Related
I'm interested in fine-tuning a Mask-RCNN model that I'm using for instance segmentation. Currently I have trained the model for 6 epochs and the various Mask-RCNN losses are as follows:
The reason I'm stopping is that the COCO evaluation metrics seem to have dipped in the last epoch:
I know this is a far reaching question, but I'm looking to gain some intuition of how to understand which parameters are going to be the most impactful in improving the evaluation metrics. I understand there are three places to consider:
Should I be looking at batch size, learning rate and momentum, this uses an SGD optimizer with a learning rate of 1e-4 and batch size 2?
Should I be looking at using more training data or adding augmentation (I don't currently use any) and my dataset is current pretty large 40K images?
Should I be looking at the specific MaskRCNN parameters?
I thing I'll likely be asked to me more specific on what I want to improve so let me say that I would like to improve the recall of the individual masks. The model is performing well but doesn't quite capture the full extend of what I would like it to. I'm also leaving out details of the specific learning problem as I'd like to gain intuition of how to approach this in general.
A couple of notes:
6 epochs are too small for the network to converge even if you use a pre-trained network—especially such a big one as resnet50. I think you need at least 50 epochs. On a pre-trained resnet18 I started to get good results after 30 epochs, resnet34 needed +10-20 epochs and your resnet50 + 40k images of the train set - definitely need more epochs than 6;
definitely use a pre-trained network;
in my experience, I failed to get the results I like with SGD. I started using AdamW + ReduceLROnPlateau scheduler. The network converges quite fast, like 50-60% AP on epoch 7 or 8 but then it comes up to 80-85 after 50-60 epochs using very small improvements from epoch to epoch, only if the LR is small enough. You must be familiar with the gradient descent notion. I used to think of it as if you have more augmentation, your "hill" is covered with "boulders" that you have to be able to bypass and this is only possible if you control the LR. Additionally, AdamW helps with the overfitting.
This is how I do it. For networks with higher input resolution (your input images are scaled on input by the net itself), I use higher LR.
init_lr = 0.00005
weight_decay = init_lr * 100
optimizer = torch.optim.AdamW(params, lr=init_lr, weight_decay=weight_decay)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, verbose=True, patience=3, factor=0.75)
for epoch in range(epochs):
# train for one epoch, printing every 10 iterations
metric_logger = train_one_epoch(model, optimizer, train_loader, scaler, device,
epoch, print_freq=10)
scheduler.step(metric_logger.loss.global_avg)
optimizer.param_groups[0]["weight_decay"] = optimizer.param_groups[0]["lr"] * 100
# scheduler.step()
# evaluate on the test dataset
evaluate(model, test_loader, device=device)
print("[INFO] serializing model to '{}' ...".format(args["model"]))
save_and_print_size_of_model(model, args["model"], script=False)
Find such an LR and weight decay that the training exhausts LR to a very small value, like 1/10 of your initial LR, at the end of the training. If you will have a plateau too often, the scheduler quickly brings it to very small values and the network will learn nothing all the rest of the epochs.
Your plots indicate that your LR is too high at some point in the training, the network stops training and then AP is going down. You need constant improvements, even small ones. The more network trains the more subtle details it learns about your domain and the smaller the learning rate. Imho, constant LR will not allow doing that correctly.
anchor generator settings. Here is how I initialize the network.
def get_maskrcnn_resnet_model(name, num_classes, pretrained, res='normal'):
print('Using maskrcnn with {} backbone...'.format(name))
backbone = resnet_fpn_backbone(name, pretrained=pretrained, trainable_layers=5)
sizes = ((4,), (8,), (16,), (32,), (64,))
aspect_ratios = ((0.25, 0.5, 1.0, 2.0, 4.0),) * len(sizes)
anchor_generator = AnchorGenerator(
sizes=sizes, aspect_ratios=aspect_ratios
)
roi_pooler = torchvision.ops.MultiScaleRoIAlign(featmap_names=['0', '1', '2', '3'],
output_size=7, sampling_ratio=2)
default_min_size = 800
default_max_size = 1333
if res == 'low':
min_size = int(default_min_size / 1.25)
max_size = int(default_max_size / 1.25)
elif res == 'normal':
min_size = default_min_size
max_size = default_max_size
elif res == 'high':
min_size = int(default_min_size * 1.25)
max_size = int(default_max_size * 1.25)
else:
raise ValueError('Invalid res={} param'.format(res))
model = MaskRCNN(backbone, min_size=min_size, max_size=max_size, num_classes=num_classes,
rpn_anchor_generator=anchor_generator, box_roi_pool=roi_pooler)
model.roi_heads.detections_per_img = 512
return model
I need to find small objects here why I use such anchor params.
classes in-balancing issue. If you have only your object and bg - no problem. If you have more classes then make sure that your training split (as 80% for train and 20% for the test) is more or less precisely applied to all the classes used in your particular training.
Good luck!
Below is the code to configure TrainingArguments consumed from the HuggingFace transformers library to finetune the GPT2 language model.
training_args = TrainingArguments(
output_dir="./gpt2-language-model", #The output directory
num_train_epochs=100, # number of training epochs
per_device_train_batch_size=8, # batch size for training #32, 10
per_device_eval_batch_size=8, # batch size for evaluation #64, 10
save_steps=100, # after # steps model is saved
warmup_steps=500,# number of warmup steps for learning rate scheduler
prediction_loss_only=True,
metric_for_best_model = "eval_loss",
load_best_model_at_end = True,
evaluation_strategy="epoch",
learning_rate=0.00004, # learning rate
)
early_stop_callback = EarlyStoppingCallback(early_stopping_patience = 3)
trainer = Trainer(
model=gpt2_model,
args=training_args,
data_collator=data_collator,
train_dataset=train_dataset,
eval_dataset=test_dataset,
callbacks = [early_stop_callback],
)
The number of epochs as 100 and learning_rate as 0.00004 and also the early_stopping is configured with the patience value as 3.
The model ran for 5/100 epochs and noticed that the difference in loss_value is negligible. The latest checkpoint is saved as checkpoint-latest.
Now Can I modify the learning_rate may be to 0.01 from 0.00004 and resume the training from the latest saved checkpoint - checkpoint-latest? Doing that will be efficient?
Or to train with the new learning_rate value should I start the training from the beginning?
No, you don't have to restart your training.
Changing the learning rate is like changing how big a step your model take in the direction determined by your loss function.
You can also think of it as transfer learning where the model has some experience (no matter how little or irrelevant) and the weights are in a state most likely better than a randomly initialised one.
As a matter of fact, changing the learning rate mid-training is considered an art in deep learning and you should change it if you have a very very good reason to do it.
You would probably want to write down when (why, what, etc) you did it if you or someone else wants to "reproduce" the result of your model.
Pytorch provides several methods to adjust the learning_rate: torch.optim.lr_scheduler.
Check the docs for usage https://pytorch.org/docs/stable/optim.html#how-to-adjust-learning-rate
I've been trying to figure out how to compute the number of Flops in backward pass of ResNet. For forward pass, it seems straightforward: apply the conv filters to the input for each layer. But how does one do the Flops counts for gradient computation and update of all weights during the backward pass?
Specifically,
how to compute Flops in gradient computations for each layer?
what all gradients need to be computed so Flops for each of those can be counted?
How many Flops in computation of gradient for Pool, BatchNorm, and Relu layers?
I understand the chain rule for gradient computation, but having a hard time formulating how it'd apply to weight filters in conv layers of ResNet and how many Flops each of those would take. It'd be very useful to get any comments about method to compute total Flops for Backward pass. Thanks
You can definitely count the number of multiplication, addition for the backward pass manually, but I guess that's an exhaustive process for complex models.
Usually, most models are benchmarked with flops for a forward pass instead of backward flop count for CNN and other models. I guess the reason has to do with the inference being more important in terms of different CNN variants and other deep learning models in the application.
The backward pass is only important while training, and for most of the simple models, the backward and forward flops should be close with some constant factors.
So, I tried a hacky approach to calculate the gradients for the whole resnet model in the graph to get the flop counts for both forward pass and gradient calculation and then subtracted the forward flops. It's not an exact measurement, may miss many operations for a complex graph/model.
But this may give a flop estimate for most models.
[Following code snippet works with tensorflow 2.0]
import tensorflow as tf
def get_flops():
for_flop = 0
total_flop = 0
session = tf.compat.v1.Session()
graph = tf.compat.v1.get_default_graph()
# forward
with graph.as_default():
with session.as_default():
model = tf.keras.applications.ResNet50() # change your model here
run_meta = tf.compat.v1.RunMetadata()
opts = tf.compat.v1.profiler.ProfileOptionBuilder.float_operation()
# We use the Keras session graph in the call to the profiler.
flops = tf.compat.v1.profiler.profile(graph=graph,
run_meta=run_meta, cmd='op', options=opts)
for_flop = flops.total_float_ops
# print(for_flop)
# forward + backward
with graph.as_default():
with session.as_default():
model = tf.keras.applications.ResNet50() # change your model here
outputTensor = model.output
listOfVariableTensors = model.trainable_weights
gradients = tf.gradients(outputTensor, listOfVariableTensors)
run_meta = tf.compat.v1.RunMetadata()
opts = tf.compat.v1.profiler.ProfileOptionBuilder.float_operation()
# We use the Keras session graph in the call to the profiler.
flops = tf.compat.v1.profiler.profile(graph=graph,
run_meta=run_meta, cmd='op', options=opts)
total_flop = flops.total_float_ops
# print(total_flop)
return for_flop, total_flop
for_flops, total_flops = get_flops()
print(f'forward: {for_flops}')
print(f'backward: {total_flops - for_flops}')
Out:
51112224
102224449
forward: 51112224
backward: 51112225
I'm currently working on a project in pytorch on Wasserstein GAN (https://arxiv.org/pdf/1701.07875.pdf).
In Wasserstain GAN a new objective function is defined using the wasserstein distance as :
Which leads to the following algorithms for training the GAN:
My question is :
When implementing line 5 and 6 of the algorithm in pytorch should I be multiplying my loss -1 ? As in my code (I use RMSprop as my optimizer for both the generator and critic):
############################
# (1) Update D network: maximize (D(x)) + (D(G(x)))
###########################
for n in range(n_critic):
D.zero_grad()
real_cpu = data[0].to(device)
b_size = real_cpu.size(0)
output = D(real_cpu)
#errD_real = -criterion(output, label) #DCGAN
errD_real = torch.mean(output)
# Calculate gradients for D in backward pass
errD_real.backward()
D_x = output.mean().item()
## Train with all-fake batch
# Generate batch of latent vectors
noise = torch.randn(b_size, 100, device=device) #Careful here we changed shape of input (original : torch.randn(4, 100, 1, 1, device=device))
# Generate fake image batch with G
fake = G(noise)
# Classify all fake batch with D
output = D(fake.detach())
# Calculate D's loss on the all-fake batch
errD_fake = torch.mean(output)
# Calculate the gradients for this batch
errD_fake.backward()
D_G_z1 = output.mean().item()
# Add the gradients from the all-real and all-fake batches
errD = -(errD_real - errD_fake)
# Update D
optimizerD.step()
#Clipping weights
for p in D.parameters():
p.data.clamp_(-0.01, 0.01)
As you can see, I do the operation errD = -(errD_real - errD_fake), with errD_real and errD_fake being respectively the mean of the predictions of the critic on real and fake samples.
To my understanding RMSprop should optimize the weights of the critic the following way :
w <- w - alpha*gradient(w)
(alpha being the learning rate divided by the square root of the weighted moving average of the squared gradient)
Since the optimization problem requires to "go" in the same direction as the gradient it should be required to multiply gradient(w) by -1 before optimizing the weights.
Do you think that my reasoning is right ?
The program runs but my results are quiet poor.
I follow the same logic for the generator's weights but this time in order to go in the opposite direction of the gradient:
############################
# (2) Update G network: minimize -D(G(x))
###########################
G.zero_grad()
noise = torch.randn(b_size, 100, device=device)
fake = G(noise)
#label.fill_(fake_label) # fake labels are real for generator cost
# Since we just updated D, perform another forward pass of all-fake batch through D
output = D(fake).view(-1)
# Calculate G's loss based on this output
#errG = criterion(output, label) #DCGAN
errG = -torch.mean(output)
# Calculate gradients for G
errG.backward()
D_G_z2 = output.mean().item()
# Update G
optimizerG.step()
Sorry for the long question, I tried to explain my doubt as clear as possible. Thank you everyone.
I noticed some errors in the implementation of your discriminator training protocol. You call your backward functions twice with both the real and fake values loss being backpropagated at different time steps.
Technically an implementation using this scheme is possible but highly unreadable. There was a mistake with your errD_real in which your output is going to be positive instead of negative as an optimal D(G(z))>0 and so you penalize it for being correct. Overall your model converges simply by predicting D(x)<0 for all inputs.
To fix this do not call your errD_readl.backward() or your errD_fake.backward(). Simply using an errD.backward() after you define errD would work perfectly fine. Otherwise, your generator seems to be correct.
When using a Keras LSTM to predict on time series data I've been getting errors when I'm trying to train the model using a batch size of 50, while then trying to predict on the same model using a batch size of 1 (ie just predicting the next value).
Why am I not able to train and fit the model with multiple batches at once, and then use that model to predict for anything other than the same batch size. It doesn't seem to make sense, but then I could easily be missing something about this.
Edit: this is the model. batch_size is 50, sl is sequence length, which is set at 20 currently.
model = Sequential()
model.add(LSTM(1, batch_input_shape=(batch_size, 1, sl), stateful=True))
model.add(Dense(1))
model.compile(loss='mean_squared_error', optimizer='adam')
model.fit(trainX, trainY, epochs=epochs, batch_size=batch_size, verbose=2)
here is the line for predicting on the training set for RMSE
# make predictions
trainPredict = model.predict(trainX, batch_size=batch_size)
here is the actual prediction of unseen time steps
for i in range(test_len):
print('Prediction %s: ' % str(pred_count))
next_pred_res = np.reshape(next_pred, (next_pred.shape[1], 1, next_pred.shape[0]))
# make predictions
forecastPredict = model.predict(next_pred_res, batch_size=1)
forecastPredictInv = scaler.inverse_transform(forecastPredict)
forecasts.append(forecastPredictInv)
next_pred = next_pred[1:]
next_pred = np.concatenate([next_pred, forecastPredict])
pred_count += 1
This issue is with the line:
forecastPredict = model.predict(next_pred_res, batch_size=batch_size)
The error when batch_size here is set to 1 is:
ValueError: Cannot feed value of shape (1, 1, 2) for Tensor 'lstm_1_input:0', which has shape '(10, 1, 2)' which is the same error that throws when batch_size here is set to 50 like the other batch sizes as well.
The total error is:
forecastPredict = model.predict(next_pred_res, batch_size=1)
File "/home/entelechy/tf_keras/lib/python3.5/site-packages/keras/models.py", line 899, in predict
return self.model.predict(x, batch_size=batch_size, verbose=verbose)
File "/home/entelechy/tf_keras/lib/python3.5/site-packages/keras/engine/training.py", line 1573, in predict
batch_size=batch_size, verbose=verbose)
File "/home/entelechy/tf_keras/lib/python3.5/site-packages/keras/engine/training.py", line 1203, in _predict_loop
batch_outs = f(ins_batch)
File "/home/entelechy/tf_keras/lib/python3.5/site-packages/keras/backend/tensorflow_backend.py", line 2103, in __call__
feed_dict=feed_dict)
File "/home/entelechy/tf_keras/lib/python3.5/site-packages/tensorflow/python/client/session.py", line 767, in run
run_metadata_ptr)
File "/home/entelechy/tf_keras/lib/python3.5/site-packages/tensorflow/python/client/session.py", line 944, in _run
% (np_val.shape, subfeed_t.name, str(subfeed_t.get_shape())))
ValueError: Cannot feed value of shape (1, 1, 2) for Tensor 'lstm_1_input:0', which has shape '(10, 1, 2)'
Edit: Once I set the model to stateful=False then I am able to use different batch sizes for fitting/training and prediction. What is the reason for this?
Unfortunately what you want to do is impossible with Keras ... I've also struggle a lot of time on this problems and the only way is to dive into the rabbit hole and work with Tensorflow directly to do LSTM rolling prediction.
First, to be clear on terminology, batch_size usually means number of sequences that are trained together, and num_steps means how many time steps are trained together. When you mean batch_size=1 and "just predicting the next value", I think you meant to predict with num_steps=1.
Otherwise, it should be possible to train and predict with batch_size=50 meaning you are training on 50 sequences and make 50 predictions every time step, one for each sequence (meaning training/prediction num_steps=1).
However, I think what you mean is that you want to use stateful LSTM to train with num_steps=50 and do prediction with num_steps=1. Theoretically this make senses and should be possible, and it is possible with Tensorflow, just not Keras.
The problem: Keras requires an explicit batch size for stateful RNN. You must specify batch_input_shape (batch_size, num_steps, features).
The reason: Keras must allocate a fixed-size hidden state vector in the computation graph with shape (batch_size, num_units) in order to persist the values between training batches. On the other hand, when stateful=False, the hidden state vector can be initialized dynamically with zeroes at the beginning of each batch so it does not need to be a fixed size. More details here: http://philipperemy.github.io/keras-stateful-lstm/
Possible work around: Train and predict with num_steps=1. Example: https://github.com/keras-team/keras/blob/master/examples/lstm_stateful.py. This might or might not work at all for your problem as the gradient for back propagation will be computed on only one time step. See: https://github.com/fchollet/keras/issues/3669
My solution: use Tensorflow: In Tensorflow you can train with batch_size=50, num_steps=100, then do predictions with batch_size=1, num_steps=1. This is possible by creating a different model graph for training and prediction sharing the same RNN weight matrices. See this example for next-character prediction: https://github.com/sherjilozair/char-rnn-tensorflow/blob/master/model.py#L11 and blog post http://karpathy.github.io/2015/05/21/rnn-effectiveness/. Note that one graph can still only work with one specified batch_size, but you can setup multiple model graphs sharing weights in Tensorflow.
Sadly what you wish for is impossible because you specify the batch_size when you define the model...
However, I found a simple way around this problem: create 2 models! The first is used for training and the second for predictions, and have them share weights:
train_model = Sequential([Input(batch_input_shape=(batch_size,...),
<continue specifying your model>])
predict_model = Sequential([Input(batch_input_shape=(1,...),
<continue specifying exact same model>])
train_model.compile(loss='sparse_categorical_crossentropy', optimizer=Adam())
predict_model.compile(loss='sparse_categorical_crossentropy', optimizer=Adam())
Now you can use any batch size you want. after you fit your train_model just save it's weights and load them with the predict_model:
train_model.save_weights('lstm_model.h5')
predict_model.load_weights('lstm_model.h5')
notice that you only want to save and load the weights, and not the whole model (which includes the architecture, optimizer etc...). This way you get the weights but you can input one batch at a time...
more on keras save/load models:
https://keras.io/getting-started/faq/#how-can-i-save-a-keras-model
notice that you need to install h5py to use "save weights".
Another easy workaround is:
def create_model(batch_size):
model = Sequential()
model.add(LSTM(1, batch_input_shape=(batch_size, 1, sl), stateful=True))
model.add(Dense(1))
return model
model_train = create_model(batch_size=50)
model_train.compile(loss='mean_squared_error', optimizer='adam')
model_train.fit(trainX, trainY, epochs=epochs, batch_size=batch_size)
model_predict = create_model(batch_size=1)
weights = model_train.get_weights()
model_predict.set_weights(weights)
The best solution to this problem is "Copy Weights". It can be really helpful if you want to train & predict with your LSTM model with different batch sizes.
For example, once you have trained your model with 'n' batch size as shown below:
# configure network
n_batch = len(X)
n_epoch = 1000
n_neurons = 10
# design network
model = Sequential()
model.add(LSTM(n_neurons, batch_input_shape=(n_batch, X.shape[1], X.shape[2]), stateful=True))
model.add(Dense(1))
model.compile(loss='mean_squared_error', optimizer='adam')
And now you want to want predict values fewer than your batch size where n=1.
What you can do is that, copy the weights of your fit model and reinitialize the new model LSTM model with same architecture and set batch size equal to 1.
# re-define the batch size
n_batch = 1
# re-define model
new_model = Sequential()
new_model.add(LSTM(n_neurons, batch_input_shape=(n_batch, X.shape[1], X.shape[2]), stateful=True))
new_model.add(Dense(1))
# copy weights
old_weights = model.get_weights()
new_model.set_weights(old_weights)
Now you can easily predict and train LSTMs with different batch sizes.
For more information please read: https://machinelearningmastery.com/use-different-batch-sizes-training-predicting-python-keras/
I found below helpful (and fully inline with above). The section "Solution 3: Copy Weights" worked for me:
How to use Different Batch Sizes when Training and Predicting with LSTMs, by Jason Brownlee
n_neurons = 10
# design network
model = Sequential()
model.add(LSTM(n_neurons, batch_input_shape=(n_batch, X.shape[1], X.shape[2]), stateful=True))
model.add(Dense(1))
model.compile(loss='mean_squared_error', optimizer='adam')
# fit network
for i in range(n_epoch):
model.fit(X, y, epochs=1, batch_size=n_batch, verbose=1, shuffle=False)
model.reset_states()
# re-define the batch size
n_batch = 1
# re-define model
new_model = Sequential()
new_model.add(LSTM(n_neurons, batch_input_shape=(n_batch, X.shape[1], X.shape[2]), stateful=True))
new_model.add(Dense(1))
# copy weights
old_weights = model.get_weights()
new_model.set_weights(old_weights)
# compile model
new_model.compile(loss='mean_squared_error', optimizer='adam')
I also have same problem and resolved it.
In another way, you can save your weights, when you test your result, you can reload your model with same architecture and set batch_size=1 as below:
n_neurons = 10
# design network
model = Sequential()
model.add(LSTM(n_neurons, batch_size=1, batch_input_shape=(n_batch,X.shape[1], X.shape[2]), statefull=True))
model.add(Dense(1))
model.compile(loss='mean_squared_error', optimizer='adam')
model.load_weights("w.h5")
It will work well. I hope it will helpfull for you.
If you don't have access to the code that created the model or if you just don't want your prediction/validation code to depend on your model creation and training code there is another way:
You could create a new model from a modified version of the loaded model's config like this:
loaded_model = tf.keras.models.load_model('model_file.h5')
config = loaded_model.get_config()
old_batch_input_shape = config['layers'][0]['config']['batch_input_shape']
config['layers'][0]['config']['batch_input_shape'] = (new_batch_size, old_batch_input_shape[1])
new_model = loaded_model.__class__.from_config(config)
new_model.set_weights(loaded_model.get_weights())
This works well for me in a situation where I have several different models with state-full RNN layers working together in a graph network but being trained separately with different networks leading to different batch sizes. It allows me to experiment with the model structures and training batches without needing to change anything in my validation script.