append, index):
x, y, z = module(input)
list1_append.append((index, x))
list2_append.append((index, y))
list3_append.append((index, z))
def threader(modules, input_tensor1, input_tensor2):
list_append = []
threads = []
for i, t in enumerate(input_tensor1):
threads.append(Thread(target=threaderAppender, args=(modules[i], t, input_tensor2[i], list_append, i)))
[t.start() for t in threads]
[t.join() for t in threads]
list_append.sort()
list_append = list(map(lambda x: x[1], list_append))
return torch.stack(list_append)
def threaderAppender(module, t1, t2, list_append, index):
list_append.append((index, module(t1, t2)))
class Classifier(nn.Module):
def __init__(self, shrinkingLayersStack: [ShrinkingUnitStack], mlp: nn.Module):
super().__init__()
self.neuralNet = nn.Sequential(*shrinkingLayersStack)
self.mlp = mlp
def forward(self, x: torch.Tensor, pos: torch.Tensor):
feature_matrix_batch = pos.unsqueeze(0)
output = self.neuralNet(feature_matrix_batch)
output = torch.mean(output, dim=0)
return self.mlp(output)
Error:
thank you for your help

The attribute labels_ of a KMeans object is created once you actually compute the clusters by running .fit() (or .fit_predict(), or .fit_transform()).
Simple example:
>>> from sklearn.cluster import KMeans
>>> from numpy.random import random
>>> X = random((10,2))
>>> X
array([[0.2096706 , 0.69704806],
[0.31732618, 0.29607599],
[0.10372159, 0.56911046],
[0.30922255, 0.07952464],
[0.21190404, 0.46823665],
[0.67134948, 0.95702692],
[0.14781526, 0.24619197],
[0.89931979, 0.96301003],
[0.88256126, 0.07569739],
[0.70776912, 0.92997521]])
>>> clustering = KMeans(n_clusters=3)
>>> clustering.labels_
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
AttributeError: 'KMeans' object has no attribute 'labels_'
>>> clustering.fit(X)
KMeans(n_clusters=3)
>>> clustering.labels_
array([0, 0, 0, 0, 0, 1, 0, 1, 2, 1], dtype=int32)

How to use torch.nn.transformer with pytroch lightning?

I am trying to use the vanilla transformer from PyTorch using Pytorch Lightning. I tried to test the model with a reverse number task. So given [1, 3, 5, 4, 13, 19] it returns [1, 13, 4, 5, 3, 19] with 1, 19 being start and end token respectively. The full code is below. The code can run without error but there seems to be a problem with the backpropagation. The training loss does go down at first but it doesn't go beyond 2.8 and the accuracy doesn't go beyond 11%.
It seems that part of the model is able to optimize, I am guessing it is because the weights located in Embeddings and Generator can backpropagate, but weights located in nn.Transformer cannot? I am really not sure.
import math
import torch.nn.functional as F
import numpy as np
import torch
import torch.nn as nn
from torch.utils.data import DataLoader
import pytorch_lightning as pl
from pytorch_lightning.callbacks import EarlyStopping
class Embeddings(pl.LightningModule):
def __init__(self, d_model, vocab):
super(Embeddings, self).__init__()
self.lut = nn.Embedding(vocab, d_model)
self.d_model = d_model
def forward(self, x):
a = self.lut(x) * math.sqrt(self.d_model)
return a
class PositionalEncoding(pl.LightningModule):
def __init__(self, d_model, dropout, max_len=5000):
super(PositionalEncoding, self).__init__()
self.dropout = nn.Dropout(p=dropout)
# Compute the positional encodings once in log space.
pe = torch.zeros(max_len, d_model)
position = torch.arange(0, max_len).unsqueeze(1)
div_term = torch.exp(torch.arange(0, d_model, 2) *
-(math.log(10000.0) / d_model))
pe[:, 0::2] = torch.sin(position * div_term)
pe[:, 1::2] = torch.cos(position * div_term)
pe = pe.unsqueeze(0)
self.register_buffer('pe', pe)
def forward(self, x):
x = x + self.pe[:, :x.size(1)]
return self.dropout(x)
class Generator(pl.LightningModule):
def __init__(self, size):
super(Generator, self).__init__()
self.proj = nn.Linear(512, size)
def forward(self, x):
return F.log_softmax(self.proj(x), dim=-1)
class Model(pl.LightningModule):
def __init__(self, src_embed, tgt_embed, transformer, generator):
super(Model, self).__init__()
self.src_embed = src_embed
self.tgt_embed = tgt_embed
self.transformer = transformer
self.generator = generator
self.valLoss = 0
self.valAcc = 0
self.automatic_optimization = False
self.optimizer = None
for p in self.parameters():
if p.dim() > 1:
nn.init.xavier_uniform_(p)
def forward(self, x, y, tgt_mask=None):
x = self.src_embed(x)
y = self.tgt_embed(y)
return self.generator(self.transformer(x, y, tgt_mask=tgt_mask))
def training_step(self, batch, batch_idx):
if self.optimizer is None:
self.optimizer = self.optimizers()
batch = Batch(batch[0], batch[1])
tgt_mask = batch.trg_mask.squeeze(0)
tgt_mask = (tgt_mask != True)
output = self(batch.src, batch.trg, tgt_mask)
criterion = LossCompute(V)
loss = criterion.forward(output.contiguous().view(-1, output.size(-1)), batch.trg_y.contiguous().view(-1)) / batch.ntokens
loss.backward()
self.optimizer.step()
self.optimizer.zero_grad()
self.log('train_loss', loss)
print(loss)
def validation_step(self, batch, batch_idx):
batch = Batch(batch[0], batch[1])
tgt_mask = batch.trg_mask.squeeze(0)
tgt_mask = (tgt_mask != True)
output = self(batch.src, batch.trg, tgt_mask)
criterion = LossCompute(V)
loss = criterion.forward(output.view(-1, output.size(-1)), batch.trg_y.contiguous().view(-1)) / batch.ntokens
self.log('val_loss', loss)
self.valLoss += loss
if batch_idx % 10 == 0:
print(loss)
if batch_idx == 99:
print(self.valLoss/100)
self.valLoss = 0
return {"x": output, "trg": batch.trg_y, "index": batch_idx}
def validation_step_end(self, batch):
output, trg, idx = batch["x"], batch["trg"], batch["index"]
accuracy = getAccuracy(output, trg)
self.log("accuracy", accuracy)
self.valAcc += accuracy
if idx == 99:
print(self.valAcc/100)
self.valAcc = 0
def train_dataloader(self):
data = data_gen(V, 0, 3000)
return DataLoader(data, batch_size=30, shuffle=False, num_workers=2, pin_memory=True)
def val_dataloader(self):
data = data_gen(V, 1, 1000)
return DataLoader(data, batch_size=10, shuffle=False, num_workers=2, pin_memory=True)
def configure_optimizers(self):
return torch.optim.Adam(self.parameters(), lr=1e-3, betas=(0.9, 0.98), eps=1e-9)
class LossCompute(pl.LightningModule):
def __init__(self, size):
super(LossCompute, self).__init__()
self.criterion = nn.KLDivLoss(reduction='sum')
self.size = size
self.true_dist = None
def forward(self, x, target):
# x has size (batch_size x length, vocab_size)
assert x.size(1) == self.size
true_dist = x.data.clone()
true_dist.fill_(0)
true_dist.scatter_(1, target.data.unsqueeze(1).long(), 1)
self.true_dist = true_dist
return self.criterion(x, true_dist)
# prepare data
class Batch:
"Object for holding a batch of data with mask during training."
def __init__(self, src, trg=None):
self.src = src
if trg is not None:
self.trg = trg[:, :-1]
self.trg_y = trg[:, 1:]
self.trg_mask = \
self.make_std_mask(self.trg)
self.ntokens = self.trg_y.size(0) * self.trg_y.size(1)
print("")
#staticmethod
def make_std_mask(tgt):
"Create a mask to hide padding and future words."
tgt_mask = subsequent_mask(tgt.size(-1)).type_as(tgt.data)
return tgt_mask
def subsequent_mask(size):
"Mask out subsequent positions."
attn_shape = (1, size, size)
subsequent_mask = np.triu(np.ones(attn_shape), k=1).astype('uint8')
return torch.from_numpy(subsequent_mask) == 0
def data_gen(V, randomSeed, totalTrainingSample):
np.random.seed(randomSeed)
x = torch.from_numpy(np.random.randint(2, V - 2, size=(totalTrainingSample, 10)))
y = torch.flip(torch.flip(x, [0, 1]), [0])
x[:, 0] = 1
y[:, 0] = 1
x[:, -1] = V - 1
y[:, -1] = V - 1
return list(zip(x, y))
def getAccuracy(x, trg):
totalValAcc = 0
totalValAccToken = 0
trg = trg.contiguous().view(-1)
out = x.view(-1, x.size(-1)) # (batch_size * tgt_length, src_vocab)
_, index = torch.max(out, dim=-1) # index (batch_size * tgt_length)
correct = list((trg == index)).count(True)
totalValAcc += correct
totalValAccToken += index.size(0)
return totalValAcc / totalValAccToken
V = 20
transformer = nn.Transformer(num_encoder_layers=2, num_decoder_layers=2, batch_first=True)
PositionEnc = PositionalEncoding(512, 0.1)
src_emb = Embeddings(512, V)
tgt_emb = Embeddings(512, V)
gen = Generator(V)
if __name__ == '__main__':
model = Model(nn.Sequential(src_emb, PositionEnc), nn.Sequential(tgt_emb, PositionEnc), transformer, gen)
earlyStopping = EarlyStopping(monitor='val_loss', patience=3)
trainer = pl.Trainer(max_epochs=10, callbacks=[earlyStopping])
trainer.fit(model)

error X = X.reshape(1, X.shape[0]) IndexError: tuple index out of range. How to fix that?

this is fragment of my code
def train(self, features, targets):
for X, y in zip(features, targets):
X = X.reshape(1, X.shape[0])
outputs = self.feed_forward(X)
when I try to use the method with data:
train(np.array([gameDataList[n].ball_position, gameDataList[n].wall_position]), np.array(gameDataList[n].upOrDown))
where gameDataList[n].upOrDown is an array e.g. [0.1, 0.9], and gameDataList[n].ball_position and gameDataList[n].wall_position are floats, I get this error.
Full code:
#### Imports ####
import numpy as np
#### Neural Network Class ####
class MLP:
##### Constructor ####
def __init__(self, n_input_nodes, hidden_nodes, n_output_nodes, lr):
## Network ##
self.n_input_nodes = n_input_nodes
self.n_output_nodes = n_output_nodes
self.nodes = hidden_nodes
self.nodes.insert(0, n_input_nodes)
self.nodes.append(n_output_nodes)
## Weights and Biases##
self.weights = []
self.biases = []
for i in range(1, len(self.nodes)):
self.weights.append(np.random.uniform(-1.0, 1.0, (self.nodes[i - 1], self.nodes[i])))
self.biases.append(np.random.uniform(-1.0, 1.0, (1, self.nodes[i])))
## Learning Rate ##
self.lr = lr
## Activation Functions ##
# Linear Activation
self.linear = lambda x: x
self.d_linear = lambda x: np.ones(x.shape)
# Relu Activation
def relu(x):
x[x < 0] = 0
return x
def d_relu(out):
out: x[x > 0] = 1
return out
self.relu = relu
self.d_relu = d_relu
# Sigmoid Activation
self.sigmoid = lambda x: 1 / (1 + np.exp(-x))
self.d_sigmoid = lambda out: out * (1 - out) # assumes out is tanh(x)
# Hyperbolic Tangent Activation
self.tanh = lambda x: np.tanh(x)
self.d_tanh = lambda out: 1 - out ** 2 # assumes out is tanh(x)
def getWeights(self):
return self.weights.copy()
def getBiases(self):
return self.biases.copy()
def setWeights(self, weights):
self.weights = weights.copy()
def setBiases(self, biases):
self.biases = biases.copy()
#### Feed Forward ####
def feed_forward(self, X):
outputs = [X]
logits = np.dot(X, self.weights[0]) + self.biases[0]
for i in range(1, len(self.nodes) - 1):
out = self.sigmoid(logits)
outputs.append(out)
logits = np.dot(out, self.weights[i]) + self.biases[i]
out = self.sigmoid(logits)
outputs.append(out)
return outputs
#### Backpropagation ####
def backpropagation(self, X, y, outputs):
weights_gradients = []
biases_gradients = []
d1 = y - outputs[-1]
d2 = self.d_sigmoid(outputs[-1])
error = d1 * d2
grad = outputs[-2].T * error
weights_gradients.append(grad)
biases_gradients.append(error)
for i in range(len(self.weights) - 2, 1, -1):
d = self.d_sigmoid(outputs[i])
error = np.dot(error, self.weights[i + 1].T) * d
grad = outputs[i - 1].T * error
weights_gradients.append(grad)
biases_gradients.append(error)
return weights_gradients, biases_gradients
#### Training ####
def train(self, features, targets):
# Batch Size for weight update step
batch_size = features.shape[0]
# Delta Weights Variables
delta_weights = [np.zeros(weight.shape) for weight in self.weights]
delta_biases = [np.zeros(bias.shape) for bias in self.biases]
# For every data point, forward pass, backpropogation, store weights change
for X, y in zip(features, targets):
# Forward pass
X = X.reshape(1, X.shape[0])
outputs = self.feed_forward(X)
# Back propogation
weights_gradients, biases_gradients = self.backpropagation(X, y, outputs)
for i in range(len(weights_gradients)):
delta_weights[-(i + 1)] += weights_gradients[i]
delta_biases[-(i + 1)] += biases_gradients[i]
for i in range(len(delta_weights)):
self.weights[i] += (self.lr * delta_weights[i]) / batch_size
self.biases[i] += (self.lr * delta_biases[i]) / batch_size
#### Testing Methods ####
def predict(self, X):
# Gives prediction
return self.feed_forward(X)[-1]
def test(self, features, targets):
predictions = self.predict(features)
n_correct = 0
for i in range(len(predictions)):
prediction = np.argmax(predictions[i])
correct = np.argmax(targets[i])
if prediction == correct:
n_correct += 1
return n_correct / len(targets)
class GameData:
def __init__(self, ball_position, wall_position, upOrDown):
self.wall_position = wall_position
self.ball_position = ball_position
self.upOrDown = upOrDown
I collect data, and train my network, in this way:
gameDataList.append(GameData(ball.trt.ycor(), b.trt.ycor(), [0.1, 0.9]))
mlp = MLP(2, [32, 32], 2, 0.0001)
n = random.randint(0, 999)
mlp.train(np.array([gameDataList[n].ball_position, gameDataList[n].wall_position]), np.array(gameDataList[n].upOrDown))

Problem solved. It was needed to write two square brackets instead of one.
wrong example:
np.array([gameDataList[n].ball_position, gameDataList[n].wall_position])
correct example:
np.array([[gameDataList[n].ball_position, gameDataList[n].wall_position]])

Why gradient check gives high difference (almost 1)?

I'm trying to implement a Neural Net in python without the use of libraries like Keras or Tensorflow. I still have to test the net, right now I just tried to train it on Iris dataset and check afterwards the correctness of the backpropagation algorithm.
To do so, I wrote the gradient checking procedure, calculating the analytical gradients and comparing them with the gradients from backpropagation.
The point is that, even if the backpropagation algorithm seems correct to me, the difference between the gradients is always high (around 0.8, instead of the classic 1e-7).
Layer class
class Dense(Layer):
def __init__(self, input_shape, name=None, activation='relu', regularization='l2'):
self.name = name
self.is_output = False
self.weights = np.random.uniform(low=0.01, high=0.10, size=input_shape)
self.biases = np.ones((1,input_shape[1]))
if activation == 'sigmoid':
self.activation = Activation_Sigmoid()
else: #activation == 'relu':
self.activation = Activation_ReLU()
self.cost = Categorical_CrossEntropyLoss()
def set_as_output(self, is_output=True):
self.is_output = is_output
def forward(self, inputs, debug=False, epsilon=None):
self.net_input = inputs
if debug:
augmented_parameters = np.zeros(epsilon.shape)
weights_column_vector = np.reshape(self.weights,(-1,1))
biases_column_vector = np.reshape(self.biases,(-1,1))
concatenated_parameters = np.concatenate((weights_column_vector, biases_column_vector))
for i in range(concatenated_parameters.shape[0]):
augmented_parameters[i] = concatenated_parameters[i]
# make the augmented parameter long as theta in order to sum them
# this because epsilon is a standard basis vector
augmented_parameters += epsilon
# rebuild the weights matrix and biases vector to apply forward propagation
weights_end = self.weights.shape[0] * self.weights.shape[1]
biases_end = self.biases.shape[0] * self.biases.shape[1] + weights_end
weights = np.reshape(augmented_parameters[0:weights_end],self.weights.shape)
biases = np.reshape(augmented_parameters[weights_end:biases_end], self.biases.shape)
output = np.dot(inputs, weights) + biases
activated_output = self.activation.forward(output)
return activated_output
self.output = np.dot(inputs, self.weights) + self.biases
self.activated_output = self.activation.forward(self.output)
return self.activated_output
def backward(self, X, y, output, step, l2=0.5): #backpropagation
m = X.shape[0] # number of examples
if self.is_output:
error = self.cost.backward(output, y) #(a_k - y_hat_k)
delta_k = self.activation.backward(self.output)* error
# net input for neuron k is a_j^(l-1)
grad = np.dot(self.net_input.T, delta_k)
#update weights with l2 regularization
self.grad_w = grad + (l2 / m)*self.weights
self.grad_b = np.sum(delta_k * 1,axis=0)
self.weights -= step * self.grad_w
self.biases -= step * self.grad_b
return np.dot(delta_k ,self.weights.T)
else:
delta_j = self.activation.backward(self.output) * output
grad = np.dot(self.net_input.T, delta_j)
self.grad_w = grad + (l2 / m) * self.weights
self.grad_b = np.sum(delta_j * 1, axis=0)
self.weights -= step * self.grad_w
self.biases -= step * self.grad_b
return np.dot(delta_j, self.weights.T)
def get_parameters(self):
return self.weights, self.biases
def get_gradients(self):
return self.grad_w, self.grad_b
Neural Net class
class NeuralNet():
def __init__(self):
self.layers = []
self.layers_output = []
self.cost = None
self.regularization = L2_Regularization()
def add(self,layer):
self.layers.append(layer)
def forward(self, inputs, debug=False, epsilon=None):
input = np.copy(inputs)
for layer in self.layers:
output = layer.forward(input, debug=debug, epsilon=epsilon)
input = output
return input
def backward(self, X, y, output, step):
prev_delta = None
out = output
for layer in self.layers[::-1]:
prev_delta = layer.backward(X, y, out, step)
out = prev_delta
def fit(self, X, y, batch_size=1, epochs=10, step=0.05, shuffle=True):
self.layers[-1].set_as_output()
self.error = []
i = 0.005 * epochs
for epoch in range(epochs):
if shuffle:
X = np.random.permutation(X)
batches = int(np.ceil(X.shape[0]/batch_size))
batches_error = []
for t in range(batches):
batch_X = X[t*batch_size:np.min([X.shape[0],(t+1)*batch_size]),:]
batch_y = y[t*batch_size:np.min([y.shape[0],(t+1)*batch_size]),:]
output = self.forward(batch_X)
cost = self.cost.forward(output,batch_y)
cost += self.regularization.forward(X, self.layers)
batches_error.append(cost)
self.backward(batch_X, batch_y, output, step)
self.error.append(np.mean(batches_error))
if epoch % i == 0:
print('epoch:', epoch, 'error:', np.mean(self.error))
return self
def parameters_to_theta(self):
theta = []
for layer in self.layers:
w, b = layer.get_parameters()
#flatten parameter w
new_vector = np.reshape(w, (-1,1))
theta.append(new_vector)
#flatten parameter b
new_vector = np.reshape(b, (-1,1))
theta.append(new_vector)
return np.vstack(theta)
def gradients_to_theta(self):
theta = []
for layer in self.layers:
grad_w, grad_b = layer.get_gradients()
new_vector = np.reshape(grad_w, (-1,1))
theta.append(new_vector)
new_vector = np.reshape(grad_b, (-1,1))
theta.append(new_vector)
return np.vstack(theta)
def gradient_check(self, X, y, epsilon=1e-7):
theta = self.parameters_to_theta()
dtheta = self.gradients_to_theta()
num_parameters = theta.shape[0]
J_plus = np.zeros((num_parameters, 1))
J_minus = np.zeros((num_parameters, 1))
dtheta_approx = np.zeros((num_parameters, 1))
for i in range(num_parameters):
theta_plus = np.zeros((num_parameters,1))
theta_plus[i] = epsilon
J_plus[i] = self.cost.forward(self.forward(X, debug=True, epsilon=theta_plus),y)
theta_minus = np.zeros((num_parameters,1))
theta_minus[i] = - epsilon
J_minus[i] = self.cost.forward(self.forward(X, debug=True, epsilon=theta_minus),y)
dtheta_approx[i] = (J_plus[i] - J_minus[i])/ (2 * epsilon)
numerator = np.linalg.norm(dtheta - dtheta_approx)
denominator = np.linalg.norm(dtheta_approx) + np.linalg.norm(dtheta)
difference = numerator / denominator
return difference
I'm using ReLU and Sigmoid as activation functions, and Categorical Cross Entropy for the cost
import numpy as np
from scipy.special import expit as sigmoid
class Activation_ReLU:
def forward(self, inputs):
return np.maximum(0, inputs)
def backward(self, inputs):
return np.greater(inputs,0).astype(int)
class Activation_Sigmoid:
def forward(self, inputs):
return sigmoid(inputs)
def backward(self, inputs):
return sigmoid(inputs) * (1 - sigmoid(inputs))
class Categorical_CrossEntropyLoss():
def forward(self, y_pred, y_real):
predictions = np.copy(y_pred)
predictions = np.clip(predictions, 1e-12, 1 - 1e-12) # avoid zero values for log
n = y_real.shape[0]
return - (1 / n) * np.sum(y_real * np.log(y_pred))
def backward(self, y_pred, y_real):
return y_real - y_pred
These are the main classes that define the net. The model that I create to train on Iris dataset is a NN with 1 hidden layer.
# random seed is 1
X, y = load_iris(return_X_y=True)
X = (X - np.mean(X)) / np.std(X) # standardize data to improve network convergence
y = y.reshape((-1,1))
encoder = OneHotEncoder(sparse=False)
y = encoder.fit_transform(y)
X_train, X_test, y_train, y_test = train_test_split(X,y,train_size=0.8)
model = NeuralNet()
model.add(Dense((4,10),name='input_layer',activation='relu'))
model.add(Dense((10,10),name='hidden_layer',activation='relu'))
model.add(Dense((10,3),name='output_layer',activation='sigmoid'))
model.fit(X_train,y_train, batch_size=5, epochs=200, step=1e-3)
difference = model.gradient_check(X_train, y_train)
And then, the result of print(difference) is
0.7992920544491866
So there is something wrong with my implementation. What things I have to check to determine the causes of this high difference between gradients?

i can't understand this lambda functions

i am studying machine learning with python.
and this code is from Standford Uiv classes.
i was trying to grasp these codes but failed.
The problem is loss_W = lambda W: self.loss(x,t).
isn't it True that loss_W(1) or loss_W(2) or anything cannot change the result?
i can't understand that the results of these two codes are different.
grads['W1'] = numerical_gradient(loss_W, self.params['W1'])
grads['b1'] = numerical_gradient(loss_W, self.params['b1'])
def numerical_gradient(f, x):
h = 1e-4 # 0.0001
grad = np.zeros_like(x)
for idx in range(x.size):
tmp_val = x[idx]
# f(x+h)
x[idx] = float(tmp_val) + h
fxh1 = f(x)
# f(x-h)
x[idx] = tmp_val - h
fxh2 = f(x)
grad[idx] = (fxh1 - fxh2) / (2*h)
x[idx] = tmp_val
return grad
class TwoLayerNet:
def __init__(self, input_size, hidden_size, output_size, weight_init_std=0.01):
self.params = {}
self.params['W1'] = weight_init_std * np.random.randn(input_size, hidden_size)
self.params['b1'] = np.zeros(hidden_size)
self.params['W2'] = weight_init_std * np.random.randn(hidden_size, output_size)
self.params['b2'] = np.zeros(output_size)
def predict(self, x):
W1, W2 = self.params['W1'], self.params['W2']
b1, b2 = self.params['b1'], self.params['b2']
a1 = np.dot(x, W1)
z1 = sigmoid(a1)
a2 = np.dot(z1,W2)
y = softmax(a2)
return y
def loss(self, x, t):
y = self.predict(x)
return cross_entropy_error(y,t)
def accuracy(self, x,t):
y = self.predict(x)
y = np.argmax(y, axis=0)
t = np.argmax(t, axis=0)
data_len = len(x)
accuracy = np.sum(y==t)/float(data_len)
return accuracy
def numerical_gradient(self, x, t):
loss_W = lambda W: self.loss(x,t)
grads = {}
grads['W1'] = numerical_gradient(loss_W, self.params['W1'])
grads['b1'] = numerical_gradient(loss_W, self.params['b1'])
grads['W2'] = numerical_gradient(loss_W, self.params['W2'])
grads['b2'] = numerical_gradient(loss_W, self.params['b2'])
return grads

This lambda alone loss_W = lambda W: self.loss(x,t) is indifferent from the value of W. This function can be simplified like this:
x = 1 # Just some random value
t = 5 # Just some random value
def simplified_lambda_function(W):
return (x,t)
The code snippet you posted indicates that there is a class definition outside somewhere as
grads['W1'] = numerical_gradient(loss_W, self.params['W1'])
grads['b1'] = numerical_gradient(loss_W, self.params['b1']
self is undefined here. Because of this we cant be sure if they are truly identical, but most likely they are.