Chainer batchnormalization does not work well with my code although batch normalization of tensorflow works. I use the dataset mnist the code below shows.
Using chainer(version=6.1.0), without batchnormalization, the validation accuracy is within 0.97 and 0.98 after 100 epochs whereas with batchnormalization, it is less than 0.80 after 100 epochs.
When I use the same way with tensorflow(version=1.14.0), the validation accuracy is around 0.98 in a both way, with batchnormalization or without batchnormalization.
This is the part of my code.
The number of epoch is 100, and its batchsize is 1000. I use the Adam as the optimizer with learning_rate 0.01.
dataset, train data, validation data
(train_images, train_labels), (test_images, test_labels) = mnist.load_data()
train_images=train_images.reshape(60000, 28*28)
test_images = test_images.reshape(10000, 28*28)
x_train = train_images.astype('float32')/255
y_train = train_labels.astype('int32')
x_val = test_images.astype('float32')/255
y_val = test_labels.astype('int32')
model and condition (chainer)
# Define model
class MyModel(Chain):
def __init__(self,n_in=784,n_hidden=100,n_out=10):
initializer = chainer.initializers.HeNormal()
super().__init__()
with self.init_scope():
self.l1=L.Linear(n_in, n_hidden, initialW=initializer)
self.l2=L.Linear(n_hidden, n_hidden, initialW=initializer)
self.l3=L.Linear(n_hidden, n_out, initialW=initializer)
self.bn=L.BatchNormalization(n_hidden, decay=0.99, eps=0.001)
def forward(self,x):
h = F.relu(self.bn(self.l1(x)))
h = F.relu(self.bn(self.l2(h)))
return self.l3(h)
model = MyModel()
optimizer = optimizers.Adam()
optimizer.setup(model)
n_epoch = 100
n_batchsize = 1000
model and condition (tensorflow)
n_inputs = 28 * 28
n_hidden1 = 100
n_hidden2 = 100
n_outputs = 10
batch_norm_momentum = 0.9
X = tf.placeholder(tf.float32, shape=(None, n_inputs), name="X")
y = tf.placeholder(tf.int32, shape=(None), name="y")
training = tf.placeholder_with_default(False, shape=(), name='training')
with tf.name_scope("dnn"):
he_init = tf.variance_scaling_initializer()
my_batch_norm_layer = partial(tf.layers.batch_normalization,
training=training,
momentum=batch_norm_momentum)
my_dense_layer = partial(tf.layers.dense,
kernel_initializer=he_init)
hidden1 = my_dense_layer(X, n_hidden1, name="hidden1")
bn1 = tf.nn.relu(my_batch_norm_layer(hidden1))
hidden2 = my_dense_layer(bn1, n_hidden2, name="hidden2")
bn2 = tf.nn.relu(my_batch_norm_layer(hidden2))
logits_before_bn = my_dense_layer(bn2, n_outputs, name="outputs")
logits = my_batch_norm_layer(logits_before_bn)
with tf.name_scope("loss"):
xentropy = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=y, logits=logits)
loss = tf.reduce_mean(xentropy, name="loss")
with tf.name_scope("train"):
optimizer = tf.train.AdamOptimizer(learning_rate=0.001)
training_op = optimizer.minimize(loss)
with tf.name_scope("eval"):
correct = tf.nn.in_top_k(logits, y, 1)
accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
All code I use with chainer
import numpy as np
import chainer
from chainer import cuda, Function, gradient_check, report, training, utils, Variable
from chainer import datasets, iterators, optimizers, serializers
from chainer import Link, Chain, ChainList
import chainer.functions as F
import chainer.links as L
from chainer.training import extensions
from keras.datasets import mnist
import cupy as cp
(train_images, train_labels), (test_images, test_labels) = mnist.load_data()
train_images=train_images.reshape(60000, 28*28)
test_images = test_images.reshape(10000, 28*28)
x_train = train_images.astype('float32')/255
y_train = train_labels.astype('int32')
x_val = test_images.astype('float32')/255
y_val = test_labels.astype('int32')
# Define model
class MyModel(Chain):
def __init__(self,n_in=784,n_hidden=100,n_out=10):
initializer = chainer.initializers.HeNormal()
super().__init__()
with self.init_scope():
self.l1=L.Linear(n_in, n_hidden, initialW=initializer)
self.l2=L.Linear(n_hidden, n_hidden, initialW=initializer)
self.l3=L.Linear(n_hidden, n_out, initialW=initializer)
self.bn=L.BatchNormalization(n_hidden, decay=0.9, eps=0.001)
def forward(self,x):
h = F.relu(self.bn(self.l1(x)))
h = F.relu(self.bn(self.l2(h)))
return self.l3(h)
# define optimizer
model = MyModel()
optimizer = optimizers.Adam(alpha=0.01)
optimizer.setup(model)
## learn network
n_epoch = 100
n_batchsize = 1000
iteration = 0
gpu_id = 0
cuda.get_device(gpu_id).use()
# send the network to gpu memory
model.to_gpu(gpu_id)
print("epoch train/loss val/loss train/acc val/acc")
for epoch in range(n_epoch):
# order dataset randomly
order = np.random.permutation(range(len(x_train)))
loss_list = []
accuracy_list = []
for i in range(0, len(order), n_batchsize):
index = order[i:i+n_batchsize]
x_train_batch = x_train[index,:]
y_train_batch = y_train[index]
x_train_batch = cp.asarray(x_train_batch)
y_train_batch = cp.asarray(y_train_batch)
output_train = model(x_train_batch)
loss_train_batch = F.softmax_cross_entropy(output_train, y_train_batch)
accuracy_train_batch = F.accuracy(output_train, y_train_batch)
loss_list.append(cuda.to_cpu(loss_train_batch.array))
accuracy_list.append(cuda.to_cpu(accuracy_train_batch.array))
model.cleargrads()
loss_train_batch.backward()
optimizer.update()
iteration += 1
loss_train = np.mean(loss_list)
accuracy_train = np.mean(accuracy_list)
# after one epoch, evaluate with validation data
x_val = cp.asarray(x_val)
y_val = cp.asarray(y_val)
with chainer.using_config('train', False), chainer.using_config('enable_backprop', False):
output_val = model(x_val)
loss_val = F.softmax_cross_entropy(output_val, y_val)
loss_val = cuda.to_cpu(loss_val.array)
accuracy_val = F.accuracy(output_val, y_val)
accuracy_val = cuda.to_cpu(accuracy_val.array)
print('{0:>4d} {1:>10.4f} {2:>10.4f} {3:>10.4f} {4:>10.4f}'.format(epoch,loss_train,loss_val,accuracy_train,accuracy_val))
All code I use with tensorflow
python
import tensorflow as tf
from keras.datasets import mnist
from functools import partial
import numpy as np
def shuffle_batch(X, y, batch_size):
rnd_idx = np.random.permutation(len(X))
n_batches = len(X) // batch_size
for batch_idx in np.array_split(rnd_idx, n_batches):
X_batch, y_batch = X[batch_idx], y[batch_idx]
yield X_batch, y_batch
(train_images, train_labels), (test_images, test_labels) = mnist.load_data()
train_images=train_images.reshape(60000, 28*28)
test_images = test_images.reshape(10000, 28*28)
X_train = train_images.astype('float32')/255
y_train = train_labels.astype('int32')
X_valid = test_images.astype('float32')/255
y_valid = test_labels.astype('int32')
n_inputs = 28 * 28
n_hidden1 = 100
n_hidden2 = 100
n_outputs = 10
batch_norm_momentum = 0.9
X = tf.placeholder(tf.float32, shape=(None, n_inputs), name="X")
y = tf.placeholder(tf.int32, shape=(None), name="y")
training = tf.placeholder_with_default(False, shape=(), name='training')
with tf.name_scope("dnn"):
he_init = tf.variance_scaling_initializer()
my_batch_norm_layer = partial(tf.layers.batch_normalization,
training=training,
momentum=batch_norm_momentum)
my_dense_layer = partial(tf.layers.dense)
hidden1 = my_dense_layer(X, n_hidden1, name="hidden1")
bn1 = tf.nn.relu(my_batch_norm_layer(hidden1))
hidden2 = my_dense_layer(bn1, n_hidden2, name="hidden2")
bn2 = tf.nn.relu(my_batch_norm_layer(hidden2))
logits_before_bn = my_dense_layer(bn2, n_outputs, name="outputs")
logits = my_batch_norm_layer(logits_before_bn)
with tf.name_scope("loss"):
xentropy = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=y, logits=logits)
loss = tf.reduce_mean(xentropy, name="loss")
learning_rate = 0.01
with tf.name_scope("train"):
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
training_op = optimizer.minimize(loss)
with tf.name_scope("eval"):
correct = tf.nn.in_top_k(logits, y, 1)
accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
init = tf.global_variables_initializer()
saver = tf.train.Saver()
n_epochs = 100
batch_size = 1000
extra_update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
print("epoch train/loss val/loss train/acc val/acc")
with tf.Session() as sess:
init.run()
for epoch in range(n_epochs):
loss_list = []
accuracy_list = []
for X_batch, y_batch in shuffle_batch(X_train, y_train, batch_size):
sess.run([training_op, extra_update_ops],
feed_dict={training: True, X: X_batch, y: y_batch})
loss_batch = loss.eval(feed_dict={X: X_batch, y: y_batch})
accuracy_batch = accuracy.eval(feed_dict={X: X_batch, y: y_batch})
loss_list.append(loss_batch)
accuracy_list.append(accuracy_batch)
loss_val = loss.eval(feed_dict={X: X_valid, y: y_valid})
accuracy_val = accuracy.eval(feed_dict={X: X_valid, y: y_valid})
print('{0:>4d} {1:>10.4f} {2:>10.4f} {3:>10.4f} {4:>10.4f}'
.format(epoch,np.mean(loss_list),loss_val,np.mean(accuracy_list),accuracy_val))
I expect batch normalization with chainer would be around 98% but it got less than 80%.
Am I using the batchnormalization with chainer in a wrong way, or the structure of batchnormalization differs a lot between chainer and tensorflow??
In order to use different batch statistics between the layers, the model definition has to be like the following code, which achieves 98% validation accuracy after 100 epochs in my environment.
class MyModel(Chain):
def __init__(self,n_in=784,n_hidden=100,n_out=10):
initializer = chainer.initializers.HeNormal()
super().__init__()
with self.init_scope():
self.l1=L.Linear(n_in, n_hidden, initialW=initializer)
self.l2=L.Linear(n_hidden, n_hidden, initialW=initializer)
self.l3=L.Linear(n_hidden, n_out, initialW=initializer)
self.bn1=L.BatchNormalization(n_hidden, decay=0.9, eps=0.001)
self.bn2=L.BatchNormalization(n_hidden, decay=0.9, eps=0.001)
def forward(self,x):
h = F.relu(self.bn1(self.l1(x)))
h = F.relu(self.bn2(self.l2(h)))
return self.l3(h)
Related
This is my code, and image of training accuracy
# !curl -L -s -o '/content/ESC-50-master-16k.tar' 'https://drive.google.com/uc?id=1hFt-qarD_Ihjb3jW7pd3tje2pcY5tfuJ&confirm=t'
# !tar -xvf "ESC-50-master-16k.tar"
from torch.utils.data import Dataset, TensorDataset, DataLoader
import torch
import torch.nn as nn
import torch.nn.functional as F
import librosa
import numpy as np
import pandas as pd
from glob import glob
Device Configuration
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
Load ESC-50 Data & Preprocess
metadata = pd.read_csv("./ESC-50-master-16k/meta/esc50.csv")
wav_list = sorted(glob("./ESC-50-master-16k/resample/\*.wav"))
def spec_to_image(spec, eps=1e-6):
mean = spec.mean()
std = spec.std()
spec_norm = (spec - mean) / (std + eps)
spec_min, spec_max = spec_norm.min(), spec_norm.max()
spec_scaled = 255 \* (spec_norm - spec_min) / (spec_max - spec_min)
spec_scaled = spec_scaled.astype(np.uint8)
return spec_scaled
class esc50dataset(Dataset):
def __init__(self, wavlist, metadata):
self.labels = np.array(metadata.target).astype(int)
self.audio = \[\]
for f in wavlist:
wav, sr = librosa.load(f, sr=None)
spec=librosa.feature.melspectrogram(wav, sr=sr, n_fft=1024, hop_length=640, n_mels=126) # \[126,126\]
spec_db=librosa.power_to_db(spec,top_db=80)
spec_image = np.expand_dims(spec_to_image(spec_db), axis=0)
self.audio.append(spec_image.tolist())
self.audio = np.array(self.audio)
def __len__(self):
return len(self.audio)
def __getitem__(self, idx):
return self.audio\[idx\], self.labels\[idx\]
dataset = esc50dataset(wav_list, metadata)
features = dataset\[:\]\[0\]
labels = dataset\[:\]\[1\]
train test split
train_size = 0.7
val_size = 0.5
`make train set`
split_id = int(len(features) \* train_size) # 2000 \* 0.7 = 1400
train_features, remain_features = features\[:split_id\], features\[split_id:\] # train_features = 1400, remain_features = 600
train_labels, remain_labels = labels\[:split_id\], labels\[split_id:\] # train_labels = 1400, remain_labels = 600
`make val and test set`
split_val_id = int(len(remain_features) \* val_size) # 600 \* 0.5 = 300
val_features, test_features = remain_features\[:split_val_id\], remain_features\[split_val_id:\] # val_features = 300, test_features = 300
val_labels, test_labels = remain_labels\[:split_val_id\], remain_labels\[split_val_id:\] # val_labels = 300, test_labels = 300
`define batch size`
batch_size = 32
`create tensor datasets`
train_set = TensorDataset(torch.from_numpy(train_features), torch.from_numpy(train_labels))
valid_set = TensorDataset(torch.from_numpy(val_features), torch.from_numpy(val_labels))
test_set = TensorDataset(torch.from_numpy(test_features), torch.from_numpy(test_labels))
`create dataloaders`
train_loader = DataLoader(train_set, shuffle=True, batch_size=batch_size) # 1400 / 32
val_loader = DataLoader(valid_set, shuffle=True, batch_size=batch_size) # 300 / 32
test_loader = DataLoader(test_set, shuffle=True, batch_size=batch_size) # 300 / 32
`Model Hyperparameter`
in_channel = features\[0\].shape\[0\] # in_channel = 1
max_pool_kernel = 3
hidden_size = 256
hidden_layers = 10
output_classes = labels.max() + 1 # output_classes = 50
dropout = 0.25
learning_rate = 0.001
num_epochs = 10
Model
class CRNN(nn.Module):
def __init__(self, in_channel, hidden_size, hidden_layers, output_classes, dropout):
super(CRNN, self).__init__()
self.hidden_layers = hidden_layers
self.hidden_size = hidden_size
self.CNN_layer1 = nn.Sequential(
nn.Conv2d(in_channels=in_channel, out_channels=32, kernel_size=5, stride=1, padding=2),
nn.BatchNorm2d(num_features=32),
nn.ReLU(),
nn.MaxPool2d(kernel_size=max_pool_kernel), # 32, 42, 42
)
self.CNN_layer2 = nn.Sequential(
nn.Conv2d(in_channels=32, out_channels=64, kernel_size=5, stride=1, padding=2),
nn.BatchNorm2d(num_features=64),
nn.ReLU(),
nn.MaxPool2d(kernel_size=max_pool_kernel), # 64, 14, 14
)
self.lstm = nn.LSTM(input_size=64, hidden_size=hidden_size, num_layers=hidden_layers, batch_first=True)
self.dropout = nn.Dropout(dropout)
self.fc = nn.Linear(hidden_size, output_classes)
self.sigmoid = nn.Sigmoid()
def forward(self, x):
x = x.float()
# print("start : {}".format(x))
x = self.CNN_layer1(x) # 1, 126, 126 -> 32, 42, 42
# print("after CNN_1 : {}".format(x))
x = self.CNN_layer2(x) # 32, 42, 42 -> 64, 14, 14
# print("after CNN_2 : {}".format(x))
x = x.reshape(x.size(0), -1, 64) # 64, 14, 14 -> 12544 , x.size(0) = Batchsize = 32
# print("after reshape : {}".format(x))
h0 = torch.zeros(self.hidden_layers, x.size(0), self.hidden_size).to(device) # torch.size([10, 32(batch_size), 512])
c0 = torch.zeros(self.hidden_layers, x.size(0), self.hidden_size).to(device) # torch.size([10, 32(batch_size), 512])
x, _ = self.lstm(x, (h0, c0))
# print("after lstm : {}".format(x))
x = self.dropout(x)
# print("after dropout : {}".format(x))
x = self.fc(x[:,-1,:])
# print("after fc : {}".format(x))
x = self.sigmoid(x)
# print("after sigmoid : {}".format(x))
return x
model initialization
model = CRNN(in_channel, hidden_size, hidden_layers, output_classes, dropout)
Criterion & Optimizer
criterion = nn.CrossEntropyLoss().to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
Train
from tqdm import tqdm
tqdm.pandas()
tr_loss = \[\]
tr_acc = \[\]
v_loss = \[\]
v_acc = \[\]
best_valid_loss = torch.inf
best_epoch = 0
model = model.to(device)
epochloop = tqdm(range(num_epochs), position=0, desc='Training', leave=True)
for epoch in epochloop:
model.train()
train_loss = 0
train_acc = 0
epochloop.set_postfix_str(f'Training epoch {epoch}/{len(epochloop)}') # visualize
for idx, (audio, label) in enumerate(train_loader):
audio, label = audio.to(device), label.to(device)
out = model(audio)
\_, pred = torch.max(out.data, 1)
# acc
train_acc += (pred==label).sum()
# loss
optimizer.zero_grad()
loss = criterion(out, label)
train_loss += loss.item()
loss.backward()
optimizer.step()
Validation
model.eval()
val_loss = 0
val_acc = 0
with torch.no_grad():
for idx, (audio, label) in enumerate(val_loader):
epochloop.set_postfix_str(f'Validation batch {idx}/{len(val_loader)}')
audio, label = audio.to(device), label.to(device)
# forward pass
out = model(audio)
_, pred = torch.max(out.data, 1)
# acc
val_acc += (pred==label).sum()
# loss
loss = criterion(out, label)
val_loss += loss.item()
model.train()
save model if validation loss decrease
if val_loss / len(valid_set) \<= best_valid_loss:
best_valid_loss = val_loss / len(val_loader)
best_epoch = epoch
torch.save(model.state_dict(), "LSTM_epoch\_{}.pth".format(epoch))
print epoch loss & accuracy
print(f'Epoch {epoch+1}/{num_epochs} | Train Loss: {train_loss / len(train_loader):.3f} Train Acc: {train_acc / len(train_set) \* 100}% | Val Loss: {val_loss / len(val_loader):.3f} Val Acc: {val_acc / len(valid_set) \* 100}%')
tr_loss.append(train_loss / len(train_loader))
tr_acc.append(train_acc.cpu().numpy() / len(train_set) \* 100)
v_loss.append(val_loss / len(val_loader))
v_acc.append(val_acc.cpu().numpy() / len(valid_set) \* 100)
my training accuracy
Please tell me why my training accuracy is too low..
By your dataset, I'm inferring you are trying to identify sound data, and you are applying a CRNN. If you go to the repo Have you seen the repo for this dataset you will find a table with the accuracy for each method, the CRNN has about 60% accuracy, this is not so great since other methods achieve more than 95%.
You might need to change your model to get better results. Check the table you will find also some models with code
A bit difficult to give you an answer without reviewing your whole code, and re-running the model. Given this situation, I would suggest you create a Github repo, with appropriate the requirements.txt file for your code.
Even more helpful, and more likely to be answered would be if your code is in a jupyter notebook, and you can show the results of your loss curve and epochs. Your results could be more understandable shown like the output of tensorboard.
Hope this helps you to get some answers
I am trying to implement the code from a Pytorch beginner's tutorial. But I have written the code for loading the saved model in another Python file.
The FashionClassify file contains the code exactly as its in the tutorial.
Below is the code:
from FashionClassify import NeuralNetwork
from FashionClassify import test_data
import torch
model = NeuralNetwork()
model.load_state_dict(torch.load("model.pth"))
classes = [
"T-shirt/top", "Trouser","Pullover","Dress","Coat","Sandal","Shirt","Sneaker","Bag","Ankle boot",
]
model.eval()
x, y = test_data[0][0], test_data[0][1]
with torch.no_grad():
pred = model(x)
predicted, actual = classes[pred[0].argmax(0)],classes[y]
print(f'Predicted: "{predicted}", Actual: "{actual}"')
However, when I run this, the entire training process starts again. Why is that so ?
OR
Is it an expected behavior ?
(I have gone through a couple of webpages and StackOverflow answers but couldn't find my problem)
FashionClassify file code:
import torch
from torch import nn
from torch.utils.data import DataLoader # wraps an iterable around dataset
from torchvision import datasets # stores samples and their label
from torchvision.transforms import ToTensor, Lambda, Compose
import matplotlib as plt
training_data = datasets.FashionMNIST(root='data', train=True, download=True, transform=ToTensor(), )
test_data = datasets.FashionMNIST(root='data', train=False, download=True, transform=ToTensor(), )
batch_size = 64
train_dataLoader = DataLoader(training_data, batch_size=batch_size)
test_dataLoader = DataLoader(test_data, batch_size=batch_size)
for X, y in test_dataLoader:
print('Shape of X [N,C,H,W]:', X.size())
print('Shape of y:', y.shape, y.dtype)
break
device = 'cuda' if torch.cuda.is_available() else 'cpu'
print('Using {} device'.format(device))
# to define a NN, we inherit a class from nn.Module
class NeuralNetwork(nn.Module):
def __init__(self):
# will specify how data will proceed in the forward pass
super(NeuralNetwork, self).__init__()
self.flatten = nn.Flatten()
self.linear_relu_stack = nn.Sequential(
nn.Linear(28 * 28, 512),
nn.ReLU(),
nn.Linear(512, 512),
nn.ReLU(),
nn.Linear(512, 10),
nn.ReLU()
)
def forward(self, x):
x = self.flatten(x)
logits = self.linear_relu_stack(x)
return logits
model = NeuralNetwork().to(device)
print(model)
loss_fn = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr=1e-3)
def train(dataloader, model, loss_fn, optimizer):
size = len(dataloader.dataset)
for batch, (X,y) in enumerate(dataloader):
X,y = X.to(device), y.to(device)
#compute prediction error
pred = model(X)
loss = loss_fn(pred, y)
#backprop
optimizer.zero_grad()
loss.backward()
optimizer.step()
if batch%100 ==0:
loss,current = loss.item(), batch * len(X)
print(f"loss: {loss:>7f} [{current:>5d}/{size:>5d}]")
def test(dataloader, model):
size = len(dataloader.dataset)
model.eval()
test_loss, correct = 0,0
with torch.no_grad():
for X, y in dataloader:
X,y = X.to(device), y.to(device)
pred = model(X)
test_loss += loss_fn(pred, y).item()
correct += (pred.argmax(1) == y).type(torch.float).sum().item()
test_loss /= size
correct /= size
print(f"Test Error: \n Accuracy: {(100*correct):>0.1f}%, Avg loss: {test_loss:>8f} \n")
epochs = 5
for t in range(epochs):
print(f"Epoch {t+1}\n-------------------------------")
train(train_dataLoader, model, loss_fn, optimizer)
test(test_dataLoader, model)
print("Done!")
torch.save(model.state_dict(), "model.pth")
print("Saved PyTorch Model State to model.pth")
That's what happens when you import another file. All the code gets rerun.
Instead, in your training file:
class FancyNetwork(nn.Module):
[...]
def train():
[train code]
if __name__ == "__main__":
train()
Now when you run this file train() will get called, but when you import this file in another one, train won't get called automatically.
I built this acoustic model with features dim = [1124823,13] and labels dim = [1124823,1] and I split both to train, test, and dev. The problem that when I try to run my model I get this error
RuntimeError: expected scalar type Long but found Int in
loss = criterion(outputs, y_train)
import torch
import torch.nn as nn
from fela import feat, labels
from Dataloader import train_loader, test_loader, X_train, X_test, X_val, y_train, y_test, y_val
################################################################################################
input_size = 13
hidden1_size = 13
hidden2_size = 128
hidden3_size = 64
output_size = 50
################################################################################################
class DNN(nn.Module):
def __init__(self, input_size, hidden2_size, hidden3_size, output_size):
super(DNN, self).__init__()
self.fc1 = nn.Linear(input_size, hidden1_size)
self.relu1 = nn.ReLU()
self.fc2 = nn.Linear(hidden1_size, hidden2_size)
self.relu2 = nn.ReLU()
self.fc3 = nn.Linear(hidden2_size, hidden3_size)
self.relu3 = nn.ReLU()
self.fc4 = nn.Linear(hidden3_size, output_size)
self.relu4 = nn.ReLU()
def forward(self, x):
out = self.fc1(x)
out = self.relu1(out)
out = self.fc2(out)
out = self.relu2(out)
out = self.fc3(out)
out = self.relu3(out)
out = self.fc4(out)
out = self.relu4(out)
return out
################################################################################################
# Instantiate the model
batch_size = 50
n_iterations = 50
no_epochs = 80
model = DNN(input_size, hidden2_size, hidden3_size, output_size)
################################################################################################
# Define the loss criterion and optimizer
criterion = nn.CrossEntropyLoss()
learning_rate = 0.01
optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
print(model)
########################################################################################################################
# train the network
iter = 0
for epoch in range(no_epochs):
for i, (X_train, y_train) in enumerate(train_loader):
optimizer.zero_grad()
outputs = model(X_train)
loss = criterion(outputs, torch.max(labels, 1)[1])
loss.backward()
optimizer.step()
iter += 1
if iter % 500 == 0:
correct = 0
total = 0
for X_test, y_test in test_loader:
outputs = model(X_test)
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum()
accuracy = 100 * correct / total
print(iter, loss.data[0], accuracy)
I think no_epochs=0 with this initialization. Possibly (len(train_loader) / batch_size) > n_iterations. Then int(no_eps) = 0. Try to change no_epochs to 100 manually, for example.
no_eps = n_iterations / (len(train_loader) / batch_size)
no_epochs = int(no_eps)
for epoch in range(no_epochs):
I am training my own model using Tensorflow. However, I got some trouble when I change my activation function from Relu to Selu.
This is what happened. Learning curve drops accidentally and I have no idea about what's going on.
my learning curve
like this.
For what I have known, Selu can prevent overfitting, so I try to implement it in my model. Is there any tips, or any condition when I want to use Selu?
This is my code:
this is the place where I change my activation function
-----
def conv2d_maxpool(x_tensor, conv_num_outputs, conv_ksize, conv_strides, pool_ksize, pool_strides, layer_name):
conv_layer = tf.layers.conv2d(x_tensor, conv_num_outputs, kernel_size=conv_ksize, strides=conv_strides, activation=tf.nn.selu, name = layer_name)
conv_layer = tf.layers.max_pooling2d(conv_layer, pool_size=pool_ksize, strides=pool_strides)
return conv_layer
-----
graph
tf.reset_default_graph()
#### placeholder ####
input_img = tf.placeholder(dtype=tf.float32, shape=(None, img_size, img_size, 3))
y_true = tf.placeholder(dtype=tf.float32, shape=(None, num_class))
keep_prob = tf.placeholder(tf.float32, name="keep_prob")
lr_in = tf.placeholder(dtype = tf.float32, name = 'learning_rate')
conv_ksize = (3,3)
conv_strides = (1,1)
pool_ksize = (2,2)
pool_strides = (2,2)
n_filters_1 = 32
n_filters_2 = 64
n_filters_3 = 128
n_filters_4 = 256
onebyone_ksize = (1,1)
#CNN
conv_1 = conv2d_maxpool(input_img, n_filters_1, conv_ksize, conv_strides, pool_ksize, pool_strides, layer_name = "conv1")
# conv_1 = tf.layers.conv2d(conv_1, conv_num_outputs, kernel_size=conv_ksize, strides=conv_strides, activation=tf.nn.relu)
# conv_1_norm = tf.layers.batch_normalization(conv_1, name = "batch_norm1")
# conv_1_dropout = tf.layers.dropout(conv_1_norm, rate = keep_prob, training = True, name = "dropout1")
conv_2 = conv2d_maxpool(conv_1, n_filters_2, conv_ksize, conv_strides, pool_ksize, pool_strides, layer_name = "conv2")
# conv_2_norm = tf.layers.batch_normalization(conv_2)
conv_3 = conv2d_maxpool(conv_2, n_filters_3, conv_ksize, conv_strides, pool_ksize, pool_strides, layer_name = "conv3")
# conv_3_norm = tf.layers.batch_normalization(conv_3, name = "batch_norm3")
# conv_3_dropout = tf.layers.dropout(conv_3_norm, rate = keep_prob, training = True, name = "dropout3")
conv_4 = conv2d_maxpool(conv_3, n_filters_4, conv_ksize, conv_strides, pool_ksize, pool_strides, layer_name = "conv4")
flatten = tf.layers.flatten(conv_4)
fc1 = tf.layers.dense(flatten, 256, activation = tf.nn.relu)
out = tf.layers.dense(fc1, 6, activation=None, name= "logits") #logit
predict = tf.nn.softmax(out)
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits = out, labels = y_true))
optimizer = tf.train.AdamOptimizer(lr).minimize(cost)
##accuracy
correct_pred = tf.equal(tf.argmax(out, 1), tf.argmax(y_true, 1))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32), name='accuracy')
Training
#history/record
train_loss, train_acc = [], []
valid_loss, valid_acc = [], []
update_per_epoch = int(np.floor(X_train.shape[0] / batch_size))
## early stopping and learning rate congig
es_patience = 10
es_n = 0
lr_patience = 3
lr_n = 0
save_model_path = './save'
saver = tf.train.Saver()
sess = tf.Session()
init = tf.global_variables_initializer()
sess.run(init)
# Initializing the variables
batch_gen = img_gen.flow(generator_input(X_train), y_train, batch_size = 32)
val_batch_gen = img_gen.flow(generator_input(X_valid), y_valid, batch_size = len(X_valid))
for i in range(epoch):
epoch_loss = 0
epoch_acc = 0
for j in range(update_per_epoch):
image, label = next(batch_gen)
_, this_loss, this_acc = sess.run([optimizer, cost, accuracy], feed_dict={
input_img : image,
y_true : label,
lr_in: lr,
keep_prob : keep_probability
})
epoch_loss += this_loss
epoch_acc += this_acc
## end of epoch
epoch_loss /= update_per_epoch
epoch_acc /= update_per_epoch
train_loss.append(epoch_loss)
train_acc.append(epoch_acc)
print('Epoch {:>2} Loss: {:>4.4f} Training Accuracy: {:.6f}'.format(i + 1, epoch_loss, epoch_acc))
valid_image, valid_label = next(val_batch_gen)
valid_this_loss, valid_this_acc = sess.run([cost, accuracy], feed_dict = {
input_img: valid_image,
y_true: valid_label,
lr_in: lr,
keep_prob: 1.
})
valid_loss.append(valid_this_loss)
valid_acc.append(valid_this_acc)
print('Epoch {:>2} Loss: {:>4.4f} Validation Accuracy: {:.6f}'.format(i + 1,valid_this_loss, valid_this_acc))
# early stop
if valid_this_loss > np.min(valid_loss):
es_n += 1
lr_n += 1
else:
es_n = 0
lr_n = 0
saver.save(sess, os.path.join(os.getcwd(), 'bestsession.ckpt'))
# early stop
if es_n >= es_patience:
print("-----------early stopping-------------")
break
# adaptive learning rate
if lr_n >= lr_patience:
lr *= lr_decay_rate
lr_n = 0
print("-----------adjust learning rate------------")
# Save Model
save_path = saver.save(sess, save_model_path)
print('-----model save ------')
----------- 18/09/07------------
I can always reproduce the same result.
And this is my code, I wrote it in Jupyter. But sorry I can't upload the training data:
https://drive.google.com/open?id=1uUE32KrNmWnhLbV8z-fyHSMu6zGCCG_e
I started learning tensorflow one week ago and i am struggling a bit with the possibility of restoring a model.
I am constructing a class for a CNN (from stanford course code) where i want to have a run and a test function (first one to train the model, second one to make a prediction on another set of datas).
The function run is working correctly, mais i have a hard time restoring the model to make the prediction in function test (here i have a problem shape and i guess it comes from the way I get the operation).
import os
import tensorflow as tf
from time import time
class Cnn:
def __init__(self, batch_size=128, skip_step=10, epochs=1, dropout_ratio=0.75):
self.batch_size = batch_size
self.skip_step = skip_step
self.epochs = epochs
self.dropout_ratio = dropout_ratio
self.x = tf.placeholder(tf.float32, [None, 784], name="X_placeholder")
self.images = tf.reshape(self.x, shape=[-1, 28, 28, 1])
self.y = tf.placeholder(tf.float32, [None, 10], name="Y_placeholder")
self.dropout = tf.placeholder(tf.float32, name="dropout")
self.global_step = tf.Variable(0, dtype=tf.int32, trainable=False, name='global_step')
self.layers = [self.images]
self.n_layers = 0
def convolution(self, kernel_size, n_output, strides, scope_name):
previous_layer = self.layers[self.n_layers]
with tf.variable_scope(scope_name) as scope:
kernel_shape = [kernel_size[0], kernel_size[1], previous_layer.get_shape()[3], n_output]
kernel = tf.get_variable("kernels", kernel_shape,
initializer=tf.truncated_normal_initializer())
biases = tf.get_variable("biases", [n_output],
initializer=tf.random_normal_initializer())
convolution_2d = tf.nn.conv2d(previous_layer, kernel, strides=strides,
padding="SAME")
convolution = tf.nn.relu(convolution_2d + biases, name=scope.name)
self.layers.append(convolution)
self.n_layers += 1
return convolution
def pool(self, size, strides, scope_name, padding="SAME"):
previous_layer = self.layers[self.n_layers]
with tf.variable_scope(scope_name):
pool = tf.nn.max_pool(previous_layer, ksize=size, strides=strides,
padding=padding)
self.layers.append(pool)
self.n_layers += 1
return pool
def fully_connected(self, n_input, n_output, scope_name, relu=False, dropout=True):
previous_layer = self.layers[self.n_layers]
with tf.variable_scope(scope_name):
weights = tf.get_variable("weights", [n_input, n_output],
initializer=tf.truncated_normal_initializer())
biases = tf.get_variable("biases", [n_output],
initializer=tf.random_normal_initializer())
pool_reshaped = tf.reshape(previous_layer, [-1, n_input])
logits = tf.add(tf.matmul(pool_reshaped, weights), biases, name="logits")
fc = logits
if relu:
fc = tf.nn.relu(fc, name="relu")
if dropout:
fc = tf.nn.dropout(fc, self.dropout, name="dropout")
self.layers.append(fc)
self.n_layers += 1
if relu is False and dropout is False:
self.logits = tf.add(tf.matmul(pool_reshaped, weights), biases, name="logits")
self.predictions = tf.nn.softmax(logits, name="predictions")
return fc
def set_loss(self, scope_name):
previous_layer = self.layers[self.n_layers]
with tf.name_scope(scope_name):
entropy = tf.nn.softmax_cross_entropy_with_logits(logits=previous_layer, labels=self.y)
loss = tf.reduce_mean(entropy, name='loss')
self.layers.append(loss)
self.n_layers += 1
self.loss = loss
def set_optimizer(self):
previous_layer = self.layers[self.n_layers]
optimizer = tf.train.AdamOptimizer(0.001).minimize(previous_layer, global_step=self.global_step)
self.layers.append(optimizer)
self.n_layers += 1
self.optimizer = optimizer
def run(self, train_x, train_y):
sess = tf.Session()
with sess.as_default():
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
writer = tf.summary.FileWriter('./my_graph/mnist', sess.graph)
ckpt = tf.train.get_checkpoint_state(os.path.dirname('results/checkpoint'))
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
initial_step = self.global_step.eval()
start_time = time()
n_batches = int(train_x.shape[0] / self.batch_size)
total_correct_preds = 0
total_loss = 0.0
current_position = 0
for index in range(initial_step, int(n_batches * self.epochs)): # train the model n_epochs times
x_batch = train_x[current_position: current_position + self.batch_size, :]
y_batch = train_y[current_position: current_position + self.batch_size, :]
feed_dict = {self.x: x_batch, self.y: y_batch, self.dropout: self.dropout_ratio}
_, loss_batch, logits_batch = sess.run([self.optimizer, self.loss, self.logits],
feed_dict=feed_dict)
print(logits_batch.shape)
total_loss += loss_batch
preds = tf.nn.softmax(logits_batch)
correct_preds = tf.equal(tf.argmax(preds, 1), tf.argmax(y_batch, 1))
accuracy = tf.reduce_sum(tf.cast(correct_preds, tf.float32))
total_correct_preds += sess.run(accuracy)
if (index + 1) % self.skip_step == 0:
print('Average loss at step {}: {:5.1f}'.format(index + 1, total_loss / self.skip_step))
total_loss = 0.0
saver.save(sess, 'results/mnist-convnet', index)
current_position += self.batch_size
print("Optimization Finished!") # should be around 0.35 after 25 epochs
print("Total time: {0} seconds".format(time() - start_time))
print("Accuracy {0}".format(total_correct_preds / train_x.shape[0]))
def test(self, val_x, val_y):
checkpoint_file = tf.train.latest_checkpoint("results/")
graph = tf.Graph()
with graph.as_default():
sess = tf.Session()
with sess.as_default():
saver = tf.train.import_meta_graph("{}.meta".format(checkpoint_file))
saver.restore(sess, checkpoint_file)
self.x = graph.get_operation_by_name("X_placeholder").outputs[0]
self.y = graph.get_operation_by_name("Y_placeholder").outputs[0]
self.dropout = graph.get_operation_by_name("dropout").outputs[0]
self.global_step = graph.get_operation_by_name("global_step").outputs[0]
self.logits = graph.get_operation_by_name("logits").outputs[0]
current_position = 0
n_batches = int(val_x.shape[0] / self.batch_size)
total_correct_preds = 0
for i in range(n_batches):
x_batch = val_x[current_position: current_position + self.batch_size]
y_batch = val_y[current_position: current_position + self.batch_size]
feed_dict = {self.x: x_batch, self.y: y_batch, self.dropout: self.dropout_ratio}
logits_batch = sess.run([self.logits], feed_dict=feed_dict)
preds = tf.nn.softmax(logits_batch)
# correct_preds = tf.equal(tf.argmax(preds, 1), tf.argmax(y_batch, 1))
"""
accuracy = tf.reduce_sum(tf.cast(correct_preds, tf.float32))
total_correct_preds += sess.run(accuracy)
current_position += self.batch_size
print("Accuracy {0}".format(total_correct_preds / val_x.shape[0]))
"""
and testing it like this (sorry if there is a lot of code) :
from cnn import Cnn
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
tf.set_random_seed(1)
N_CLASSES = 10
# load datas
mnist = input_data.read_data_sets("mnist", one_hot=True)
train_x, train_y = mnist.train.images, mnist.train.labels
val_x, val_y = mnist.test.images, mnist.test.labels
model = Cnn(batch_size=128, skip_step=10, epochs=0.25, dropout_ratio=0.75)
model.convolution(kernel_size=[5, 5], strides=[1, 1, 1, 1], n_output=32, scope_name="conv1")
model.pool(size=[1, 2, 2, 1], strides=[1, 2, 2, 1], scope_name="pool1")
model.convolution(kernel_size=[5, 5], strides=[1, 1, 1, 1], n_output=64, scope_name="conv2")
model.pool(size=[1, 2, 2, 1], strides=[1, 2, 2, 1], scope_name="pool2")
input_features = 7 * 7 * 64
model.fully_connected(n_input=input_features, n_output=1024, scope_name="fc", relu=True, dropout=True)
model.fully_connected(n_input=1024, n_output=N_CLASSES, scope_name="softmax_linear", dropout=False)
model.set_loss(scope_name="loss")
model.set_optimizer()
model.run(train_x=train_x, train_y=train_y)
model.test(val_x, val_y)
I would also appreciate any comments regarding the better way to create a class like this (in particular the handling of the placeholders and the sessions)
Thanks !
Nicolas