Develop Reference

Pytorch TypeError: scatter_add() takes from 2 to 5 positional arguments but 6 were given

Could someone please explain how to fix the situation where I take an example straight from the Pytorch documentation here:
import torch
from torch_geometric.datasets import TUDataset
from torch_geometric.data import Data, Dataset,DataLoader
dataset = TUDataset(root='data/TUDataset', name='MUTAG')
print()
print(f'Dataset: {dataset}:')
print('====================')
print(f'Number of graphs: {len(dataset)}')
print(f'Number of features: {dataset.num_features}')
print(f'Number of classes: {dataset.num_classes}')
data = dataset[0] # Get the first graph object.
print()
print(data)
print('=============================================================')
# Gather some statistics about the first graph.
print(f'Number of nodes: {data.num_nodes}')
print(f'Number of edges: {data.num_edges}')
print(f'Average node degree: {data.num_edges / data.num_nodes:.2f}')
#print(f'Has isolated nodes: {data.has_isolated_nodes()}')
#print(f'Has self-loops: {data.has_self_loops()}')
#print(f'Is undirected: {data.is_undirected()}')
torch.manual_seed(12345)
dataset = dataset.shuffle()
train_dataset = dataset[:150]
test_dataset = dataset[150:]
print(f'Number of training graphs: {len(train_dataset)}')
print(f'Number of test graphs: {len(test_dataset)}')
train_loader = DataLoader(train_dataset, batch_size=64, shuffle=True)
test_loader = DataLoader(test_dataset, batch_size=64, shuffle=False)
for step, data in enumerate(train_loader):
print(f'Step {step + 1}:')
print('=======')
print(f'Number of graphs in the current batch: {data.num_graphs}')
print(data)
print()
from torch.nn import Linear
import torch.nn.functional as F
from torch_geometric.nn import GCNConv
from torch_geometric.nn import global_mean_pool
class GCN(torch.nn.Module):
def __init__(self, hidden_channels):
super(GCN, self).__init__()
torch.manual_seed(12345)
self.conv1 = GCNConv(dataset.num_node_features, hidden_channels)
self.conv2 = GCNConv(hidden_channels, hidden_channels)
self.conv3 = GCNConv(hidden_channels, hidden_channels)
self.lin = Linear(hidden_channels, dataset.num_classes)
def forward(self, x, edge_index, batch):
# 1. Obtain node embeddings
x = self.conv1(x, edge_index)
x = x.relu()
x = self.conv2(x, edge_index)
x = x.relu()
x = self.conv3(x, edge_index)
# 2. Readout layer
x = global_mean_pool(x, batch) # [batch_size, hidden_channels]
# 3. Apply a final classifier
x = F.dropout(x, p=0.5, training=self.training)
x = self.lin(x)
return x
model = GCN(hidden_channels=64)
print(model)
model = GCN(hidden_channels=64)
optimizer = torch.optim.Adam(model.parameters(), lr=0.01)
criterion = torch.nn.CrossEntropyLoss()
def train():
model.train()
for data in train_loader: # Iterate in batches over the training dataset.
out = model(data.x, data.edge_index, data.batch) # Perform a single forward pass.
loss = criterion(out, data.y) # Compute the loss.
loss.backward() # Derive gradients.
optimizer.step() # Update parameters based on gradients.
optimizer.zero_grad() # Clear gradients.
def test(loader):
model.eval()
correct = 0
for data in loader: # Iterate in batches over the training/test dataset.
out = model(data.x, data.edge_index, data.batch)
pred = out.argmax(dim=1) # Use the class with highest probability.
correct += int((pred == data.y).sum()) # Check against ground-truth labels.
return correct / len(loader.dataset) # Derive ratio of correct predictions.
for epoch in range(1, 171):
train()
train_acc = test(train_loader)
test_acc = test(test_loader)
print(f'Epoch: {epoch:03d}, Train Acc: {train_acc:.4f}, Test Acc: {test_acc:.4f}')
I get the error:
out = model(data.x, data.edge_index, data.batch) # Perform a single forward pass.
File "/root/miniconda3/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
return forward_call(*input, **kwargs)
File "base_test.py", line 67, in forward
x = self.conv1(x, edge_index)
File "/root/miniconda3/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
return forward_call(*input, **kwargs)
File "/root/miniconda3/lib/python3.7/site-packages/torch_geometric/nn/conv/gcn_conv.py", line 103, in forward
return self.propagate(edge_index, x=x, norm=norm)
File "/root/miniconda3/lib/python3.7/site-packages/torch_geometric/nn/conv/message_passing.py", line 127, in propagate
out = scatter_(self.aggr, out, edge_index[i], dim, dim_size=size[i])
File "/root/miniconda3/lib/python3.7/site-packages/torch_geometric/utils/scatter.py", line 34, in scatter_
out = op(src, index, dim, None, dim_size, fill_value)
TypeError: scatter_add() takes from 2 to 5 positional arguments but 6 were given
I am using:
torch 1.11.0
torch-cluster 1.6.0
torch-geometric 1.3.2
torch-scatter 2.0.9
torch-sparse 0.6.13
torchmetrics 0.9.1

Just to mention, the answer was just to uninstall torch-scatter and then re-install the exact same version.

TypeError: 'GraphModule' object is not subscriptable in PyTorch quantization

Based on these tutorials, I want to apply torch.quantization API on Federated Learning:
https://colab.research.google.com/github/pytorch/tutorials/blob/gh-pages/_downloads/static_quantization_tutorial.ipynb#scrollTo=UBwjcUDJ7L_1
https://pytorch.org/tutorials/prototype/fx_graph_mode_ptq_static.html
But it seems that it doesn't test with FL yet! Although it supposes to be applicable "if the the code is symbolic traceable"! Based on a torch expert's answer here https://discuss.pytorch.org/t/quantization-in-federated-learning/152885
(what does that means?)
I tried to test it, but got an error! This is the code:
# model
class CNNCifar(nn.Module):
def __init__(self):
super(CNNCifar, self).__init__()
self.conv1 = nn.Conv2d(3, 6, 5)
self.pool = nn.MaxPool2d(2, 2)
self.conv2 = nn.Conv2d(6, 16, 5)
self.fc1 = nn.Linear(16 * 5 * 5, 120)
self.fc2 = nn.Linear(120, 84)
self.fc3 = nn.Linear(84, 100)
def forward(self, x):
x = self.pool(F.relu(self.conv1(x)))
x = self.pool(F.relu(self.conv2(x)))
x = x.reshape(-1, 16 * 5 * 5)
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = self.fc3(x)
return F.log_softmax(x, dim=1)
# training CIFAR100 dataset with this model is not shown here, but it works with no issues
# quantization
import torch.quantization.quantize_fx as quantize_fx
import copy
# client models quantize
c_model_q = []
for i in range(3):
model= client_models[i]
model.train()
model_to_quantize = copy.deepcopy(model)
qconfig_dict = {"": torch.quantization.get_default_qat_qconfig('qnnpack')}
model_prepared = quantize_fx.prepare_qat_fx(model_to_quantize, qconfig_dict)
c_model_q.append(model_prepared)
# golbal quantization
model_fp = global_model
# quantization aware training for static quantization
model_to_quantize = copy.deepcopy(model_fp)
qconfig_dict = {"": torch.quantization.get_default_qat_qconfig('qnnpack')} # or ...("fbgemm")
model_to_quantize.train()
# prepare
g_model_q = quantize_fx.prepare_qat_fx(model_to_quantize, qconfig_dict)
print(g_model_q.graph)
# train the model
# global model
g_model_q = g_model_q.cpu()
cmodel_q = []
for i in range(3):
cmodel = c_model_q[i]
cmodel.cpu()
cmodel_q.append(cmodel)
for model in cmodel_q:
model.load_state_dict(g_model_q.state_dict()) # initial synchronizing with global model
opt = [optim.SGD(model.parameters(), lr=0.1) for model in cmodel_q]
# helper function for local training in num_selected
def q_client_update(cmodel_q, optimizer, train_loader, epoch=5):
model.train()
for e in range(epoch):
for batch_idx, (data, target) in enumerate(train_loader):
data, target = data.cpu(), target.cpu()
optimizer.zero_grad()
output = cmodel_q(data)
loss = F.nll_loss(output, target)
loss.backward()
optimizer.step()
return loss.item()
# take the mean of the weights and aggregated into the global weights
def q_server_aggregate(g_model_q, cmodel_q):
global_dict = g_model_q.state_dict()
for k in global_dict.keys():
global_dict[k] = torch.stack([cmodel_q[i].state_dict()[k].float() for i in range(len(cmodel_q))], 0).mean(0)
g_model_q.load_state_dict(global_dict)
for model in cmodel_q:
model.load_state_dict(g_model_q.state_dict())
def q_test(g_model_q, test_loader):
model.eval()
test_loss = 0
correct = 0
with torch.no_grad():
for data, target in test_loader:
data, target = data.cpu(), target.cpu()
output = g_model_q(data)
test_loss += F.nll_loss(output, target, reduction='sum').item()
pred = output.argmax(dim=1, keepdim=True) # get the index of the max log-probability
correct += pred.eq(target.view_as(pred)).sum().item()
test_loss /= len(test_loader.dataset)
acc = correct / len(test_loader.dataset)
return test_loss, acc
losses_train = []
losses_test = []
acc_train = []
acc_test = []
# Runnining FL
for r in range(num_rounds):
# select random clients
client_idx = np.random.permutation(num_clients)[:num_selected]
# client update
loss = 0
for i in tqdm(range(num_selected)):
loss += q_client_update(cmodel_q[i], opt[i], train_loader[client_idx[i]], epoch=epochs)
losses_train.append(loss)
# server aggregate
q_server_aggregate(g_model_q, cmodel_q)
cmodel_q = quantize_fx.convert_fx(cmodel_q[i])
cmodel_q.eval()
g_model_q = quantize_fx.convert_fx(g_model_q)
g_model_q.eval()
test_loss, acc = q_test(g_model_q, test_loader)
losses_test.append(test_loss)
acc_test.append(acc)
print('%d-th round' % r)
print('average train loss %0.3g | test loss %0.3g | test acc: %0.3f' % (loss / num_selected, test_loss, acc))
100%|██████████| 3/3 [01:02<00:00, 20.82s/it]
/usr/local/lib/python3.7/dist-packages/torch/nn/quantized/_reference/modules/linear.py:41: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).
torch.tensor(weight_qparams["scale"], dtype=torch.float, device=device))
/usr/local/lib/python3.7/dist-packages/torch/nn/quantized/_reference/modules/linear.py:46: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).
dtype=torch.int, device=device))
0-th round
average train loss 3.3 | test loss 3.23 | test acc: 0.220
0%| | 0/3 [00:00<?, ?it/s]
---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
<ipython-input-21-0b68c58ee902> in <module>()
11 loss = 0
12 for i in tqdm(range(num_selected)):
---> 13 loss += q_client_update(cmodel_q[i], opt[i], train_loader[client_idx[i]], epoch=epochs)
14
15 losses_train.append(loss)
TypeError: 'GraphModule' object is not subscriptable
The model type before quantization is: __main__.CNNCifar
After quantization: torch.fx.graph_module.GraphModule.__new__.<locals>.GraphModuleImpl
What does object is not subscriptable mean? How I can fix it?

Python(PyTorch): TypeError: string indices must be integers

I have written the following code to train a bert model on my dataset but when I execute it I get an error at the part where I implement tqdm. I have written the entire training code below with full description of the error. How to fix this?
Code
Model
TRANSFORMERS = {
"bert-multi-cased": (BertModel, BertTokenizer, "bert-base-uncased"),
}
class Transformer(nn.Module):
def __init__(self, model, num_classes=1):
"""
Constructor
Arguments:
model {string} -- Transformer to build the model on. Expects "camembert-base".
num_classes {int} -- Number of classes (default: {1})
"""
super().__init__()
self.name = model
model_class, tokenizer_class, pretrained_weights = TRANSFORMERS[model]
bert_config = BertConfig.from_json_file(MODEL_PATHS[model] + 'bert_config.json')
bert_config.output_hidden_states = True
self.transformer = BertModel(bert_config)
self.nb_features = self.transformer.pooler.dense.out_features
self.pooler = nn.Sequential(
nn.Linear(self.nb_features, self.nb_features),
nn.Tanh(),
)
self.logit = nn.Linear(self.nb_features, num_classes)
def forward(self, tokens):
"""
Usual torch forward function
Arguments:
tokens {torch tensor} -- Sentence tokens
Returns:
torch tensor -- Class logits
"""
_, _, hidden_states = self.transformer(
tokens, attention_mask=(tokens > 0).long()
)
hidden_states = hidden_states[-1][:, 0] # Use the representation of the first token of the last layer
ft = self.pooler(hidden_states)
return self.logit(ft)
Training
def fit(model, train_dataset, val_dataset, epochs=1, batch_size=8, warmup_prop=0, lr=5e-4):
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)
val_loader = DataLoader(val_dataset, batch_size=batch_size, shuffle=False)
optimizer = AdamW(model.parameters(), lr=lr)
num_warmup_steps = int(warmup_prop * epochs * len(train_loader))
num_training_steps = epochs * len(train_loader)
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps, num_training_steps)
loss_fct = nn.BCEWithLogitsLoss(reduction='mean').cuda()
for epoch in range(epochs):
model.train()
start_time = time.time()
optimizer.zero_grad()
avg_loss = 0
for step, (x, y_batch) in tqdm(enumerate(train_loader), total=len(train_loader)):
y_pred = model(x.to(device))
loss = loss_fct(y_pred.view(-1).float(), y_batch.float().to(device))
loss.backward()
avg_loss += loss.item() / len(train_loader)
xm.optimizer_step(optimizer, barrier=True)
#optimizer.step()
scheduler.step()
model.zero_grad()
optimizer.zero_grad()
model.eval()
preds = []
truths = []
avg_val_loss = 0.
with torch.no_grad():
for x, y_batch in tqdm(val_loader):
y_pred = model(x.to(device))
loss = loss_fct(y_pred.detach().view(-1).float(), y_batch.float().to(device))
avg_val_loss += loss.item() / len(val_loader)
probs = torch.sigmoid(y_pred).detach().cpu().numpy()
preds += list(probs.flatten())
truths += list(y_batch.numpy().flatten())
score = roc_auc_score(truths, preds)
dt = time.time() - start_time
lr = scheduler.get_last_lr()[0]
print(f'Epoch {epoch + 1}/{epochs} \t lr={lr:.1e} \t t={dt:.0f}s \t loss={avg_loss:.4f} \t val_loss={avg_val_loss:.4f} \t val_auc={score:.4f}')
Error
---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
<timed eval> in <module>
<ipython-input-19-e47eae808597> in fit(model, train_dataset, val_dataset, epochs, batch_size, warmup_prop, lr)
22 for step, (x, y_batch) in tqdm(enumerate(train_loader), total=len(train_loader)):
23
---> 24 y_pred = model(x.to(device))
25
26 loss = loss_fct(y_pred.view(-1).float(), y_batch.float().to(device))
/opt/conda/lib/python3.6/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
724 result = self._slow_forward(*input, **kwargs)
725 else:
--> 726 result = self.forward(*input, **kwargs)
727 for hook in itertools.chain(
728 _global_forward_hooks.values(),
<ipython-input-11-2002cc7ec843> in forward(self, tokens)
41 )
42
---> 43 hidden_states = hidden_states[-1][:, 0] # Use the representation of the first token of the last layer
44
45 ft = self.pooler(hidden_states)
TypeError: string indices must be integers

Your code is designed for an older version of the transformers library:
AttributeError: 'str' object has no attribute 'dim' in pytorch
As such you will need to either downgrade to version 3.0.0, or adapt the code to deal with the new-format output of bert.

Pytorch: 1D target tensor expected, multi-target not supported

I want to train a 1D CNN on time series. I get the following error message 1D target tensor expected, multi-target not supported
Here is the code with simulated data corresponding to the structures of my data as well as the error message
import torch
from torch.utils.data import DataLoader
import torch.utils.data as data
import torch.nn as nn
import numpy as np
import random
from tqdm.notebook import tqdm
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(device)
train_dataset = []
n_item = 20
for i in range(0,n_item):
train_data = np.random.uniform(-10, 10, 500)
train_dataset.append(train_data)
train_dataset = np.asarray(train_dataset)
train_dataset.shape
ecg_train = torch.from_numpy(train_dataset).float()
labels_train = np.random.randint(2, size=n_item)
labels_train = torch.from_numpy(labels_train).long()
val_dataset = []
n_item = 10
for i in range(0,n_item):
val_data = np.random.uniform(-10, 10, 500)
val_dataset.append(val_data)
val_dataset = np.asarray(val_dataset)
val_dataset.shape
ecg_validation = torch.from_numpy(val_dataset).float()
labels_validation = np.random.randint(2, size=n_item)
labels_validation = torch.from_numpy(labels_validation).long()
class ECGNet(data.Dataset):
"""ImageNet Limited dataset."""
def __init__(self, ecgs, labls, transform=None):
self.ecg = ecgs
self.target = labls
self.transform = transform
def __getitem__(self, idx):
ecgVec = self.ecg[idx] #.reshape(10, -1)
labelID = self.target[idx].reshape(1)
return ecgVec,labelID
def __len__(self):
return len(self.ecg)
train_data = ECGNet(ecg_train,
labels_train,
)
print("size of Training dataset: {}".format(len(train_data)))
validation_data = ECGNet(ecg_validation,
labels_validation,
)
print("size of Training dataset: {}".format(len(validation_data)))
batch_size = 1
train_dataloader = DataLoader(dataset = train_data,
batch_size=batch_size,
shuffle = True,
num_workers = 0)
val_dataloader = DataLoader(dataset = validation_data,
batch_size=batch_size,
shuffle = True,
num_workers = 0)
def train_epoch(model, train_dataloader, optimizer, loss_fn):
losses = []
correct_predictions = 0
# Iterate mini batches over training dataset
for images, labels in tqdm(train_dataloader):
images = images.to(device)
#labels = labels.squeeze_()
labels = labels.to(device)
#labels = labels.to(device=device, dtype=torch.int64)
# Run predictions
output = model(images)
# Set gradients to zero
optimizer.zero_grad()
# Compute loss
loss = loss_fn(output, labels)
# Backpropagate (compute gradients)
loss.backward()
# Make an optimization step (update parameters)
optimizer.step()
# Log metrics
losses.append(loss.item())
predicted_labels = output.argmax(dim=1)
correct_predictions += (predicted_labels == labels).sum().item()
accuracy = 100.0 * correct_predictions / len(train_dataloader.dataset)
# Return loss values for each iteration and accuracy
mean_loss = np.array(losses).mean()
return mean_loss, accuracy
def evaluate(model, dataloader, loss_fn):
losses = []
correct_predictions = 0
with torch.no_grad():
for images, labels in dataloader:
images = images.to(device)
#labels = labels.squeeze_()
labels = labels.to(device=device, dtype=torch.int64)
# Run predictions
output = model(images)
# Compute loss
loss = loss_fn(output, labels)
# Save metrics
predicted_labels = output.argmax(dim=1)
correct_predictions += (predicted_labels == labels).sum().item()
losses.append(loss.item())
mean_loss = np.array(losses).mean()
accuracy = 100.0 * correct_predictions / len(dataloader.dataset)
# Return mean loss and accuracy
return mean_loss, accuracy
def train(model, train_dataloader, val_dataloader, optimizer, n_epochs, loss_function):
# We will monitor loss functions as the training progresses
train_losses = []
val_losses = []
train_accuracies = []
val_accuracies = []
for epoch in range(n_epochs):
model.train()
train_loss, train_accuracy = train_epoch(model, train_dataloader, optimizer, loss_fn)
model.eval()
val_loss, val_accuracy = evaluate(model, val_dataloader, loss_fn)
train_losses.append(train_loss)
val_losses.append(val_loss)
train_accuracies.append(train_accuracy)
val_accuracies.append(val_accuracy)
print('Epoch {}/{}: train_loss: {:.4f}, train_accuracy: {:.4f}, val_loss: {:.4f}, val_accuracy: {:.4f}'.format(epoch+1, n_epochs,
train_losses[-1],
train_accuracies[-1],
val_losses[-1],
val_accuracies[-1]))
return train_losses, val_losses, train_accuracies, val_accuracies
class Simple1DCNN(torch.nn.Module):
def __init__(self):
super(Simple1DCNN, self).__init__()
self.layer1 = torch.nn.Conv1d(in_channels=50,
out_channels=20,
kernel_size=5,
stride=2)
self.act1 = torch.nn.ReLU()
self.layer2 = torch.nn.Conv1d(in_channels=20,
out_channels=10,
kernel_size=1)
self.fc1 = nn.Linear(10* 3, 2)
def forward(self, x):
print(x.shape)
x = x.view(1, 50,-1)
print(x.shape)
x = self.layer1(x)
print(x.shape)
x = self.act1(x)
print(x.shape)
x = self.layer2(x)
print(x.shape)
x = x.view(1,-1)
print(x.shape)
x = self.fc1(x)
print(x.shape)
print(x)
return x
model_a = Simple1DCNN()
model_a = model_a.to(device)
criterion = nn.CrossEntropyLoss()
loss_fn = torch.nn.CrossEntropyLoss()
n_epochs_a = 50
learning_rate_a = 0.01
alpha_a = 1e-5
momentum_a = 0.9
optimizer = torch.optim.SGD(model_a.parameters(),
momentum = momentum_a,
nesterov = True,
weight_decay = alpha_a,
lr=learning_rate_a)
train_losses_a, val_losses_a, train_acc_a, val_acc_a = train(model_a,
train_dataloader,
val_dataloader,
optimizer,
n_epochs_a,
loss_fn
)
Error message:
cpu
size of Training dataset: 20
size of Training dataset: 10
0%| | 0/20 [00:00<?, ?it/s]
torch.Size([1, 500])
torch.Size([1, 50, 10])
torch.Size([1, 20, 3])
torch.Size([1, 20, 3])
torch.Size([1, 10, 3])
torch.Size([1, 30])
torch.Size([1, 2])
tensor([[ 0.5785, -1.0169]], grad_fn=<AddmmBackward>)
Traceback (most recent call last):
File "SO_question.py", line 219, in <module>
train_losses_a, val_losses_a, train_acc_a, val_acc_a = train(model_a,
File "SO_question.py", line 137, in train
train_loss, train_accuracy = train_epoch(model, train_dataloader, optimizer, loss_fn)
File "SO_question.py", line 93, in train_epoch
loss = loss_fn(output, labels)
File "/Users/mymac/Documents/programming/python/mainenv/lib/python3.8/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
result = self.forward(*input, **kwargs)
File "/Users/mymac/Documents/programming/python/mainenv/lib/python3.8/site-packages/torch/nn/modules/loss.py", line 961, in forward
return F.cross_entropy(input, target, weight=self.weight,
File "/Users/mymac/Documents/programming/python/mainenv/lib/python3.8/site-packages/torch/nn/functional.py", line 2468, in cross_entropy
return nll_loss(log_softmax(input, 1), target, weight, None, ignore_index, None, reduction)
File "/Users/mymac/Documents/programming/python/mainenv/lib/python3.8/site-packages/torch/nn/functional.py", line 2264, in nll_loss
ret = torch._C._nn.nll_loss(input, target, weight, _Reduction.get_enum(reduction), ignore_index)
RuntimeError: 1D target tensor expected, multi-target not supported
What am I doing wrong?

You are using nn.CrossEntropyLoss as the criterion for your training. You correctly passed the labels as indices of the ground truth class: 0s and 1s. However, as the error message suggests, it needs to be a 1D tensor!
Simply remove the reshape in ECGNet's __getitem__:
def __getitem__(self, idx):
ecgVec = self.ecg[idx]
labelID = self.target[idx]
return ecgVec,labelID
Edit
I want to increase the batch_size to 8. But now I get the error [...]
You are doing a lot of broadcasting (flattening) which surely will affect the batch size. As a general rule of thumb never fiddle with axis=0. For instance, if you have an input shape of (8, 500), straight off you have a problem when doing x.view(1, 50, -1). Since the resulting tensor will be (1, 50, 80) (the desired shape would have been (8, 50, 10)). Instead, you could broadcast with x.view(x.size(0), 50, -1).
Same with x.view(1, -1) later down forward. You are looking to flatten the tensor, but you should not flatten it along with the batches, they need to stay separated! It's safer to use torch.flatten, yet I prefer nn.Flatten which flattens from axis=1 to axis=-1 by default.
My personal advice is to start with a simple setup (without train loops etc...) to verify the architecture and intermediate output shapes. Then, add the necessary logic to handle the training.
class ECGNet(data.Dataset):
"""ImageNet Limited dataset."""
def __init__(self, ecgs, labls, transform=None):
self.ecg = ecgs
self.target = labls
self.transform = transform
def __getitem__(self, idx):
ecgVec = self.ecg[idx]
labelID = self.target[idx]
return ecgVec, labelID
def __len__(self):
return len(self.ecg)
class Simple1DCNN(nn.Module):
def __init__(self):
super(Simple1DCNN, self).__init__()
self.layer1 = nn.Conv1d(in_channels=50,
out_channels=20,
kernel_size=5,
stride=2)
self.act1 = nn.ReLU()
self.layer2 = nn.Conv1d(in_channels=20,
out_channels=10,
kernel_size=1)
self.fc1 = nn.Linear(10*3, 2)
self.flatten = nn.Flatten()
def forward(self, x):
x = x.view(x.size(0), 50, -1)
x = self.layer1(x)
x = self.act1(x)
x = self.layer2(x)
x = self.flatten(x)
x = self.fc1(x)
return x
batch_size = 8
train_data = ECGNet(ecg_train, labels_train)
train_dl = DataLoader(dataset=train_data,
batch_size=batch_size,
shuffle=True,
num_workers=0)
model = Simple1DCNN()
criterion = nn.CrossEntropyLoss()
Then
>>> x, y = next(iter(train_dl))
>>> y_hat = model(x)
>>> y_hat.shape
torch.Size([8, 2])
Also, make sure your loss works:
>>> criterion(y_hat, y)
tensor(..., grad_fn=<NllLossBackward>)

Implementing a custom dataset with PyTorch

I'm attempting to modify this feedforward network taken from https://github.com/yunjey/pytorch-tutorial/blob/master/tutorials/01-basics/feedforward_neural_network/main.py
to utilize my own dataset.
I define a custom dataset of two 1 dim arrays as input and two scalars the corresponding output :
x = torch.tensor([[5.5, 3,3,4] , [1 , 2,3,4], [9 , 2,3,4]])
print(x)
y = torch.tensor([1,2,3])
print(y)
import torch.utils.data as data_utils
my_train = data_utils.TensorDataset(x, y)
my_train_loader = data_utils.DataLoader(my_train, batch_size=50, shuffle=True)
I've updated the hyperparameters to match new input_size (2) & num_classes (3).
I've also changed images = images.reshape(-1, 28*28).to(device) to images = images.reshape(-1, 4).to(device)
As the training set is minimal I've changed the batch_size to 1.
Upon making these modifications I receive error when attempting to train :
RuntimeError Traceback (most recent call
last) in ()
51
52 # Forward pass
---> 53 outputs = model(images)
54 loss = criterion(outputs, labels)
55
/home/.local/lib/python3.6/site-packages/torch/nn/modules/module.py in call(self, *input, **kwargs)
489 result = self._slow_forward(*input, **kwargs)
490 else:
--> 491 result = self.forward(*input, **kwargs)
492 for hook in self._forward_hooks.values():
493 hook_result = hook(self, input, result)
in forward(self, x)
31
32 def forward(self, x):
---> 33 out = self.fc1(x)
34 out = self.relu(out)
35 out = self.fc2(out)
/home/.local/lib/python3.6/site-packages/torch/nn/modules/module.py in call(self, *input, **kwargs)
489 result = self._slow_forward(*input, **kwargs)
490 else:
--> 491 result = self.forward(*input, **kwargs)
492 for hook in self._forward_hooks.values():
493 hook_result = hook(self, input, result)
/home/.local/lib/python3.6/site-packages/torch/nn/modules/linear.py in forward(self, input)
53
54 def forward(self, input):
---> 55 return F.linear(input, self.weight, self.bias)
56
57 def extra_repr(self):
/home/.local/lib/python3.6/site-packages/torch/nn/functional.py
in linear(input, weight, bias)
990 if input.dim() == 2 and bias is not None:
991 # fused op is marginally faster
--> 992 return torch.addmm(bias, input, weight.t())
993
994 output = input.matmul(weight.t())
RuntimeError: size mismatch, m1: [3 x 4], m2: [2 x 3] at
/pytorch/aten/src/THC/generic/THCTensorMathBlas.cu:249
How to amend code to match expected dimensionality ? I'm unsure what code to change as I've changed all parameters that require updating ?
Source prior to changes :
import torch
import torch.nn as nn
import torchvision
import torchvision.transforms as transforms
# Device configuration
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# Hyper-parameters
input_size = 784
hidden_size = 500
num_classes = 10
num_epochs = 5
batch_size = 100
learning_rate = 0.001
# MNIST dataset
train_dataset = torchvision.datasets.MNIST(root='../../data',
train=True,
transform=transforms.ToTensor(),
download=True)
test_dataset = torchvision.datasets.MNIST(root='../../data',
train=False,
transform=transforms.ToTensor())
# Data loader
train_loader = torch.utils.data.DataLoader(dataset=train_dataset,
batch_size=batch_size,
shuffle=True)
test_loader = torch.utils.data.DataLoader(dataset=test_dataset,
batch_size=batch_size,
shuffle=False)
# Fully connected neural network with one hidden layer
class NeuralNet(nn.Module):
def __init__(self, input_size, hidden_size, num_classes):
super(NeuralNet, self).__init__()
self.fc1 = nn.Linear(input_size, hidden_size)
self.relu = nn.ReLU()
self.fc2 = nn.Linear(hidden_size, num_classes)
def forward(self, x):
out = self.fc1(x)
out = self.relu(out)
out = self.fc2(out)
return out
model = NeuralNet(input_size, hidden_size, num_classes).to(device)
# Loss and optimizer
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
# Train the model
total_step = len(train_loader)
for epoch in range(num_epochs):
for i, (images, labels) in enumerate(train_loader):
# Move tensors to the configured device
images = images.reshape(-1, 28*28).to(device)
labels = labels.to(device)
# Forward pass
outputs = model(images)
loss = criterion(outputs, labels)
# Backward and optimize
optimizer.zero_grad()
loss.backward()
optimizer.step()
if (i+1) % 100 == 0:
print ('Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}'
.format(epoch+1, num_epochs, i+1, total_step, loss.item()))
# Test the model
# In test phase, we don't need to compute gradients (for memory efficiency)
with torch.no_grad():
correct = 0
total = 0
for images, labels in test_loader:
images = images.reshape(-1, 28*28).to(device)
labels = labels.to(device)
outputs = model(images)
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum().item()
print('Accuracy of the network on the 10000 test images: {} %'.format(100 * correct / total))
# Save the model checkpoint
torch.save(model.state_dict(), 'model.ckpt')
Source post changes :
x = torch.tensor([[5.5, 3,3,4] , [1 , 2,3,4], [9 , 2,3,4]])
print(x)
y = torch.tensor([1,2,3])
print(y)
import torch.utils.data as data_utils
my_train = data_utils.TensorDataset(x, y)
my_train_loader = data_utils.DataLoader(my_train, batch_size=50, shuffle=True)
print(my_train)
print(my_train_loader)
import torch
import torch.nn as nn
import torchvision
import torchvision.transforms as transforms
# Device configuration
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# Hyper-parameters
input_size = 2
hidden_size = 3
num_classes = 3
num_epochs = 5
batch_size = 1
learning_rate = 0.001
# MNIST dataset
train_dataset = my_train
# Data loader
train_loader = my_train_loader
# Fully connected neural network with one hidden layer
class NeuralNet(nn.Module):
def __init__(self, input_size, hidden_size, num_classes):
super(NeuralNet, self).__init__()
self.fc1 = nn.Linear(input_size, hidden_size)
self.relu = nn.ReLU()
self.fc2 = nn.Linear(hidden_size, num_classes)
def forward(self, x):
out = self.fc1(x)
out = self.relu(out)
out = self.fc2(out)
return out
model = NeuralNet(input_size, hidden_size, num_classes).to(device)
# Loss and optimizer
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
# Train the model
total_step = len(train_loader)
for epoch in range(num_epochs):
for i, (images, labels) in enumerate(train_loader):
# Move tensors to the configured device
images = images.reshape(-1, 4).to(device)
labels = labels.to(device)
# Forward pass
outputs = model(images)
loss = criterion(outputs, labels)
# Backward and optimize
optimizer.zero_grad()
loss.backward()
optimizer.step()
if (i+1) % 100 == 0:
print ('Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}'
.format(epoch+1, num_epochs, i+1, total_step, loss.item()))
# Test the model
# In test phase, we don't need to compute gradients (for memory efficiency)
with torch.no_grad():
correct = 0
total = 0
for images, labels in test_loader:
images = images.reshape(-1, 4).to(device)
labels = labels.to(device)
outputs = model(images)
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum().item()
print('Accuracy of the network on the 10000 test images: {} %'.format(100 * correct / total))
# Save the model checkpoint
torch.save(model.state_dict(), 'model.ckpt')

You need to change input_size to 4 (2*2), and not 2 as your modified code currently shows.
If you compare it to the original MNIST example, you'll see that input_size is set to 784 (28*28) and not just to 28.