I want to train on my local gpu but it's only running on cpu while torch.cuda.is_available() is actually true and i can see my gpu but it runs only on cpu , so how to fix it
my CNN model:
import torch.nn as nn
import torch.nn.functional as F
from PIL import ImageFile
ImageFile.LOAD_TRUNCATED_IMAGES = True
# define the CNN architecture
class Net(nn.Module):
### TODO: choose an architecture, and complete the class
def __init__(self):
super(Net, self).__init__()
## Define layers of a CNN
self.conv1 = nn.Conv2d(3, 16, 3, padding=1)
# convolutional layer (sees 16x16x16 tensor)
self.conv2 = nn.Conv2d(16, 32, 3, padding=1)
# convolutional layer (sees 8x8x32 tensor)
self.conv3 = nn.Conv2d(32, 64, 3, padding=1)
# max pooling layer
self.pool = nn.MaxPool2d(2, 2)
# linear layer (64 * 4 * 4 -> 500)
self.fc1 = nn.Linear(64 * 28 * 28, 500)
# linear layer (500 -> 10)
self.fc2 = nn.Linear(500, 133)
# dropout layer (p=0.25)
self.dropout = nn.Dropout(0.25)
def forward(self, x):
## Define forward behavior
x = self.pool(F.relu(self.conv1(x)))
#print(x.shape)
x = self.pool(F.relu(self.conv2(x)))
#print(x.shape)
x = self.pool(F.relu(self.conv3(x)))
#print(x.shape)
#print(x.shape)
# flatten image input
x = x.view(-1, 64 * 28 * 28)
# add dropout layer
x = self.dropout(x)
# add 1st hidden layer, with relu activation function
x = F.relu(self.fc1(x))
# add dropout layer
x = self.dropout(x)
# add 2nd hidden layer, with relu activation function
x = self.fc2(x)
return x
#-#-# You so NOT have to modify the code below this line. #-#-#
# instantiate the CNN
model_scratch = Net()
# move tensors to GPU if CUDA is available
if use_cuda:
print("TRUE")
model_scratch = model_scratch.cuda()
train function :
def train(n_epochs, loaders, model, optimizer, criterion, use_cuda, save_path):
"""returns trained model"""
# initialize tracker for minimum validation loss
valid_loss_min = np.Inf
loaders_scratch = {'train': train_loader,'valid': valid_loader,'test': test_loader}
for epoch in range(1, n_epochs+1):
# initialize variables to monitor training and validation loss
train_loss = 0.0
valid_loss = 0.0
###################
# train the model #
###################
model.train()
for batch_idx, (data, target) in enumerate(loaders['train']):
# move to GPU
if use_cuda:
data, target = data.cuda(), target.cuda()
## find the loss and update the model parameters accordingly
## record the average training loss, using something like
## train_loss = train_loss + ((1 / (batch_idx + 1)) * (loss.data - train_loss))
# clear the gradients of all optimized variables
optimizer.zero_grad()
# forward pass: compute predicted outputs by passing inputs to the model
output = model(data)
# calculate the batch loss
loss = criterion(output, target)
# backward pass: compute gradient of the loss with respect to model parameters
loss.backward()
# perform a single optimization step (parameter update)
optimizer.step()
# update training loss
train_loss += loss.item()*data.size(0)
######################
# validate the model #
######################
model.eval()
for batch_idx, (data, target) in enumerate(loaders['valid']):
# move to GPU
if use_cuda:
data, target = data.cuda(), target.cuda()
## update the average validation loss
output = model(data)
# calculate the batch loss
loss = criterion(output, target)
# update average validation loss
valid_loss += loss.item()*data.size(0)
# calculate average losses
train_loss = train_loss/len(train_loader.dataset)
valid_loss = valid_loss/len(valid_loader.dataset)
# print training/validation statistics
print('Epoch: {} \tTraining Loss: {:.6f} \tValidation Loss: {:.6f}'.format(
epoch,
train_loss,
valid_loss
))
## TODO: save the model if validation loss has decreased
if valid_loss <= valid_loss_min:
print('Validation loss decreased ({:.6f} --> {:.6f}). Saving model ...'.format(
valid_loss_min,
valid_loss))
torch.save(model.state_dict(), save_path)
valid_loss_min = valid_loss
# return trained model
return model
# train the model
loaders_scratch = {'train': train_loader,'valid': valid_loader,'test': test_loader}
model_scratch = train(100, loaders_scratch, model_scratch, optimizer_scratch,
criterion_scratch, use_cuda, 'model_scratch.pt')
# load the model that got the best validation accuracy
model_scratch.load_state_dict(torch.load('model_scratch.pt'))
while i am getting "TRUE" in torch.cuda.is_available() but still not running on GPU
i am only running on CPU
the below picture shows that i am running on cpu with 62%
To utilize cuda in pytorch you have to specify that you want to run your code on gpu device.
a line of code like:
use_cuda = torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
will determine whether you have cuda available and if so, you will have it as your device.
later in the code you have to pass your tensors and model to this device:
net = net.to(device)
and do the same for your other tensors that need to go to gpu, like test and training values.
If you are experiencing an issue where your model is only using the CPU during training even though your GPU is available, it's likely due to the data loading and transformation process. When loading images from your local directory and applying transforms on your data, the majority of the time during training is spent on the data loading process, which is performed on the CPU.
To resolve this issue, you can preprocess your data by applying your custom transforms once and then saving the results. This way, when you load the preprocessed data, you can take advantage of the GPU's performance during training. This can help to significantly improve the training time of your model.
In summary, if you are facing a problem with model using CPU instead of GPU during training, it could be due to the data loading process. To fix this, preprocess your data and save the results, then use the preprocessed data while training. This will allow you to take advantage of the GPU's performance and reduce training time.
I am building a simple Sequential model in Keras (tensorflow backend). During training I want to inspect the individual training batches and model predictions. Therefore, I am trying to create a custom Callback that saves the model predictions and targets for each training batch. However, the model is not using the current batch for prediction, but the entire training data.
How can I hand over only the current training batch to the Callback?
And how can I access the batches and targets that the Callback saves in self.predhis and self.targets?
My current version looks as follows:
callback_list = [prediction_history((self.x_train, self.y_train))]
self.model.fit(self.x_train, self.y_train, batch_size=self.batch_size, epochs=self.n_epochs, validation_data=(self.x_val, self.y_val), callbacks=callback_list)
class prediction_history(keras.callbacks.Callback):
def __init__(self, train_data):
self.train_data = train_data
self.predhis = []
self.targets = []
def on_batch_end(self, epoch, logs={}):
x_train, y_train = self.train_data
self.targets.append(y_train)
prediction = self.model.predict(x_train)
self.predhis.append(prediction)
tf.logging.info("Prediction shape: {}".format(prediction.shape))
tf.logging.info("Targets shape: {}".format(y_train.shape))
NOTE: this answer is outdated and only works with TF1. Check #bers's answer for a solution tested on TF2.
After model compilation, the placeholder tensor for y_true is in model.targets and y_pred is in model.outputs.
To save the values of these placeholders at each batch, you can:
First copy the values of these tensors into variables.
Evaluate these variables in on_batch_end, and store the resulting arrays.
Now step 1 is a bit involved because you'll have to add an tf.assign op to the training function model.train_function. Using current Keras API, this can be done by providing a fetches argument to K.function() when the training function is constructed.
In model._make_train_function(), there's a line:
self.train_function = K.function(inputs,
[self.total_loss] + self.metrics_tensors,
updates=updates,
name='train_function',
**self._function_kwargs)
The fetches argument containing the tf.assign ops can be provided via model._function_kwargs (only works after Keras 2.1.0).
As an example:
from keras.layers import Dense
from keras.models import Sequential
from keras.callbacks import Callback
from keras import backend as K
import tensorflow as tf
import numpy as np
class CollectOutputAndTarget(Callback):
def __init__(self):
super(CollectOutputAndTarget, self).__init__()
self.targets = [] # collect y_true batches
self.outputs = [] # collect y_pred batches
# the shape of these 2 variables will change according to batch shape
# to handle the "last batch", specify `validate_shape=False`
self.var_y_true = tf.Variable(0., validate_shape=False)
self.var_y_pred = tf.Variable(0., validate_shape=False)
def on_batch_end(self, batch, logs=None):
# evaluate the variables and save them into lists
self.targets.append(K.eval(self.var_y_true))
self.outputs.append(K.eval(self.var_y_pred))
# build a simple model
# have to compile first for model.targets and model.outputs to be prepared
model = Sequential([Dense(5, input_shape=(10,))])
model.compile(loss='mse', optimizer='adam')
# initialize the variables and the `tf.assign` ops
cbk = CollectOutputAndTarget()
fetches = [tf.assign(cbk.var_y_true, model.targets[0], validate_shape=False),
tf.assign(cbk.var_y_pred, model.outputs[0], validate_shape=False)]
model._function_kwargs = {'fetches': fetches} # use `model._function_kwargs` if using `Model` instead of `Sequential`
# fit the model and check results
X = np.random.rand(10, 10)
Y = np.random.rand(10, 5)
model.fit(X, Y, batch_size=8, callbacks=[cbk])
Unless the number of samples can be divided by the batch size, the final batch will have a different size than other batches. So K.variable() and K.update() can't be used in this case. You'll have to use tf.Variable(..., validate_shape=False) and tf.assign(..., validate_shape=False) instead.
To verify the correctness of the saved arrays, you can add one line in training.py to print out the shuffled index array:
if shuffle == 'batch':
index_array = _batch_shuffle(index_array, batch_size)
elif shuffle:
np.random.shuffle(index_array)
print('Index array:', repr(index_array)) # Add this line
batches = _make_batches(num_train_samples, batch_size)
The shuffled index array should be printed out during fitting:
Epoch 1/1
Index array: array([8, 9, 3, 5, 4, 7, 1, 0, 6, 2])
10/10 [==============================] - 0s 23ms/step - loss: 0.5670
And you can check if cbk.targets is the same as Y[index_array]:
index_array = np.array([8, 9, 3, 5, 4, 7, 1, 0, 6, 2])
print(Y[index_array])
[[ 0.75325592 0.64857277 0.1926653 0.7642865 0.38901153]
[ 0.77567689 0.13573623 0.4902501 0.42897559 0.55825652]
[ 0.33760938 0.68195038 0.12303088 0.83509441 0.20991668]
[ 0.98367778 0.61325065 0.28973401 0.28734073 0.93399794]
[ 0.26097574 0.88219054 0.87951941 0.64887846 0.41996446]
[ 0.97794604 0.91307569 0.93816428 0.2125808 0.94381495]
[ 0.74813435 0.08036688 0.38094272 0.83178364 0.16713736]
[ 0.52609421 0.39218962 0.21022047 0.58569125 0.08012982]
[ 0.61276627 0.20679494 0.24124858 0.01262245 0.0994412 ]
[ 0.6026137 0.25620512 0.7398164 0.52558182 0.09955769]]
print(cbk.targets)
[array([[ 0.7532559 , 0.64857274, 0.19266529, 0.76428652, 0.38901153],
[ 0.77567691, 0.13573623, 0.49025011, 0.42897558, 0.55825651],
[ 0.33760938, 0.68195039, 0.12303089, 0.83509439, 0.20991668],
[ 0.9836778 , 0.61325067, 0.28973401, 0.28734073, 0.93399793],
[ 0.26097575, 0.88219053, 0.8795194 , 0.64887846, 0.41996446],
[ 0.97794604, 0.91307569, 0.93816429, 0.2125808 , 0.94381493],
[ 0.74813437, 0.08036689, 0.38094273, 0.83178365, 0.16713737],
[ 0.5260942 , 0.39218962, 0.21022047, 0.58569127, 0.08012982]], dtype=float32),
array([[ 0.61276627, 0.20679495, 0.24124858, 0.01262245, 0.0994412 ],
[ 0.60261369, 0.25620511, 0.73981643, 0.52558184, 0.09955769]], dtype=float32)]
As you can see, there are two batches in cbk.targets (one "full batch" of size 8 and the final batch of size 2), and the row order is the same as Y[index_array].
Long edit (almost a new answer) for the following reasons:
Yu-Yang's 2017 answer relies on the private _make_train_function and _function_kwargs APIs, which work only in TF1 (and maybe in TF1 compatibility, so-called non-eager mode).
Similarly, Binyan Hu's 2020 answer relies on _make_test_function and does not work in TF2 by default (requiring non-eager mode as well).
My own Jan 2020 answer, which was already subject to several required configuration settings, seems to have stopped working with (or before) TF 2.5, and I was not able to make model.inputs or model.outputs work any longer.
Finally, the earlier version of this answer requires potentially expensive model evaluation to obtain the predictions for each batch. A similar solution to obtain activation histograms even led to OOM issues with repeated training of different models.
So I set out find a way to obtain all possible quantities (inputs, targets, predictions, activations), batch-wise, without using any private APIs. The aim was to be able to call .numpy() on the intended quantities, so Keras callbacks can run ordinary Python code to ease debugging (I suppose that is what this question is mainly about - for maximum performance, one would probably try to integrate as many computations as possible into TensorFlow's graph operations anyway).
This is the common base model for all solutions:
"""Demonstrate batch data access."""
import tensorflow as tf
from tensorflow import keras
class DataCallback(keras.callbacks.Callback):
"""This class is where all implementations differ."""
def tf_nan(dtype):
"""Create NaN variable of proper dtype and variable shape for assign()."""
return tf.Variable(float("nan"), dtype=dtype, shape=tf.TensorShape(None))
def main():
"""Run main."""
model = keras.Sequential([keras.layers.Dense(1, input_shape=(2,))])
callback = DataCallback()
model.compile(loss="mse", optimizer="adam")
model.fit(
x=tf.transpose(tf.range(7.0) + [[0.2], [0.4]]),
y=tf.transpose(tf.range(7.0) + 10 + [[0.5]]),
validation_data=(
tf.transpose(tf.range(11.0) + 30 + [[0.6], [0.7]]),
tf.transpose(tf.range(11.0) + 40 + [[0.9]]),
),
shuffle=False,
batch_size=3,
epochs=2,
verbose=0,
callbacks=[callback],
)
model.save("tmp.tf")
if __name__ == "__main__":
main()
The following three snippets show one possible solution each, each with their own pros and cons. The core trick is always the same: allocate a tf.Variable and use tf.Variable.assign to export the intended quantity, from some Keras code run in graph mode, into the callback. The methods differ slightly in callback initialization and (in one case) model compilation, and most importantly, in the quantities they can access, which is why I summarize them above each snippet.
Custom metric
Using a custom (fake) metric (similar to my Jan 2020 answer), while we cannot seem to access model.inputs nor model.outputs any more (and model.(_)targets does not even exist any longer), we can access y_true and y_pred, which represent the model targets and outputs:
[ ] Inputs/Samples (x)
[ ] Weights (w)
[+] Targets/Labels (y_true)
[+] Outputs/Predictions (y_pred)
[ ] All layers (or only final input/output layers)
"""Demonstrate batch data access using a custom metric."""
import tensorflow as tf
from tensorflow import keras
class DataCallback(keras.callbacks.Callback): # diff
"""Callback to operate on batch data from metric."""
def __init__(self):
"""Offer a metric to access batch data."""
super().__init__()
self.y_true = None
self.y_pred = None
def set_model(self, model):
"""Initialize variables when model is set."""
self.y_true = tf_nan(model.output.dtype)
self.y_pred = tf_nan(model.output.dtype)
def metric(self, y_true, y_pred):
"""Fake metric."""
self.y_true.assign(y_true)
self.y_pred.assign(y_pred)
return 0
def on_train_batch_end(self, _batch, _logs=None):
"""See keras.callbacks.Callback.on_train_batch_end."""
print("y_true =", self.y_true.numpy())
print("y_pred =", self.y_pred.numpy())
def on_train_end(self, _logs=None):
"""Clean up."""
del self.y_true, self.y_pred
def tf_nan(dtype):
"""Create NaN variable of proper dtype and variable shape for assign()."""
return tf.Variable(float("nan"), dtype=dtype, shape=tf.TensorShape(None))
def main():
"""Run main."""
model = keras.Sequential([keras.layers.Dense(1, input_shape=(2,))])
callback = DataCallback()
model.compile(loss="mse", optimizer="adam", metrics=[callback.metric]) # diff
model.fit(
x=tf.transpose(tf.range(7.0) + [[0.2], [0.4]]),
y=tf.transpose(tf.range(7.0) + 10 + [[0.5]]),
validation_data=(
tf.transpose(tf.range(11.0) + 30 + [[0.6], [0.7]]),
tf.transpose(tf.range(11.0) + 40 + [[0.9]]),
),
shuffle=False,
batch_size=3,
epochs=2,
verbose=0,
callbacks=[callback],
)
model.save("tmp.tf")
if __name__ == "__main__":
main()
Custom training step
A custom training step is what I used in an earlier version of this answer. The idea still works in principle, but y_pred can be expensive and it might make sense to use a custom metric (see above) if that is required.
[+] Inputs/Samples (x)
[+] Weights (w)
[+] Targets/Labels (y_true)
[~] Outputs/Predictions (y_pred) [expensive!]
[ ] All layers (or only final input/output layers)
"""Demonstrate batch data access using a custom training step."""
import tensorflow as tf
from tensorflow import keras
class DataCallback(keras.callbacks.Callback): # diff
"""Callback to operate on batch data from training step."""
def __init__(self):
"""Initialize tf.Variables."""
super().__init__()
self.x = None
self.w = None
self.y_true = None
self.y_pred = None
def set_model(self, model):
"""Wrap the model.train_step function to access training batch data."""
self.x = tf_nan(model.input.dtype)
# pylint:disable=protected-access (replace by proper dtype if you know it)
if model.compiled_loss._user_loss_weights is not None:
self.w = tf_nan(model.compiled_loss._user_loss_weights.dtype)
self.y_true = tf_nan(model.output.dtype)
self.y_pred = tf_nan(model.output.dtype)
model_train_step = model.train_step
def outer_train_step(data):
# https://github.com/keras-team/keras/blob/v2.7.0/keras/engine/training.py
x, y_true, w = keras.utils.unpack_x_y_sample_weight(data)
self.x.assign(x)
if w is not None:
self.w.assign(w)
self.y_true.assign(y_true)
result = model_train_step(data)
y_pred = model(x)
self.y_pred.assign(y_pred)
return result
model.train_step = outer_train_step
def on_train_batch_end(self, _batch, _logs=None):
"""See keras.callbacks.Callback.on_train_batch_end."""
print("x =", self.x.numpy())
if self.w is not None:
print("w =", self.w.numpy())
print("y_true =", self.y_true.numpy())
print("y_pred =", self.y_pred.numpy())
def on_train_end(self, _logs=None):
"""Clean up."""
del self.x, self.w, self.y_true, self.y_pred
def tf_nan(dtype):
"""Create NaN variable of proper dtype and variable shape for assign()."""
return tf.Variable(float("nan"), dtype=dtype, shape=tf.TensorShape(None))
def main():
"""Run main."""
model = keras.Sequential([keras.layers.Dense(1, input_shape=(2,))])
callback = DataCallback()
model.compile(loss="mse", optimizer="adam")
model.fit(
x=tf.transpose(tf.range(7.0) + [[0.2], [0.4]]),
y=tf.transpose(tf.range(7.0) + 10 + [[0.5]]),
validation_data=(
tf.transpose(tf.range(11.0) + 30 + [[0.6], [0.7]]),
tf.transpose(tf.range(11.0) + 40 + [[0.9]]),
),
shuffle=False,
batch_size=3,
epochs=2,
verbose=0,
callbacks=[callback],
)
model.save("tmp.tf")
if __name__ == "__main__":
main()
Custom layer call
A custom layer call is a super-flexible way of accessing each layer's inputs and outputs. The callback handles patching of the call functions for a list of layers. While we cannot access weights and targets (as these quantitities do not make sense at the level of individual layers), it allows us to access individual layer activations, which can be handy for questions such as How does one log activations using `tf.keras.callbacks.TensorBoard`?.
[+] Inputs/Samples (x)
[ ] Weights (w)
[ ] Targets/Labels (y_true)
[+] Outputs/Predictions (y_pred)
[+] All layers (or only final input/output layers)
"""Demonstrate batch data access using custom layer calls."""
import tensorflow as tf
from tensorflow import keras
class DataCallback(keras.callbacks.Callback): # diff
"""Callback to operate on batch data from selected (to be wrapped) layers."""
def __init__(self, layers):
"""Wrap the calls of an iterable of model layers to access layer batch data."""
super().__init__()
self.data = {}
self.inner_calls = {}
self.outer_calls = {}
for layer in layers:
self.data[layer] = {
"inputs": tf_nan(layer.input.dtype),
"outputs": tf_nan(layer.output.dtype),
}
self.inner_calls[layer] = layer.call
def outer_call(inputs, layer=layer, layer_call=layer.call):
self.data[layer]["inputs"].assign(inputs)
outputs = layer_call(inputs)
self.data[layer]["outputs"].assign(outputs)
return outputs
self.outer_calls[layer] = outer_call
def on_train_batch_begin(self, _epoch, _logs=None):
"""Wrap layer calls during each batch."""
for layer, call in self.outer_calls.items():
layer.call = call
def on_train_batch_end(self, _epoch, _logs=None):
"""Restore original layer calls for ModelCheckpoint, model.save, ..."""
for layer, call in self.inner_calls.items():
layer.call = call
for layer, data in self.data.items():
print("Layer =", layer)
print("Inputs =", data["inputs"].numpy())
print("Outputs =", data["outputs"].numpy())
def tf_nan(dtype):
"""Create NaN variable of proper dtype and variable shape for assign()."""
return tf.Variable(float("nan"), dtype=dtype, shape=tf.TensorShape(None))
def main():
"""Run main."""
model = keras.Sequential([keras.layers.Dense(1, input_shape=(2,))])
callback = DataCallback(model.layers) # diff
model.compile(loss="mse", optimizer="adam")
model.fit(
x=tf.transpose(tf.range(7.0) + [[0.2], [0.4]]),
y=tf.transpose(tf.range(7.0) + 10 + [[0.5]]),
validation_data=(
tf.transpose(tf.range(11.0) + 30 + [[0.6], [0.7]]),
tf.transpose(tf.range(11.0) + 40 + [[0.9]]),
),
shuffle=False,
batch_size=3,
epochs=2,
verbose=0,
callbacks=[callback],
)
model.save("tmp.tf")
if __name__ == "__main__":
main()
When to use which and open to-dos
I think the snippets above each solution nicely summarize what each approach is capable of. Generally,
a custom training step will be ideal to access the model input, such as batched dataset generators, effects of shuffling, etc;
a custom layer call is ideal to access the in-betweens of the model; and
a custom metric is ideal to access the outputs of the model.
I am fairly certain (but have not tried) that one can combine all approaches to be able to access all batch quantities simultaneously. I have not tested anything but training mode - each method can have further pros and cons relating to their usefulness in testing or prediction mode. Finally, I assume, but have not tested either, that their should be only minor differences between tf.keras and keras. Having tested this code on TF2.8.rc1 and Keras 2.8.0, which has moved the tf.keras code back into the keras pip package, and not using any private APIs, I believe this assumption is justified.
It would be great if this approach could be extended to access model.inputs and model.outputs again. Currently, I am getting errors such as this one:
TypeError: You are passing KerasTensor(...), an intermediate Keras symbolic input/output, to a TF API that does not allow registering custom dispatchers, such as tf.cond, tf.function, gradient tapes, or tf.map_fn. Keras Functional model construction only supports TF API calls that do support dispatching, such as tf.math.add or tf.reshape. Other APIs cannot be called directly on symbolic Kerasinputs/outputs. You can work around this limitation by putting the operation in a custom Keras layer call and calling that layer on this symbolic input/output.
Previous answer
From TF 2.2 on, you can use custom training steps rather than callbacks to achieve what you want. Here's a demo that works with tensorflow==2.2.0rc1, using inheritance to improve the keras.Sequential model. Performance-wise, this is not ideal as predictions are made twice, once in self(x, training=True) and once in super().train_step(data). But you get the idea.
This works in eager mode and does not use private APIs, so it should be pretty stable. One caveat is that you have to use tf.keras (standalone keras does not support Model.train_step), but I feel standalone keras is becoming more and more deprecated anyway. (In fact, tf.keras migrates to keras in TF2.8.)
"""Demonstrate access to Keras batch tensors in a tf.keras custom training step."""
import numpy as np
from tensorflow import keras
from tensorflow.keras import backend as K
from tensorflow.python.keras.engine import data_adapter
in_shape = (2,)
out_shape = (1,)
batch_size = 3
n_samples = 7
class SequentialWithPrint(keras.Sequential):
def train_step(self, original_data):
# Basically copied one-to-one from https://git.io/JvDTv
data = data_adapter.expand_1d(original_data)
x, y_true, w = data_adapter.unpack_x_y_sample_weight(data)
y_pred = self(x, training=True)
# this is pretty much like on_train_batch_begin
K.print_tensor(w, "Sample weight (w) =")
K.print_tensor(x, "Batch input (x) =")
K.print_tensor(y_true, "Batch output (y_true) =")
K.print_tensor(y_pred, "Prediction (y_pred) =")
result = super().train_step(original_data)
# add anything here for on_train_batch_end-like behavior
return result
# Model
model = SequentialWithPrint([keras.layers.Dense(out_shape[0], input_shape=in_shape)])
model.compile(loss="mse", optimizer="adam")
# Example data
X = np.random.rand(n_samples, *in_shape)
Y = np.random.rand(n_samples, *out_shape)
model.fit(X, Y, batch_size=batch_size)
print("X: ", X)
print("Y: ", Y)
Finally, here is a simpler example without inheritance:
"""Demonstrate access to Keras batch tensors in a tf.keras custom training step."""
import tensorflow as tf
IN_SHAPE = (2,)
OUT_SHAPE = (1,)
BATCH_SIZE = 3
N_SAMPLES = 7
def make_print_data_and_train_step(keras_model):
"""Return a train_step function that prints data batches."""
original_train_step = keras_model.train_step
def print_data_and_train_step(data):
# Adapted from https://git.io/JvDTv, skipping data_adapter.expand_1d
x, y_true, w = tf.keras.utils.unpack_x_y_sample_weight(data)
y_pred = keras_model(x, training=True)
# this is pretty much like on_train_batch_begin
tf.keras.backend.print_tensor(w, "Sample weight (w) =")
tf.keras.backend.print_tensor(x, "Batch input (x) =")
tf.keras.backend.print_tensor(y_true, "Batch output (y_true) =")
tf.keras.backend.print_tensor(y_pred, "Prediction (y_pred) =")
result = original_train_step(data)
# add anything here for on_train_batch_end-like behavior
return result
return print_data_and_train_step
# Model
model = tf.keras.Sequential([tf.keras.layers.Dense(OUT_SHAPE[0], input_shape=IN_SHAPE)])
model.train_step = make_print_data_and_train_step(model)
model.compile(loss="mse", optimizer="adam")
# Example data
X = tf.random.normal((N_SAMPLES, *IN_SHAPE))
Y = tf.random.normal((N_SAMPLES, *OUT_SHAPE))
model.fit(X, Y, batch_size=BATCH_SIZE)
print("X: ", X)
print("Y: ", Y)
Update: This approach has stopped working. See my other answer a number of solutions compatible with TF2.8 (and hopefully beyond).
One problem with #Yu-Yang's solution is that it relies on model._function_kwargs, which is not guaranteed to work as it is not part of the API. In particular, in TF2 with eager execution, session kwargs seem to be either not accepted at all or run preemptively due to eager mode.
Therefore, here is my solution tested on tensorflow==2.1.0. The trick is to replace fetches by a Keras metric, in which the assignment operations from fetches are made during training.
This even enables a Keras-only solution if the batch size divides the number of samples; otherwise, another trick has to be applied when initializing TensorFlow variables with a None shape, similar to validate_shape=False in earlier solutions (compare https://github.com/tensorflow/tensorflow/issues/35667).
Importantly, tf.keras behaves differently from keras (sometimes just ignoring assignments, or seeing variables as Keras symbolic tensors), so this updated solution takes care of both implementations (Keras==2.3.1 and tensorflow==2.1.0).
"""Demonstrate access to Keras symbolic tensors in a (tf.)keras.Callback."""
import numpy as np
import tensorflow as tf
use_tf_keras = True
if use_tf_keras:
from tensorflow import keras
from tensorflow.keras import backend as K
tf.config.experimental_run_functions_eagerly(False)
compile_kwargs = {"run_eagerly": False, "experimental_run_tf_function": False}
else:
import keras
from keras import backend as K
compile_kwargs = {}
in_shape = (2,)
out_shape = (1,)
batch_size = 3
n_samples = 7
class CollectKerasSymbolicTensorsCallback(keras.callbacks.Callback):
"""Collect Keras symbolic tensors."""
def __init__(self):
"""Initialize intermediate variables for batches and lists."""
super().__init__()
# Collect batches here
self.inputs = []
self.targets = []
self.outputs = []
# # For a pure Keras solution, we need to know the shapes beforehand;
# # in particular, batch_size must divide n_samples:
# self.input = K.variable(np.empty((batch_size, *in_shape)))
# self.target = K.variable(np.empty((batch_size, *out_shape)))
# self.output = K.variable(np.empty((batch_size, *out_shape)))
# If the shape of these variables will change (e.g., last batch), initialize
# arbitrarily and specify `shape=tf.TensorShape(None)`:
self.input = tf.Variable(0.0, shape=tf.TensorShape(None))
self.target = tf.Variable(0.0, shape=tf.TensorShape(None))
self.output = tf.Variable(0.0, shape=tf.TensorShape(None))
def on_batch_end(self, batch, logs=None):
"""Evaluate the variables and save them into lists."""
self.inputs.append(K.eval(self.input))
self.targets.append(K.eval(self.target))
self.outputs.append(K.eval(self.output))
def on_train_end(self, logs=None):
"""Print all variables."""
print("Inputs: ", *self.inputs)
print("Targets: ", *self.targets)
print("Outputs: ", *self.outputs)
#tf.function
def assign_keras_symbolic_tensors_metric(_foo, _bar):
"""
Return the assignment operations as a metric to have them evaluated by Keras.
This replaces `fetches` from the TF1/non-eager-execution solution.
"""
# Collect assignments as list of (dest, src)
assignments = (
(callback.input, model.inputs[0]),
(callback.target, model._targets[0] if use_tf_keras else model.targets[0]),
(callback.output, model.outputs[0]),
)
for (dest, src) in assignments:
dest.assign(src)
return 0
callback = CollectKerasSymbolicTensorsCallback()
metrics = [assign_keras_symbolic_tensors_metric]
# Example model
model = keras.Sequential([keras.layers.Dense(out_shape[0], input_shape=in_shape)])
model.compile(loss="mse", optimizer="adam", metrics=metrics, **compile_kwargs)
# Example data
X = np.random.rand(n_samples, *in_shape)
Y = np.random.rand(n_samples, *out_shape)
model.fit(X, Y, batch_size=batch_size, callbacks=[callback])
print("X: ", X)
print("Y: ", Y)
Inspired by the way tf.keras.callbacks.TesnsorBoard saves v1 (graph) summaries.
No variable assignments and no redundant metrics.
For use with tensorflow>=2.0.0, graph (disable eager) mode during evaluating.
Extensive operations on the numpy predictions can be implemented by overriding SavePrediction._pred_callback.
import numpy as np
import tensorflow as tf
from tensorflow import keras
tf.compat.v1.disable_eager_execution()
in_shape = (2,)
out_shape = (1,)
batch_size = 2
n_samples = 32
class SavePrediction(keras.callbacks.Callback):
def __init__(self):
super().__init__()
self._get_pred = None
self.preds = []
def _pred_callback(self, preds):
self.preds.append(preds)
def set_model(self, model):
super().set_model(model)
if self._get_pred is None:
self._get_pred = self.model.outputs[0]
def on_test_begin(self, logs):
# pylint: disable=protected-access
self.model._make_test_function()
# pylint: enable=protected-access
if self._get_pred not in self.model.test_function.fetches:
self.model.test_function.fetches.append(self._get_pred)
self.model.test_function.fetch_callbacks[self._get_pred] = self._pred_callback
def on_test_end(self, logs):
if self._get_pred in self.model.test_function.fetches:
self.model.test_function.fetches.remove(self._get_pred)
if self._get_pred in self.model.test_function.fetch_callbacks:
self.model.test_function.fetch_callbacks.pop(self._get_pred)
print(self.preds)
model = keras.Sequential([
keras.layers.Dense(out_shape[0], input_shape=in_shape)
])
model.compile(loss="mse", optimizer="adam")
X = np.random.rand(n_samples, *in_shape)
Y = np.random.rand(n_samples, *out_shape)
model.evaluate(X, Y,
batch_size=batch_size,
callbacks=[SavePrediction()])
I am trying the distributed tensorflow, and my code is shown as follow. The problem is that the chief worker can run as expected. However, non-chief worker will blocked at :
sess = sv.prepare_or_wait_for_session(target, config=sess_config)
Could anybody help me solve this problem?
# Copyright 2016 Google Inc. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""A library to train Inception using multiple replicas with synchronous update.
Please see accompanying README.md for details and instructions.
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from datetime import datetime
import os.path
import time
import numpy as np
import tensorflow as tf
from inception.slim.datasets import dataset_factory
from inception.slim.nets import nets_factory
from inception.slim.preprocessing import preprocessing_factory
from inception import inception_model as inception
from inception.slim import slim
#from inception import image_processing
sslim = tf.contrib.slim
FLAGS = tf.app.flags.FLAGS
tf.app.flags.DEFINE_string(
'dataset_name', 'imagenet', 'The name of the dataset to load.')
tf.app.flags.DEFINE_string(
'dataset_split_name', 'train', 'The name of the train/test split.')
tf.app.flags.DEFINE_integer(
'train_image_size', None, 'Train image size')
tf.app.flags.DEFINE_string(
'dataset_dir', None, 'The directory where the dataset files are stored.')
tf.app.flags.DEFINE_string('job_name', '', 'One of "ps", "worker"')
tf.app.flags.DEFINE_string('ps_hosts', '',
"""Comma-separated list of hostname:port for the """
"""parameter server jobs. e.g. """
"""'machine1:2222,machine2:1111,machine2:2222'""")
tf.app.flags.DEFINE_string('worker_hosts', '',
"""Comma-separated list of hostname:port for the """
"""worker jobs. e.g. """
"""'machine1:2222,machine2:1111,machine2:2222'""")
tf.app.flags.DEFINE_float(
'weight_decay', 0.00004, 'The weight decay on the model weights.')
tf.app.flags.DEFINE_string('train_dir', '/tmp/imagenet_train',
"""Directory where to write event logs """
"""and checkpoint.""")
tf.app.flags.DEFINE_integer('max_steps', 100, 'Number of batches to run.')
tf.app.flags.DEFINE_string('subset', 'train', 'Either "train" or "validation".')
tf.app.flags.DEFINE_boolean('log_device_placement', False,
'Whether to log device placement.')
tf.app.flags.DEFINE_string(
'model_name', 'inception_v3', 'The name of the architecture to train.')
tf.app.flags.DEFINE_integer(
'batch_size', 32, 'The number of samples in each batch.')
tf.app.flags.DEFINE_string(
'preprocessing_name', None, 'The name of the preprocessing to use. If left '
'as `None`, then the model_name flag is used.')
# Task ID is used to select the chief and also to access the local_step for
# each replica to check staleness of the gradients in sync_replicas_optimizer.
tf.app.flags.DEFINE_integer(
'task_id', 0, 'Task ID of the worker/replica running the training.')
# More details can be found in the sync_replicas_optimizer class:
# tensorflow/python/training/sync_replicas_optimizer.py
tf.app.flags.DEFINE_integer('num_replicas_to_aggregate', -1,
"""Number of gradients to collect before """
"""updating the parameters.""")
tf.app.flags.DEFINE_integer('save_interval_secs', 10 * 60,
'Save interval seconds.')
tf.app.flags.DEFINE_integer('save_summaries_secs', 10 * 60,
'Save summaries interval seconds.')
# **IMPORTANT**
# Please note that this learning rate schedule is heavily dependent on the
# hardware architecture, batch size and any changes to the model architecture
# specification. Selecting a finely tuned learning rate schedule is an
# empirical process that requires some experimentation. Please see README.md
# more guidance and discussion.
#
# Learning rate decay factor selected from https://arxiv.org/abs/1604.00981
tf.app.flags.DEFINE_float('initial_learning_rate', 0.045,
'Initial learning rate.')
tf.app.flags.DEFINE_float('num_epochs_per_decay', 2.0,
'Epochs after which learning rate decays.')
tf.app.flags.DEFINE_float('learning_rate_decay_factor', 0.94,
'Learning rate decay factor.')
# Constants dictating the learning rate schedule.
RMSPROP_DECAY = 0.9 # Decay term for RMSProp.
RMSPROP_MOMENTUM = 0.9 # Momentum in RMSProp.
RMSPROP_EPSILON = 1.0 # Epsilon term for RMSProp.
def train(target, dataset, cluster_spec):
"""Train Inception on a dataset for a number of steps."""
# Number of workers and parameter servers are infered from the workers and ps
# hosts string.
num_workers = len(cluster_spec.as_dict()['worker'])
num_parameter_servers = len(cluster_spec.as_dict()['ps'])
# If no value is given, num_replicas_to_aggregate defaults to be the number of
# workers.
if FLAGS.num_replicas_to_aggregate == -1:
num_replicas_to_aggregate = num_workers
else:
num_replicas_to_aggregate = FLAGS.num_replicas_to_aggregate
# Both should be greater than 0 in a distributed training.
assert num_workers > 0 and num_parameter_servers > 0, (' num_workers and '
'num_parameter_servers'
' must be > 0.')
# Choose worker 0 as the chief. Note that any worker could be the chief
# but there should be only one chief.
is_chief = (FLAGS.task_id == 0)
# Ops are assigned to worker by default.
with tf.device('/job:worker/task:%d' % FLAGS.task_id):
# Variables and its related init/assign ops are assigned to ps.
with slim.scopes.arg_scope(
[slim.variables.variable, slim.variables.global_step],
device=slim.variables.VariableDeviceChooser(num_parameter_servers)):
# Create a variable to count the number of train() calls. This equals the
# number of updates applied to the variables.
global_step = slim.variables.global_step()
# Calculate the learning rate schedule.
num_batches_per_epoch = (dataset.num_examples_per_epoch() /
FLAGS.batch_size)
# Decay steps need to be divided by the number of replicas to aggregate.
decay_steps = int(num_batches_per_epoch * FLAGS.num_epochs_per_decay /
num_replicas_to_aggregate)
# Decay the learning rate exponentially based on the number of steps.
lr = tf.train.exponential_decay(FLAGS.initial_learning_rate,
global_step,
decay_steps,
FLAGS.learning_rate_decay_factor,
staircase=True)
# Add a summary to track the learning rate.
tf.summary.scalar('learning_rate', lr)
# Create an optimizer that performs gradient descent.
opt = tf.train.RMSPropOptimizer(lr,
RMSPROP_DECAY,
momentum=RMSPROP_MOMENTUM,
epsilon=RMSPROP_EPSILON)
dataset = dataset_factory.get_dataset(
FLAGS.dataset_name, FLAGS.dataset_split_name, FLAGS.dataset_dir)
network_fn = nets_factory.get_network_fn(
FLAGS.model_name,
num_classes=(dataset.num_classes),
weight_decay=FLAGS.weight_decay,
is_training=True)
preprocessing_name = FLAGS.preprocessing_name or FLAGS.model_name
image_preprocessing_fn = preprocessing_factory.get_preprocessing(
preprocessing_name,
is_training=True)
provider = sslim.dataset_data_provider.DatasetDataProvider(
dataset,
num_readers=4,
common_queue_capacity=20 * FLAGS.batch_size,
common_queue_min=10 * FLAGS.batch_size)
[image, label] = provider.get(['image', 'label'])
train_image_size = FLAGS.train_image_size or network_fn.default_image_size
image = image_preprocessing_fn(image, train_image_size, train_image_size)
images, labels = tf.train.batch(
[image, label],
batch_size=FLAGS.batch_size,
num_threads=4,
capacity=5 * FLAGS.batch_size)
# Number of classes in the Dataset label set plus 1.
# Label 0 is reserved for an (unused) background class.
num_classes = 1001
logits, end_points = network_fn(images)
batch_size=FLAGS.batch_size
# Add classification loss.
sparse_labels = tf.reshape(labels, [batch_size, 1])
indices = tf.reshape(tf.range(batch_size), [batch_size, 1])
#concated = tf.concat(1, [indices, sparse_labels])
sparse_labels = tf.cast(sparse_labels, tf.int32)
concated = tf.concat([indices, sparse_labels], 1)
dense_labels = tf.sparse_to_dense(concated,
[batch_size, 1001],
1.0, 0.0)
slim.losses.cross_entropy_loss(
logits, dense_labels, label_smoothing=0.01, weight=1.0)
# Gather all of the losses including regularization losses.
losses = tf.get_collection(slim.losses.LOSSES_COLLECTION)
losses += tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
total_loss = tf.add_n(losses, name='total_loss')
if is_chief:
# Compute the moving average of all individual losses and the
# total loss.
loss_averages = tf.train.ExponentialMovingAverage(0.9, name='avg')
loss_averages_op = loss_averages.apply(losses + [total_loss])
# Attach a scalar summmary to all individual losses and the total loss;
# do the same for the averaged version of the losses.
for l in losses + [total_loss]:
loss_name = l.op.name
# Name each loss as '(raw)' and name the moving average version of the
# loss as the original loss name.
tf.summary.scalar(loss_name + '_raw', l)
tf.summary.scalar(loss_name, loss_averages.average(l))
# Add dependency to compute loss_averages.
with tf.control_dependencies([loss_averages_op]):
total_loss = tf.identity(total_loss)
# Track the moving averages of all trainable variables.
# Note that we maintain a 'double-average' of the BatchNormalization
# global statistics.
# This is not needed when the number of replicas are small but important
# for synchronous distributed training with tens of workers/replicas.
exp_moving_averager = tf.train.ExponentialMovingAverage(
inception.MOVING_AVERAGE_DECAY, global_step)
variables_to_average = (
tf.trainable_variables() + tf.moving_average_variables())
# Add histograms for model variables.
for var in variables_to_average:
tf.summary.histogram(var.op.name, var)
# Create synchronous replica optimizer.
opt = tf.train.SyncReplicasOptimizer(
opt,
replicas_to_aggregate=num_replicas_to_aggregate,
total_num_replicas=num_workers,
variable_averages=exp_moving_averager,
variables_to_average=variables_to_average)
# Compute gradients with respect to the loss.
grads = opt.compute_gradients(total_loss)
# Add histograms for gradients.
for grad, var in grads:
if grad is not None:
tf.summary.histogram(var.op.name + '/gradients', grad)
apply_gradients_op = opt.apply_gradients(grads, global_step=global_step)
with tf.control_dependencies([apply_gradients_op]):
train_op = tf.identity(total_loss, name='train_op')
# Get chief queue_runners, init_tokens and clean_up_op, which is used to
# synchronize replicas.
# More details can be found in sync_replicas_optimizer.
chief_queue_runners = [opt.get_chief_queue_runner()]
init_tokens_op = opt.get_init_tokens_op()
# Build the summary operation based on the TF collection of Summaries.
summary_op = tf.summary.merge_all()
# Build an initialization operation to run below.
#init_op = tf.global_variables_initializer()
# We run the summaries in the same thread as the training operations by
# passing in None for summary_op to avoid a summary_thread being started.
# Running summaries and training operations in parallel could run out of
# GPU memory.
sv = tf.train.Supervisor(is_chief=is_chief,
logdir=FLAGS.train_dir,
init_op=init_op,
summary_op=None,
global_step=global_step,
#saver=saver,
saver=None,
save_model_secs=FLAGS.save_interval_secs)
tf.logging.info('%s Supervisor' % datetime.now())
sess_config = tf.ConfigProto(
allow_soft_placement=True,
log_device_placement=FLAGS.log_device_placement)
# Get a session.
sess = sv.prepare_or_wait_for_session(target, config=sess_config)
# Start the queue runners.
queue_runners = tf.get_collection(tf.GraphKeys.QUEUE_RUNNERS)
sv.start_queue_runners(sess, queue_runners)
tf.logging.info('Started %d queues for processing input data.',
len(queue_runners))
if is_chief:
sv.start_queue_runners(sess, chief_queue_runners)
sess.run(init_tokens_op)
# Train, checking for Nans. Concurrently run the summary operation at a
# specified interval. Note that the summary_op and train_op never run
# simultaneously in order to prevent running out of GPU memory.
#sess = sv.managed_session(target)
next_summary_time = time.time() + FLAGS.save_summaries_secs
while not sv.should_stop():
try:
start_time = time.time()
loss_value, step = sess.run([train_op, global_step])
assert not np.isnan(loss_value), 'Model diverged with loss = NaN'
if step > FLAGS.max_steps:
break
duration = time.time() - start_time
if step % 10 == 0:
examples_per_sec = FLAGS.batch_size / float(duration)
format_str = ('Worker %d: %s: step %d, loss = %.2f'
'(%.1f examples/sec; %.3f sec/batch)')
tf.logging.info(format_str %
(FLAGS.task_id, datetime.now(), step, loss_value,
examples_per_sec, duration))
# Determine if the summary_op should be run on the chief worker.
if is_chief and next_summary_time < time.time():
tf.logging.info('Running Summary operation on the chief.')
summary_str = sess.run(summary_op)
sv.summary_computed(sess, summary_str)
tf.logging.info('Finished running Summary operation.')
# Determine the next time for running the summary.
next_summary_time += FLAGS.save_summaries_secs
except:
if is_chief:
tf.logging.info('About to execute sync_clean_up_op!')
#sess.run(clean_up_op)
raise
# Stop the supervisor. This also waits for service threads to finish.
sv.stop()
Sync will create a local variable which will basically create the local step variable which is a local variable. But VariableDeviceChooser doesn't tell global from local so it is not functioning until we fix the device chooser. Thanks for reporting though.
Also concerned about this issue，Can you put your command line here?