Why PyTorch optimizer might fail to update its parameters? - python

I am trying to do a simple loss-minimization for a specific variable coeff using PyTorch optimizers. This variable is supposed to be used as an interpolation coefficient for two vectors w_foo and w_bar to find a third vector, w_target.
w_target = `w_foo + coeff * (w_bar - w_foo)
With w_foo and w_bar set as constant, at each optimization step I calculate w_target for the given coeff. Loss is determined from w_target using a fairly complex process beyond the scope of this question.
# w_foo.shape = [1, 16, 512]
# w_bar.shape = [1, 16, 512]
# num_layers = 16
# num_steps = 10000
vgg_loss = VGGLoss()
coeff = torch.randn([num_layers, ])
optimizer = torch.optim.Adam([coeff], lr=initial_learning_rate)
for step in range(num_steps):
w_target = w_foo + torch.matmul(coeff, (w_bar - w_foo))
optimizer.zero_grad()
target_image = generator.synthesis(w_target)
processed_target_image = process(target_image)
loss = vgg_loss(processed_target_image, source_image)
loss.backward()
optimizer.step()
However, when running this optimizer, I am met with query_opt not changing from one step to another, optimizer being essentially useless. I would like to ask for some advice on what I am doing wrong here.
Edit:
As suggested, I will try to elaborate on the loss function. Essentially, w_target is used to generate an image, and VGGLoss uses VGG feature extractor to compare this synthetic image with a certain exemplar source image.
class VGGLoss(torch.nn.Module):
def __init__(self, device, vgg):
super().__init__()
for param in self.parameters():
param.requires_grad = True
self.vgg = vgg # VGG16 in eval mode
def forward(self, source, target):
loss = 0
source_features = self.vgg(source, resize_images=False, return_lpips=True)
target_features = self.vgg(target, resize_images=False, return_lpips=True)
loss += (source_features - target_features).square().sum()
return loss

Related

Pytorch backward does not compute the gradients for requested variables

I'm trying to train a resnet18 model on pytorch (+pytorch-lightning) with the use of Virtual Adversarial Training. During the computations required for this type of training I need to obtain the gradient of D (ie. the cross-entropy loss of the model) with regard to tensor r.
This should, in theory, happen in the following code snippet:
def generic_step(self, train_batch, batch_idx, step_type):
x, y = train_batch
unlabeled_idx = y is None
d = torch.rand(x.shape).to(x.device)
d = d/(torch.norm(d) + 1e-8)
pred_y = self.classifier(x)
y[unlabeled_idx] = pred_y[unlabeled_idx]
l = self.criterion(pred_y, y)
R_adv = torch.zeros_like(x)
for _ in range(self.ip):
r = self.xi * d
r.requires_grad = True
pred_hat = self.classifier(x + r)
# pred_hat = F.log_softmax(pred_hat, dim=1)
D = self.criterion(pred_hat, pred_y)
self.classifier.zero_grad()
D.requires_grad=True
D.backward()
R_adv += self.eps * r.grad / (torch.norm(r.grad) + 1e-8)
R_adv /= 32
loss = l + R_adv * self.a
loss.backward()
self.accuracy[step_type] = self.acc_metric(torch.argmax(pred_y, 1), y)
return loss
Here, to my understanding, r.grad should in theory be the gradient of D with respect to r. However, the code throws this at D.backward():
RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn
(full traceback excluded because this error is not helpful and technically "solved" as I know the cause for it, explained just below)
After some research and debugging it seems that in this situation D.backward() attempts to calculate dD/dD disregarding any previous mention of requires_grad=True. This is confirmed when I add D.requires_grad=True and I get D.grad=Tensor(1.,device='cuda:0') but r.grad=None.
Does anyone know why this may be happening?
In Lightning, .backward() and optimizer step are all handled under the hood. If you do it yourself like in the code above, it will mess with Lightning because it doesn't know you called backward yourself.
You can enable manual optimization in the LightningModule:
def __init__(self):
super().__init__()
# put this in your init
self.automatic_optimization = False
This tells Lightning that you are taking over calling backward and handling optimizer step + zero grad yourself. Don't forget to add that in your code above. You can access the optimizer and scheduler like so in your training step:
def training_step(self, batch, batch_idx):
optimizer = self.optimizers()
scheduler = self.lr_schedulers()
# do your training step
# don't forget to call:
# 1) backward 2) optimizer step 3) zero grad
Read more about manual optimization here.

LSTM always predicts the same class

I’m trying to solve an nlp classification problem with a LSTM. The code for the model is defined here:
class LSTM(nn.Module):
def __init__(self, hidden_size, embedding_size=66 ):
super().__init__()
self.lstm = nn.LSTM(embedding_size, hidden_size, batch_first = True, bidirectional = True)
self.fc = nn.Linear(2*hidden_size,2)
def forward(self, input_seq):
output, (hidden_state, cell_state) = self.lstm(input_seq)
hidden_state = torch.cat((hidden_state[-1,:], hidden_state[-2,:]), -1)
logits = self.fc(hidden_state)
return nn.LogSoftmax(dim=1)(logits)
And the function I’m using to train this model is here:
def train_loop(dataloader, model, loss_fn, optimizer):
loss_fn = loss_fn
size = len(dataloader.dataset)
model.train()
zeros = 0
for batch, (X, y) in enumerate(dataloader):
# Transform string into tensor
tensor = torch.zeros(1,len(X[0]),66)
for i in range(len(X[0])):
tensor[0][i][ctoi[X[0][i]]] = 1
pred = model(tensor)
target = torch.zeros(2, dtype=torch.long)
target[y] = 1
if batch % 100 == 0:
print(pred.squeeze(), target)
loss = loss_fn(pred.squeeze(), target)
# Backpropagation
optimizer.zero_grad()
loss.backward()
optimizer.step()
if pred.squeeze().argmax() == 0:
zeros += 1
if batch % 100 == 0:
loss, current = loss.item(), batch * len(X)
print(f"loss: {loss:>7f} [{current:>5d}/{size:>5d}]")
print(f'In trainning predicted {zeros} zeroes out of {size} samples')
The X’s are still strings, that’s why I need to convert them to tensors before running it through the model. The y’s are either a 0 or 1 (since its a binary classification problem), that I need to convert to a tensor of shape (2,) to run through the loss function.
For some reason I keep getting the same class predicted for every input. The classes are not even that unbalanced (~45% to 55%), and I’ve tried changing the weights of the classes in the loss function with no improvements, it either converges to predicting always a 0 or always a 1. Most of the time it it converges to predicting always a 0, which makes even less sense because what happens usually is that the class 0 has less samples than class 1.
Since you're training a binary classification model, your output dim should be 1 (corresponding to a single probability P(y|x)). This means that the y you're retrieving from your dataloader should be the y used in your loss function (assuming a cross-entropy loss). The predicted class is therefore y_hat = round(pred) (i.e., is the prediction >= 0.5).
As a point of clarity, it would be much easier to follow your logic if the one-hot encoding happened within your dataset (either in __getitem__ or __iter__). It's also worth noting that you don't use embeddings, so the code of your classifier is a bit misleading.

Cost remains same at 0.6932 in Siamese Network

I am trying to implement a Siamese Network, as in this paper
In this paper, they have used cross entropy for the Loss function
I am using STL-10 Dataset for training and instead of the 3 layer network used in the paper, I replaced it with VGG-13 CNN network, except the last logit layer.
Here is my loss function code
def loss(pred,true_pred):
cross_entropy_loss = tf.multiply(-1.0,tf.reduce_mean(tf.add(tf.multiply(true_pred,tf.log(pred)),tf.multiply((1-true_pred),tf.log(tf.subtract(1.0,pred))))))
total_loss = tf.add(tf.reduce_sum(tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)),cross_entropy_loss,name='total_loss')
return cross_entropy_loss,total_loss
with tf.device('/gpu:0'):
h1 = siamese(feed_image1)
h2 = siamese(feed_image2)
l1_dist = tf.abs(tf.subtract(h1,h2))
with tf.variable_scope('pred') as scope:
predictions = tf.contrib.layers.fully_connected(l1_dist,1,activation_fn = tf.sigmoid,weights_initializer = tf.contrib.layers.xavier_initializer(uniform=False),weights_regularizer = tf.contrib.layers.l2_regularizer(tf.constant(0.001, dtype=tf.float32)))
celoss,cost = loss(predictions,feed_labels)
with tf.variable_scope('adam_optimizer') as scope:
optimizer = tf.train.AdamOptimizer(learning_rate=0.001)
opt = optimizer.minimize(cost)
However, when I run the training, the cost remains almost constant at 0.6932
I have used Adam Optimizer here.
But previously I used Momentum Optimizer.
I have tried changing the learning rate but the cost still behaves the same.
And all the prediction values converge to 0.5 after a few iterations.
After taking the output for two batches of images (input1 and input2), I take their L1 distance and to that I have connected a fully connected layer with a single output and sigmoid activation function.
[h1 and h2 contains the output of the last fully connected layer(not the logit layer) of the VGG-13 network]
Since the output activation function is sigmoid, and since the prediction values are around 0.5, we can calculate and say that the sum of the weighted L1 distance of the output of the two networks is near to zero.
I can't understand where I am going wrong.
A little help will be very much appreciated.
I thought the nonconvergence may be caused by the gradient vanishing. You can trace the gradients using tf.contrib.layers.optimize_loss and the tensorboard. You can refer to this answer for more details.
Several optimizations(maybe):
1) don't write the cross entropy yourself.
You can employ the sigmoid cross entropy with logits API, since it ensures stability as documented:
max(x, 0) - x * z + log(1 + exp(-abs(x)))
2) do some weigh normalization may would hlep.
3) keep the regularization loss small.
You can read this answer for more information.
4) I don't see the necessity of tf.abs the L1 distance.
And here is the code I modified. Hope it helps.
mode = "training"
rl_rate = .1
with tf.device('/gpu:0'):
h1 = siamese(feed_image1)
h2 = siamese(feed_image2)
l1_dist = tf.subtract(h1, h2)
# is it necessary to use abs?
l1_dist_norm = tf.layers.batch_normalization(l1_dist, training=(mode=="training"))
with tf.variable_scope('logits') as scope:
w = tf.get_variable('fully_connected_weights', [tf.shape(l1_dist)[-1], 1],
weights_initializer = tf.contrib.layers.xavier_initializer(uniform=False), weights_regularizer = tf.contrib.layers.l2_regularizer(tf.constant(0.001, dtype=tf.float32))
)
logits = tf.tensordot(l1_dist_norm, w, axis=1)
xent_loss = tf.nn.sigmoid_cross_entropy_with_logits(logits=logits, labels=feed_labels)
total_loss = tf.add(tf.reduce_sum(rl_rate * tf.abs(tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES))), (1-rl_rate) * xent_loss, name='total_loss')
# or:
# weights = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
# l1_regularizer = tf.contrib.layers.l1_regularizer()
# regularization_loss = tf.contrib.layers.apply_regularization(l1_regularizer, weights)
# total_loss = xent_loss + regularization_loss
with tf.variable_scope('adam_optimizer') as scope:
optimizer = tf.train.AdamOptimizer(learning_rate=0.0005)
opt = tf.contrib.layers.optimize_loss(total_loss, global_step, learning_rate=learning_rate, optimizer="Adam", clip_gradients=max_grad_norm, summaries=["gradients"])

CTC loss goes down and stops

I’m trying to train a captcha recognition model. Model details are resnet pretrained CNN layers + Bidirectional LSTM + Fully Connected. It reached 90% sequence accuracy on captcha generated by python library captcha. The problem is that these generated captcha seems to have similary location of each character. When I randomly add spaces between characters, the model does not work any more. So I wonder is LSTM learning segmentation during learning? Then I try to use CTC loss. At first, loss goes down pretty quick. But it stays at about 16 without significant drop later. I tried different layers of LSTM, different number of units. 2 Layers of LSTM reach lower loss, but still not converging. 3 layers are just like 2 layers. The loss curve:
#encoding:utf8
import os
import sys
import torch
import warpctc_pytorch
import traceback
import torchvision
from torch import nn, autograd, FloatTensor, optim
from torch.nn import functional as F
from torch.utils.data import DataLoader
from torch.optim.lr_scheduler import MultiStepLR
from tensorboard import SummaryWriter
from pprint import pprint
from net.utils import decoder
from logging import getLogger, StreamHandler
logger = getLogger(__name__)
handler = StreamHandler(sys.stdout)
logger.addHandler(handler)
from dataset_util.utils import id_to_character
from dataset_util.transform import rescale, normalizer
from config.config import MAX_CAPTCHA_LENGTH, TENSORBOARD_LOG_PATH, MODEL_PATH
class CNN_RNN(nn.Module):
def __init__(self, lstm_bidirectional=True, use_ctc=True, *args, **kwargs):
super(CNN_RNN, self).__init__(*args, **kwargs)
model_conv = torchvision.models.resnet18(pretrained=True)
for param in model_conv.parameters():
param.requires_grad = False
modules = list(model_conv.children())[:-1] # delete the last fc layer.
for param in modules[8].parameters():
param.requires_grad = True
self.resnet = nn.Sequential(*modules) # CNN with fixed parameters from resnet as feature extractor
self.lstm_input_size = 512 * 2 * 2
self.lstm_hidden_state_size = 512
self.lstm_num_layers = 2
self.chracter_space_length = 64
self._lstm_bidirectional = lstm_bidirectional
self._use_ctc = use_ctc
if use_ctc:
self._max_captcha_length = int(MAX_CAPTCHA_LENGTH * 2)
else:
self._max_captcha_length = MAX_CAPTCHA_LENGTH
if lstm_bidirectional:
self.lstm_hidden_state_size = self.lstm_hidden_state_size * 2 # so that hidden size for one direction in bidirection lstm is the same as vanilla lstm
self.lstm = self.lstm = nn.LSTM(self.lstm_input_size, self.lstm_hidden_state_size // 2, dropout=0.5, bidirectional=True, num_layers=self.lstm_num_layers)
else:
self.lstm = nn.LSTM(self.lstm_input_size, self.lstm_hidden_state_size, dropout=0.5, bidirectional=False, num_layers=self.lstm_num_layers) # dropout doen't work for one layer lstm
self.ouput_to_tag = nn.Linear(self.lstm_hidden_state_size, self.chracter_space_length)
self.tensorboard_writer = SummaryWriter(TENSORBOARD_LOG_PATH)
# self.dropout_lstm = nn.Dropout()
def init_hidden_status(self, batch_size):
if self._lstm_bidirectional:
self.hidden = (autograd.Variable(torch.zeros((self.lstm_num_layers * 2, batch_size, self.lstm_hidden_state_size // 2))),
autograd.Variable(torch.zeros((self.lstm_num_layers * 2, batch_size, self.lstm_hidden_state_size // 2)))) # number of layers, batch size, hidden dimention
else:
self.hidden = (autograd.Variable(torch.zeros((self.lstm_num_layers, batch_size, self.lstm_hidden_state_size))),
autograd.Variable(torch.zeros((self.lstm_num_layers, batch_size, self.lstm_hidden_state_size)))) # number of layers, batch size, hidden dimention
def forward(self, image):
'''
:param image: # batch_size, CHANNEL, HEIGHT, WIDTH
:return:
'''
features = self.resnet(image) # [batch_size, 512, 2, 2]
batch_size = image.shape[0]
features = [features.view(batch_size, -1) for i in range(self._max_captcha_length)]
features = torch.stack(features)
self.init_hidden_status(batch_size)
output, hidden = self.lstm(features, self.hidden)
# output = self.dropout_lstm(output)
tag_space = self.ouput_to_tag(output.view(-1, output.size(2))) # [MAX_CAPTCHA_LENGTH * BATCH_SIZE, CHARACTER_SPACE_LENGTH]
tag_space = tag_space.view(self._max_captcha_length, batch_size, -1)
if not self._use_ctc:
tag_score = F.log_softmax(tag_space, dim=2) # [MAX_CAPTCHA_LENGTH, BATCH_SIZE, CHARACTER_SPACE_LENGTH]
else:
tag_score = tag_space
return tag_score
def train_net(self, data_loader, eval_data_loader=None, learning_rate=0.008, epoch_num=400):
try:
if self._use_ctc:
loss_function = warpctc_pytorch.warp_ctc.CTCLoss()
else:
loss_function = nn.NLLLoss()
# optimizer = optim.SGD(filter(lambda p: p.requires_grad, self.parameters()), momentum=0.9, lr=learning_rate)
# optimizer = MultiStepLR(optimizer, milestones=[10,15], gamma=0.5)
# optimizer = optim.Adadelta(filter(lambda p: p.requires_grad, self.parameters()))
optimizer = optim.Adam(filter(lambda p: p.requires_grad, self.parameters()))
self.tensorboard_writer.add_scalar("learning_rate", learning_rate)
tensorbard_global_step=0
if os.path.exists(os.path.join(TENSORBOARD_LOG_PATH, "resume_step")):
with open(os.path.join(TENSORBOARD_LOG_PATH, "resume_step"), "r") as file_handler:
tensorbard_global_step = int(file_handler.read()) + 1
for epoch_index, epoch in enumerate(range(epoch_num)):
for index, sample in enumerate(data_loader):
optimizer.zero_grad()
input_image = autograd.Variable(sample["image"]) # batch_size, 3, 255, 255
tag_score = self.forward(input_image)
if self._use_ctc:
tag_score, target, tag_score_sizes, target_sizes = self._loss_preprocess_ctc(tag_score, sample)
loss = loss_function(tag_score, target, tag_score_sizes, target_sizes)
loss = loss / tag_score.size(1)
else:
target = sample["padded_label_idx"]
tag_score, target = self._loss_preprocess(tag_score, target)
loss = loss_function(tag_score, target)
print("Training loss: {}".format(float(loss)))
self.tensorboard_writer.add_scalar("training_loss", float(loss), tensorbard_global_step)
loss.backward()
optimizer.step()
if index % 250 == 0:
print(u"Processing batch: {} of {}, epoch: {}".format(index, len(data_loader), epoch_index))
self.evaluate(eval_data_loader, loss_function, tensorbard_global_step)
tensorbard_global_step += 1
self.save_model(MODEL_PATH + "_epoch_{}".format(epoch_index))
except KeyboardInterrupt:
print("Exit for KeyboardInterrupt, save model")
self.save_model(MODEL_PATH)
with open(os.path.join(TENSORBOARD_LOG_PATH, "resume_step"), "w") as file_handler:
file_handler.write(str(tensorbard_global_step))
except Exception as excp:
logger.error(str(excp))
logger.error(traceback.format_exc())
def predict(self, image):
# TODO ctc version
'''
:param image: [batch_size, channel, height, width]
:return:
'''
tag_score = self.forward(image)
# TODO ctc
# if self._use_ctc:
# tag_score = F.softmax(tag_score, dim=-1)
# decoder.decode(tag_score)
confidence_log_probability, indexes = tag_score.max(2)
predicted_labels = []
for batch_index in range(indexes.size(1)):
label = ""
for character_index in range(self._max_captcha_length):
if int(indexes[character_index, batch_index]) != 1:
label += id_to_character[int(indexes[character_index, batch_index])]
predicted_labels.append(label)
return predicted_labels, tag_score
def predict_pil_image(self, pil_image):
try:
self.eval()
processed_image = normalizer(rescale({"image": pil_image}))["image"].view(1, 3, 255, 255)
result, tag_score = self.predict(processed_image)
self.train()
except Exception as excp:
logger.error(str(excp))
logger.error(traceback.format_exc())
return [""], None
return result, tag_score
def evaluate(self, eval_dataloader, loss_function, step=0):
total = 0
sequence_correct = 0
character_correct = 0
character_total = 0
loss_total = 0
batch_size = eval_data_loader.batch_size
true_predicted = {}
self.eval()
for sample in eval_dataloader:
total += batch_size
input_images = sample["image"]
predicted_labels, tag_score = self.predict(input_images)
for predicted, true_label in zip(predicted_labels, sample["label"]):
if predicted == true_label: # dataloader is making label a list, use batch_size=1
sequence_correct += 1
for index, true_character in enumerate(true_label):
character_total += 1
if index < len(predicted) and predicted[index] == true_character:
character_correct += 1
true_predicted[true_label] = predicted
if self._use_ctc:
tag_score, target, tag_score_sizes, target_sizes = self._loss_preprocess_ctc(tag_score, sample)
loss_total += float(loss_function(tag_score, target, tag_score_sizes, target_sizes) / batch_size)
else:
tag_score, target = self._loss_preprocess(tag_score, sample["padded_label_idx"])
loss_total += float(loss_function(tag_score, target)) # averaged over batch index
print("True captcha to predicted captcha: ")
pprint(true_predicted)
self.tensorboard_writer.add_text("eval_ture_to_predicted", str(true_predicted), global_step=step)
accuracy = float(sequence_correct) / total
avg_loss = float(loss_total) / (total / batch_size)
character_accuracy = float(character_correct) / character_total
self.tensorboard_writer.add_scalar("eval_sequence_accuracy", accuracy, global_step=step)
self.tensorboard_writer.add_scalar("eval_character_accuracy", character_accuracy, global_step=step)
self.tensorboard_writer.add_scalar("eval_loss", avg_loss, global_step=step)
self.zero_grad()
self.train()
def _loss_preprocess(self, tag_score, target):
'''
:param tag_score: value return by self.forward
:param target: sample["padded_label_idx"]
:return: (processed_tag_score, processed_target) ready for NLLoss function
'''
target = target.transpose(0, 1)
target = target.contiguous()
target = target.view(target.size(0) * target.size(1))
tag_score = tag_score.view(-1, self.chracter_space_length)
return tag_score, target
def _loss_preprocess_ctc(self, tag_score, sample):
target_2d = [
[int(ele) for ele in sample["padded_label_idx"][row, :] if int(ele) != 0 and int(ele) != 1]
for row in range(sample["padded_label_idx"].size(0))]
target = []
for ele in target_2d:
target.extend(ele)
target = autograd.Variable(torch.IntTensor(target))
# tag_score = F.softmax(F.sigmoid(tag_score), dim=-1)
tag_score_sizes = autograd.Variable(torch.IntTensor([self._max_captcha_length] * tag_score.size(1)))
target_sizes = autograd.Variable(sample["captcha_length"].int())
return tag_score, target, tag_score_sizes, target_sizes
# def visualize_graph(self, dataset):
# '''Since pytorch use dynamic graph, an input is required to visualize graph in tensorboard'''
# # warning: Do not run this, the graph is too large to visualize...
# sample = dataset[0]
# input_image = autograd.Variable(sample["image"].view(1, 3, 255, 255))
# tag_score = self.forward(input_image)
# self.tensorboard_writer.add_graph(self, tag_score)
def save_model(self, model_path):
self.tensorboard_writer.close()
self.tensorboard_writer = None # can't be pickled
torch.save(self, model_path)
self.tensorboard_writer = SummaryWriter(TENSORBOARD_LOG_PATH)
#classmethod
def load_model(cls, model_path=MODEL_PATH, *args, **kwargs):
net = cls(*args, **kwargs)
if os.path.exists(model_path):
model = torch.load(model_path)
if model:
model.tensorboard_writer = SummaryWriter(TENSORBOARD_LOG_PATH)
net = model
return net
def __del__(self):
if self.tensorboard_writer:
self.tensorboard_writer.close()
if __name__ == "__main__":
from dataset_util.dataset import dataset, eval_dataset
data_loader = DataLoader(dataset, batch_size=2, shuffle=True)
eval_data_loader = DataLoader(eval_dataset, batch_size=2, shuffle=True)
net = CNN_RNN.load_model()
net.train_net(data_loader, eval_data_loader=eval_data_loader)
# net.predict(dataset[0]["image"].view(1, 3, 255, 255))
# predict_pil_image test code
# from config.config import IMAGE_PATHS
# import glob
# from PIL import Image
#
# image_paths = glob.glob(os.path.join(IMAGE_PATHS.get("EVAL"), "*.png"))
# for image_path in image_paths:
# pil_image = Image.open(image_path)
# predicted, score = net.predict_pil_image(pil_image)
# print("True value: {}, predicted: {}".format(os.path.split(image_path)[1], predicted))
print("Done")
The above codes are main part. If you need other components that makes it running, leave a comment. Got stuck here for quite long. Any advice for training crnn + ctc is appreciated.
I've been training with ctc loss and encountered the same problem. I know this is a rather late answer but hopefully it'll help someone else who's researching on this. After trial and error and a lot of research there are a few things that's worth knowing when it comes to training with ctc (if your model is set up correctly):
The quickest way for the model to lower cost is to predict only blanks. This is noted in a few papers and blogs: see http://www.tbluche.com/ctc_and_blank.html
The model learns to predict only blanks first, then it starts picking up on the error signal in regards to the correct underlying labels. This is also explained in the above link. In practice, I noticed that my model starts to learn the real underlying labels/targets after a couple hundred epochs and the loss starts decreasing dramatically again. Similar to what is shown for the toy example here: https://thomasmesnard.github.io/files/CTC_Poster_Mesnard_Auvolat.pdf
These parameters have a great impact on whether your model converges or not - learning rate, batch size and epoch number.
You have a few questions, so I will try to answer them one by one.
First, why does adding spaces to the captcha break the model?
A neural network learns to deal with the data it is trained on. If you change the distribution of the data (by for example adding spaces between characters) there is no guarantee that the network will generalize. As you hint at in your question. It is possible that the captchas you train on always have the characters in the same positions, or at the same distance from one another, thus your model learns that and learns to exploit this by looking in those positions. If you want your network to generalize a specific scenario, you should explicitly train on that scenario. So in your case, you should add random spaces also during training.
Second, why does the loss not go below 16?
Clearly, from the fact that your training loss is also stalled at 16 (like your validation loss), the problem is that your model simply doesn't have the capacity to deal with the complexity of the problem. In other words, your model is underfitting. You had the correct reflex to try to increase the capacity of your network. You tried to increase the capacity of the LSTM and it didn't help. Thus, the next logical step is that the convolution part of your network is not powerful enough. So here are a few things that you might want to try, from most likely to succeed in my opinion to least likely:
Make convnet trainable: I notice that you are using a pretrained convnet and that you are not fine-tuning the weights of that convnet. That could be a problem. Whatever your convnet was trained on, it might not develop the required features to deal with captchas. You should try learning the weights of the convnet too, in order to develop useful features for captchas.
Use deeper convnet: This is the naive thing to do. Your convnet doesn't have good enough features, try a more powerful deeper one. (But you should definitely use this only after you've made the convnet trainable).
From my experience, training RNN model with CTC loss is not an EASY task. The model may not converge at all if the training is not carefully setup-ed. Here are my suggestions:
Check the CTC loss output along training. For a model would converge, the CTC loss at each batch fluctuates notably. If you observed that the CTC loss shrinks almost monotonically to a stable value, then the model is most likely stuck at a local minima
Use short samples to pretrain your model. Though we have advanced RNN strucures like LSTM and GRU, it's still hard to back-propagate the RNN for long steps.
Enlarge sample variety. You can even add artificial samples to help your model escape from local minima.
F.Y.I., we've just open-sourced a new deep learning framework Dandelion which has built-in CTC objective, and interface pretty much like pytorch. You can try your model with Dandelion and compare it with your current implementation.

TensorFlow dynamic RNN not training

Problem statement
I am trying to train a dynamic RNN in TensorFlow v1.0.1 on Linux RedHat 7.3 (problem also manifests on Windows 7), and no matter what I try, I get the exact same training and validation error at every epoch, i.e. my weights are not updating.
I appreciate any help you can offer.
Example
I tried to reduce this to a minimum example that shows my issue, but the minimum example is still pretty large. I based the network structure largely on this gist.
Network definition
import functools
import numpy as np
import tensorflow as tf
def lazy_property(function):
attribute = '_' + function.__name__
#property
#functools.wraps(function)
def wrapper(self):
if not hasattr(self, attribute):
setattr(self, attribute, function(self))
return getattr(self, attribute)
return wrapper
class MyNetwork:
"""
Class defining an RNN for labeling a time series.
"""
def __init__(self, data, target, num_hidden=64):
self.data = data
self.target = target
self._num_hidden = num_hidden
self._num_steps = int(self.target.get_shape()[1])
self._num_classes = int(self.target.get_shape()[2])
self._weight_and_bias() # create weight and bias tensors
self.prediction
self.error
self.optimize
#lazy_property
def prediction(self):
"""Defines the recurrent neural network prediction scheme."""
# Dynamic LSTM.
network = tf.contrib.rnn.BasicLSTMCell(self._num_hidden)
output, _ = tf.nn.dynamic_rnn(network, data, dtype=tf.float32)
# Flatten and apply same weights to all time steps.
output = tf.reshape(output, [-1, self._num_hidden])
prediction = tf.nn.softmax(tf.matmul(output, self.weight) + self.bias)
prediction = tf.reshape(prediction,
[-1, self._num_steps, self._num_classes])
return prediction
#lazy_property
def cost(self):
"""Defines the cost function for the network."""
cross_entropy = -tf.reduce_sum(self.target * tf.log(self.prediction),
axis=[1, 2])
cross_entropy = tf.reduce_mean(cross_entropy)
return cross_entropy
#lazy_property
def optimize(self):
"""Defines the optimization scheme."""
learning_rate = 0.003
optimizer = tf.train.RMSPropOptimizer(learning_rate)
return optimizer.minimize(self.cost)
#lazy_property
def error(self):
"""Defines a measure of prediction error."""
mistakes = tf.not_equal(tf.argmax(self.target, 2),
tf.argmax(self.prediction, 2))
return tf.reduce_mean(tf.cast(mistakes, tf.float32))
def _weight_and_bias(self):
"""Returns appropriately sized weight and bias tensors for the output layer."""
self.weight = tf.Variable(tf.truncated_normal(
[self._num_hidden, self._num_classes],
mean=0.0,
stddev=0.01,
dtype=tf.float32))
self.bias = tf.Variable(tf.constant(0.1, shape=[self._num_classes]))
Training
Here is my training process. The all_data class just holds my data and labels, and uses a batch generator class to spit out batches for training when I call all_data.train.next() and all_data.train_labels.next(). You can reproduce with any batch generation scheme you like, and I can add the code if you think it is relevant; I felt like this was getting too long as it is.
tf.reset_default_graph()
data = tf.placeholder(tf.float32,
[None, all_data.num_steps, all_data.num_features])
target = tf.placeholder(tf.float32,
[None, all_data.num_steps, all_data.num_outputs])
model = MyNetwork(data, target, NUM_HIDDEN)
print('Training the model...')
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
print('Initialized.')
for epoch in range(3):
print('Epoch {} |'.format(epoch), end='', flush=True)
for step in range(all_data.train_size // BATCH_SIZE):
# Generate the next training batch and train.
d = all_data.train.next()
t = all_data.train_labels.next()
sess.run(model.optimize,
feed_dict={data: d, target: t})
# Update the user periodically.
if step % summary_frequency == 0:
print('.', end='', flush=True)
# Show training and validation error at the end of each epoch.
print('|', flush=True)
train_error = sess.run(model.error,
feed_dict={data: d, target: t})
valid_error = sess.run(model.error,
feed_dict={
data: all_data.valid,
target: all_data.valid_labels
})
print('Training error: {}%'.format(100 * train_error))
print('Validation error: {}%'.format(100 * valid_error))
# Check testing error after everything.
test_error = sess.run(model.error,
feed_dict={
data: all_data.test,
target: all_data.test_labels
})
print('Testing error after {} epochs: {}%'.format(epoch + 1, 100 * test_error))
For a simple example, I generated random data and labels, where data has shape [num_samples, num_steps, num_features], and each sample has a single label associated with the whole thing:
data = np.random.rand(5000, 1000, 2)
labels = np.random.randint(low=0, high=2, size=[5000])
I then converted my labels to one-hot vectors and tiled them so that the resulting labels tensor was the same size as the data tensor.
Results
No matter what I do, I get results like this:
Training the model...
Initialized.
Epoch 0 |.......................................................|
Training error: 56.25%
Validation error: 53.39999794960022%
Epoch 1 |.......................................................|
Training error: 56.25%
Validation error: 53.39999794960022%
Epoch 2 |.......................................................|
Training error: 56.25%
Validation error: 53.39999794960022%
Testing error after 3 epochs: 49.000000953674316%
Where I have exactly the same error at every epoch. Even if my weights were randomly walking around this should change. For the example shown here, I used random data with random labels, so I do not expect much improvement, but I do expect some change, and I am getting the exact same results every epoch. When I do this with my actual data set, I get the same behavior.
Insight
I hesitate to include this in case it proves to be a red herring, but I believe that my optimizer is calculating cost function gradients of None. When I tried a different optimizer and attempted to clip the gradients, I went ahead and used tf.Print to output the gradients as well. The network crashed with an error that tf.Print could not handle None-type values.
Attempted fixes
I have tried the following things, and the problem persists in all cases:
Using different optimizers, e.g. AdamOptimizer with and without modifications to the gradients (clipping).
Adjusting batch sizes.
Using many more and many fewer hidden nodes.
Running for more epochs.
Initializing my weights with different values assigned to stddev.
Initializing my biases to zeros (using tf.zeros) and to different constants.
Using weights and biases that are defined within the prediction method and are not member variables of the class, and a _weight_and_bias method that is defined as a #staticmethod like in this gist.
Determining logits in the prediction function instead of softmax predictions, i.e. predictions = tf.matmul(output, self.weights) + self.bias, and then using tf.nn.softmax_cross_entropy_with_logits. This requires some reshaping because the method wants its labels and targets given with shape [batch_size, num_classes], so the cost method becomes:
(line added to get code to format...)
#lazy_property
def cost(self):
"""Defines the cost function for the network."""
targs = tf.reshape(self.target, [-1, self._num_classes])
logits = tf.reshape(self.predictions, [-1, self._num_classes])
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(labels=targs, logits=logits)
cross_entropy = tf.reduce_mean(cross_entropy)
return cross_entropy
Changing which size dimension I leave as None when I create my placeholders as suggested in this answer, which requires a bit of rewriting in the network definition. Basically setting size = [all_data.batch_size, -1, all_data.num_features] and size = [all_data.batch_size, -1, all_data.num_classes].
Using tf.contrib.rnn.DropoutWrapper in my network definition and passing a dropout value set to 0.5 in training and 1.0 in validation and testing.
The problem went away when I used
output = tf.contrib.layers.flatten(output)
logits = tf.contrib.layers.fully_connected(output, some_size, activation_fn=None)
instead of flattening my network output, defining weights, and performing the tf.matmul(output, weight) + bias manually. I then used logits (instead of predictions in the question) in my cost function with
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(labels=target,
logits=logits)
If you want to get the network prediction, you will still need to do prediction = tf.nn.softmax(logits).
I have no idea why this helped, but the network would not train even on random made-up data until I made these changes.

Categories