Related
I am trying to do some transfer learning using this github DenseNet121 model (https://github.com/gaetandi/cheXpert.git). I'm running into issues resizing the classification layer from 14 to 2 outputs.
Relevant part of the github code is:
class DenseNet121(nn.Module):
"""Model modified.
The architecture of our model is the same as standard DenseNet121
except the classifier layer which has an additional sigmoid function.
"""
def __init__(self, out_size):
super(DenseNet121, self).__init__()
self.densenet121 = torchvision.models.densenet121(pretrained=True)
num_ftrs = self.densenet121.classifier.in_features
self.densenet121.classifier = nn.Sequential(
nn.Linear(num_ftrs, out_size),
nn.Sigmoid()
)
def forward(self, x):
x = self.densenet121(x)
return x
I load and init with:
# initialize and load the model
model = DenseNet121(nnClassCount).cuda()
model = torch.nn.DataParallel(model).cuda()
modeldict = torch.load("model_ones_3epoch_densenet.tar")
model.load_state_dict(modeldict['state_dict'])
It looks like DenseNet doesn't split layers up into children so model = nn.Sequential(*list(modelRes.children())[:-1]) won't work.
model.classifier = nn.Linear(1024, 2) seems to work on default DenseNets, but with the modified classifier (additional sigmoid function) here it ends up just adding an additional classifier layer without replacing the original.
I've tried
model.classifier = nn.Sequential(
nn.Linear(1024, dset_classes_number),
nn.Sigmoid()
)
But am having the same added instead of replaced classifier issue:
...
)
(classifier): Sequential(
(0): Linear(in_features=1024, out_features=14, bias=True)
(1): Sigmoid()
)
)
)
(classifier): Sequential(
(0): Linear(in_features=1024, out_features=2, bias=True)
(1): Sigmoid()
)
)
If you want to replace the classifier inside densenet121 that is a member of your model you need to assign
model.densenet121.classifier = nn.Sequential(...)
if i understand your problem, the following code will solve
import torchvision.models as models
import torch
from torch import nn
import numpy as np
np.random.seed(0)
torch.manual_seed(0)
densenet121 = models.densenet121(pretrained=True)
for param in densenet121.parameters():
param.requires_grad = False
densenet121.classifier = nn.Sequential(
nn.Linear(1024, 14),
nn.ReLU(),
nn.Dropout(0.4),
nn.Linear(14, 2),
)
densenet121.cuda()
I seem to get this error when I am using the callback function modelcheckpoint..
I read from a github issue that the solution would be make use of model.get_weight, but I am implicitly only storing that since i am only storing the one with best weight.
Keras only seem to save weights using h5, which make me question is there any other way to do store them using the eras API, if so how? If not, how do i store it?
Made an example to recreate the problem:
#!/usr/bin/python
import glob, os
import sys
from os import listdir
from os.path import isfile, join
import numpy as np
import warnings
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D
from keras.utils import np_utils
from keras import metrics
import keras
from keras import backend as K
from keras.models import Sequential
from keras.optimizers import SGD, Adam
from keras.layers.core import Dense, Activation, Lambda, Reshape,Flatten
from keras.layers import Conv1D,Conv2D,MaxPooling2D, MaxPooling1D, Reshape
#from keras.utils.visualize_util import plot
from keras.models import Model
from keras.layers import Input, Dense
from keras.layers.merge import Concatenate, Add
import h5py
import random
import tensorflow as tf
import math
from keras.callbacks import CSVLogger
from keras.callbacks import ModelCheckpoint
if len(sys.argv) < 5:
print "Missing Arguments!"
print "python keras_convolutional_feature_extraction.py <workspace> <totale_frames> <fbank-dim> <window-height> <batch_size>"
print "Example:"
print "python keras_convolutional_feature_extraction.py deltas 15 40 5 100"
sys.exit()
total_frames = int(sys.argv[2])
total_frames_with_deltas = total_frames*3
dim = int(sys.argv[3])
window_height = int(sys.argv[4])
inserted_batch_size = int(sys.argv[5])
stride = 1
splits = ((dim - window_height)+1)/stride
#input_train_data = "/media/carl/E2302E68302E443F/"+str(sys.argv[1])+"/fbank/org_train_total_frames_"+str(total_frames)+"_dim_"+str(dim)+"_winheig_"+str(window_height)+"_batch_"+str(inserted_batch_size)+"_fws_input"
#output_train_data ="/media/carl/E2302E68302E443F/"+str(sys.argv[1])+"/fbank/org_train_total_frames_"+str(total_frames)+"_dim_"+str(dim)+"_winheig_"+str(window_height)+"_batch_"+str(inserted_batch_size)+"_fws_output"
#input_test_data = "/media/carl/E2302E68302E443F/"+str(sys.argv[1])+"/fbank/org_test_total_frames_"+str(total_frames)+"_dim_"+str(dim)+"_winheig_"+str(window_height)+"_batch_"+str(1)+"_fws_input"
#output_test_data = "/media/carl/E2302E68302E443F/"+str(sys.argv[1])+"/fbank/org_test_total_frames_"+str(total_frames)+"_dim_"+str(dim)+"_winheig_"+str(window_height)+"_batch_"+str(1)+"_fws_output"
#train_files =[f for f in listdir(input_train_data) if isfile(join(input_train_data, f))]
#test_files =[f for f in listdir(input_test_data) if isfile(join(input_test_data, f))]
#print len(train_files)
np.random.seed(100)
print "hallo"
def train_generator():
while True:
# input = random.choice(train_files)
# h5f = h5py.File(input_train_data+'/'+input, 'r')
# train_input = h5f['train_input'][:]
# train_output = h5f['train_output'][:]
# h5f.close()
train_input = np.random.randint(100,size=((inserted_batch_size,splits*total_frames_with_deltas,window_height,3)))
train_list_list = []
train_input = train_input.reshape((inserted_batch_size,splits*total_frames_with_deltas,window_height,3))
train_input_list = np.split(train_input,splits*total_frames_with_deltas,axis=1)
for i in range(len(train_input_list)):
train_input_list[i] = train_input_list[i].reshape(inserted_batch_size,window_height,3)
#for i in range(len(train_input_list)):
# train_input_list[i] = train_input_list[i].reshape(inserted_batch_size,33,window_height,1,3)
train_output = np.random.randint(5, size = (1,total_frames,5))
middle = int(math.ceil(total_frames/2))
train_output = train_output[:,middle:middle+1,:].reshape((inserted_batch_size,1,5))
#print train_output.shape
#print len(train_input_list)
#print train_input_list[0].shape
yield (train_input_list, train_output)
print "hallo"
def test_generator():
while True:
# input = random.choice(test_files)
# h5f = h5py.File(input_test_data+'/'+input, 'r')
# test_input = h5f['test_input'][:]
# test_output = h5f['test_output'][:]
# h5f.close()
test_input = np.random.randint(100,size=((inserted_batch_size,splits*total_frames_with_deltas,window_height,3)))
test_input = test_input.reshape((inserted_batch_size,splits*total_frames_with_deltas,window_height,3))
test_input_list = np.split(test_input,splits*total_frames_with_deltas,axis=1)
#test_input_list = np.split(test_input,45,axis=3)
for i in range(len(test_input_list)):
test_input_list[i] = test_input_list[i].reshape(inserted_batch_size,window_height,3)
#for i in range(len(test_input_list)):
# test_input_list[i] = test_input_list[i].reshape(inserted_batch_size,33,window_height,1,3)
test_output = np.random.randint(5, size = (1,total_frames,5))
middle = int(math.ceil(total_frames/2))
test_output = test_output[:,middle:middle+1,:].reshape((inserted_batch_size,1,5))
yield (test_input_list, test_output)
print "hallo"
def fws():
#print "Inside"
# Params:
# batch , lr, decay , momentum, epochs
#
#Input shape: (batch_size,40,45,3)
#output shape: (1,15,50)
# number of unit in conv_feature_map = splitd
next(train_generator())
model_output = []
list_of_input = [Input(shape=(8,3)) for i in range(splits*total_frames_with_deltas)]
output = []
#Conv
skip = total_frames_with_deltas
for steps in range(total_frames_with_deltas):
conv = Conv1D(filters = 100, kernel_size = 8)
column = 0
for _ in range(splits):
#print "column " + str(column) + "steps: " + str(steps)
output.append(conv(list_of_input[(column*skip)+steps]))
column = column + 1
#print len(output)
#print splits*total_frames_with_deltas
conv = []
for section in range(splits):
column = 0
skip = splits
temp = []
for _ in range(total_frames_with_deltas):
temp.append(output[((column*skip)+section)])
column = column + 1
conv.append(Add()(temp))
#print len(conv)
output_conc = Concatenate()(conv)
#print output_conc.get_shape
output_conv = Reshape((splits, -1))(output_conc)
#print output_conv.get_shape
#Pool
pooled = MaxPooling1D(pool_size = 6, strides = 2)(output_conv)
reshape = Reshape((1,-1))(pooled)
#Fc
dense1 = Dense(units = 1024, activation = 'relu', name = "dense_1")(reshape)
#dense2 = Dense(units = 1024, activation = 'relu', name = "dense_2")(dense1)
dense3 = Dense(units = 1024, activation = 'relu', name = "dense_3")(dense1)
final = Dense(units = 5, activation = 'relu', name = "final")(dense3)
model = Model(inputs = list_of_input , outputs = final)
sgd = SGD(lr=0.1, decay=1e-1, momentum=0.9, nesterov=True)
model.compile(loss="categorical_crossentropy", optimizer=sgd , metrics = ['accuracy'])
print "compiled"
model_yaml = model.to_yaml()
with open("model.yaml", "w") as yaml_file:
yaml_file.write(model_yaml)
print "Model saved!"
log= CSVLogger('/home/carl/kaldi-trunk/dnn/experimental/yesno_cnn_50_training_total_frames_'+str(total_frames)+"_dim_"+str(dim)+"_window_height_"+str(window_height)+".csv")
filepath='yesno_cnn_50_training_total_frames_'+str(total_frames)+"_dim_"+str(dim)+"_window_height_"+str(window_height)+"weights-improvement-{epoch:02d}-{val_acc:.2f}.hdf5"
checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_weights_only=True, mode='max')
print "log"
#plot_model(model, to_file='model.png')
print "Fit"
hist_current = model.fit_generator(train_generator(),
steps_per_epoch=444,#len(train_files),
epochs = 10000,
verbose = 1,
validation_data = test_generator(),
validation_steps=44,#len(test_files),
pickle_safe = True,
workers = 4,
callbacks = [log,checkpoint])
fws()
Execute the script by: python name_of_script.py yens 50 40 8 1
which give me a full traceback:
full traceback
Error:
carl#ca-ThinkPad-T420s:~/Dropbox$ python mini.py yesno 50 40 8 1
Using TensorFlow backend.
Couldn't import dot_parser, loading of dot files will not be possible.
hallo
hallo
hallo
compiled
Model saved!
log
Fit
/usr/local/lib/python2.7/dist-packages/keras/backend/tensorflow_backend.py:2252: UserWarning: Expected no kwargs, you passed 1
kwargs passed to function are ignored with Tensorflow backend
warnings.warn('\n'.join(msg))
Epoch 1/10000
2017-05-26 13:01:45.851125: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.1 instructions, but these are available on your machine and could speed up CPU computations.
2017-05-26 13:01:45.851345: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.2 instructions, but these are available on your machine and could speed up CPU computations.
2017-05-26 13:01:45.851392: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX instructions, but these are available on your machine and could speed up CPU computations.
443/444 [============================>.] - ETA: 4s - loss: 100.1266 - acc: 0.3138Epoch 00000: saving model to yesno_cnn_50_training_total_frames_50_dim_40_window_height_8weights-improvement-00-0.48.hdf5
Traceback (most recent call last):
File "mini.py", line 205, in <module>
File "mini.py", line 203, in fws
File "/usr/local/lib/python2.7/dist-packages/keras/legacy/interfaces.py", line 88, in wrapper
return func(*args, **kwargs)
File "/usr/local/lib/python2.7/dist-packages/keras/engine/training.py", line 1933, in fit_generator
callbacks.on_epoch_end(epoch, epoch_logs)
File "/usr/local/lib/python2.7/dist-packages/keras/callbacks.py", line 77, in on_epoch_end
callback.on_epoch_end(epoch, logs)
File "/usr/local/lib/python2.7/dist-packages/keras/callbacks.py", line 411, in on_epoch_end
self.model.save_weights(filepath, overwrite=True)
File "/usr/local/lib/python2.7/dist-packages/keras/engine/topology.py", line 2503, in save_weights
save_weights_to_hdf5_group(f, self.layers)
File "/usr/local/lib/python2.7/dist-packages/keras/engine/topology.py", line 2746, in save_weights_to_hdf5_group
f.attrs['layer_names'] = [layer.name.encode('utf8') for layer in layers]
File "h5py/_objects.pyx", line 54, in h5py._objects.with_phil.wrapper (/tmp/pip-4rPeHA-build/h5py/_objects.c:2684)
File "h5py/_objects.pyx", line 55, in h5py._objects.with_phil.wrapper (/tmp/pip-4rPeHA-build/h5py/_objects.c:2642)
File "/usr/local/lib/python2.7/dist-packages/h5py/_hl/attrs.py", line 93, in __setitem__
self.create(name, data=value, dtype=base.guess_dtype(value))
File "/usr/local/lib/python2.7/dist-packages/h5py/_hl/attrs.py", line 183, in create
attr = h5a.create(self._id, self._e(tempname), htype, space)
File "h5py/_objects.pyx", line 54, in h5py._objects.with_phil.wrapper (/tmp/pip-4rPeHA-build/h5py/_objects.c:2684)
File "h5py/_objects.pyx", line 55, in h5py._objects.with_phil.wrapper (/tmp/pip-4rPeHA-build/h5py/_objects.c:2642)
File "h5py/h5a.pyx", line 47, in h5py.h5a.create (/tmp/pip-4rPeHA-build/h5py/h5a.c:1904)
RuntimeError: Unable to create attribute (Object header message is too large)
If you look at the amount of data Keras is trying to save under layer_names attribute (inside the output HDF5 file being create), you will find that it takes more than 64kB.
np.asarray([layer.name.encode('utf8') for layer in model.layers]).nbytes
>> 77100
I quote from https://support.hdfgroup.org/HDF5/faq/limits.html:
Is there an object header limit and how does that affect HDF5 ?
There is a limit (in HDF5-1.8) of the object header, which is 64 KB.
The datatype for a dataset is stored in the object header, so there is
therefore a limit on the size of the datatype that you can have. (See
HDFFV-1089)
The code above was (almost entirely) copied from the traceback:
File "/usr/local/lib/python2.7/dist-packages/keras/engine/topology.py", line 2746, in save_weights_to_hdf5_group
f.attrs['layer_names'] = [layer.name.encode('utf8') for layer in layers]
I am using numpy asarray method to get the figure fast but h5py gets similar figure (I guess), see https://github.com/h5py/h5py/blob/master/h5py/_hl/attrs.py#L102 if you want to find exact figure.
Anyway, either you will need to implement your own methods for saving/loading of the weights (or use existing workarounds), or you need to give a really short name to ALL the layers inside your model :), something like this:
list_of_input = [Input(shape=(8,3), name=('i%x' % i)) for i in range(splits*total_frames_with_deltas)]
conv = Conv1D(filters = 100, kernel_size = 8, name='cv%x' % steps)
conv.append(Add(name='add%x' % section)(temp))
output_conc = Concatenate(name='ct')(conv)
output_conv = Reshape((splits, -1), name='rs1')(output_conc)
pooled = MaxPooling1D(pool_size = 6, strides = 2, name='pl')(output_conv)
reshape = Reshape((1,-1), name='rs2')(pooled)
dense1 = Dense(units = 1024, activation = 'relu', name = "d1")(reshape)
dense2 = Dense(units
= 1024, activation = 'relu', name = "d2")(dense1)
dense3 = Dense(units = 1024, activation = 'relu', name = "d3")(dense1)
final = Dense(units = 5, activation = 'relu', name = "fl")(dense3)
You mustn't forget to name all the layers because the (numpy) string array into which the layer names are converted is using the size of the longest string for each individual string in it when it is saved!
After renaming the layers as proposed above (which takes almost 26kB) the model is saved successfully. Hope this elaborate answer helps someone.
Update: I have just made a PR to Keras which should fix the issue without implementing any custom loading/saving methods, see 7508
A simple solution, albeit possibly not the most elegant, could be to run a while loop with epochs = 1.
Get the weights at the end of every epoch together with the accuracy and the loss
Save the weights to file 1 with model.get_weight
if accuracy is greater than at the previous epoch (i.e. loop), store the weights to a different file (file 2)
Run the loop again loading the weights from file 1
Break the loops setting a manual early stopping so that it breaks if the loss does not improve for a certain number of loops
You can use get_weights() together with numpy.save.
It's not the best solution, because it will save several files, but it actually works.
The problem is that you won't have the "optimizer" saved with the current states. But you can perhaps work around that by using smaller learning rates after loading.
Custom callback using numpy.save:
def myCallback(epoch,logs):
global storedLoss
#do your comparisons here using the "logs" var.
print(logs)
if (logs['loss'] < storedLoss):
storedLoss = logs['loss']
for i in range(len(model.layers)):
WandB = model.layers[i].get_weights()
if len (WandB) > 0: #necessary because some layers have no weights
np.save("W" + "-" + str(i), WandB[0],False)
np.save("B" + "-" + str(i), WandB[1],False)
#remember that get and set weights use a list: [weights,biases]
#it may happen (not sure) that there is no bias, and thus you may have to check it (len(WandB)==1).
The logs var brings a dictionary with named metrics, such as "loss", and "accuracy", if you used it.
You can store the losses within the callback in a global var, and compare if each loss is better or worse than the last.
When fitting, use the lambda callback:
from keras.callbacks import LambdaCallback
model.fit(...,callbacks=[LambdaCallback(on_epoch_end=myCallback)])
In the example above, I used the LambdaCallback, which has more possibilities than just on_epoch_end.
For loading, do a similar loop:
#you have to create the model first and then set the layers
def loadModel(model):
for i in range(len(model.layers)):
WandBForCheck = model.layers[i].get_weights()
if len (WandBForCheck) > 0: #necessary because some layers have no weights
W = np.load(Wfile + str(i))
B = np.load(Bfile + str(i))
model.layers[i].set_weights([W,B])
See follow-up at https://github.com/fchollet/keras/issues/6766 and https://github.com/farizrahman4u/keras-contrib/pull/90.
I saw the YAML and the root cause is probably that you have so many Inputs. A few Inputs with many dimensions is preferred to many Inputs, especially if you can use scanning and batch operations to do everything efficiently.
Now, ignoring that entirely, here is how you can save and load your model if it has too much stuff to save as JSON efficiently:
You can pass save_weights_only=True. That won't save optimizer weights, so isn't a great solution.
Just put together a PR for saving model weights and optimizer weights but not configuration. When you want to load, first instantiate and compile the model as you did when you were going to train it, then use load_all_weights to load the model and optimizer weights into that model. I'll try to merge it soon so you can use it from the master branch.
You could use it something like this:
from keras.callbacks import LambdaCallback
from keras_contrib.utils.save_load_utils import save_all_weights, load_all_weights
# do some stuff to create and compile model
# use `save_all_weights` as a callback to checkpoint your model and optimizer weights
model.fit(..., callbacks=[LambdaCallback(on_epoch_end=lambda epoch, logs: save_all_weights(model, "checkpoint-{:05d}.h5".format(epoch))])
# use `load_all_weights` to load model and optimizer weights into an existing model
# if not compiled (no `model.optimizer`), this will just load model weights
load_all_weights(model, 'checkpoint-1337.h5')
So I don't endorse the model, but if you want to get it to save and load anyways this should probably work for you.
As a side note, if you want to save weights in a different format, something like this would work.
pickle.dump([K.get_value(w) for w in model.weights], open( "save.p", "wb" ) )
Cheers
Your model architecture must be too large to be saved.
USE get_weights AND set_weights TO SAVE AND LOAD MODEL, RESPECTIVELY.
Do not use callback model checkpoint. just once the training ends, save its weights with pickle.
Have a look at this link: Unable to save DataFrame to HDF5 ("object header message is too large")
I am trying to build a recurrent network which classifies sequences (multidimensional data streams). I must be missing something, since while running my code:
from keras.models import Sequential
from keras.layers import LSTM, Dropout, Activation
import numpy as np
ils = 10 # input layer size
ilt = 11 # input layer time steps
hls = 12 # hidden layer size
nhl = 2 # number of hidden layers
ols = 1 # output layer size
p = 0.2 # dropout probability
f_a = 'relu' # activation function
opt = 'rmsprop' # optimizing function
#
# Building the model
#
model = Sequential()
# The input layer
model.add(LSTM(hls, input_shape=(ilt, ils), return_sequences=True))
model.add(Activation(f_a))
model.add(Dropout(p))
# Hidden layers
for i in range(nhl - 1):
model.add(LSTM(hls, return_sequences=True))
model.add(Activation(f_a))
model.add(Dropout(p))
# Output layer
model.add(LSTM(ols, return_sequences=False))
model.add(Activation('softmax'))
model.compile(optimizer=opt, loss='binary_crossentropy')
#
# Making test data and fitting the model
#
m_train, n_class = 1000, 2
data = np.array(np.random.random((m_train, ilt, ils)))
labels = np.random.randint(n_class, size=(m_train, 1))
model.fit(data, labels, nb_epoch=10, batch_size=32)
I get output (truncated):
Using Theano backend.
line 611, in __call__
node = self.make_node(*inputs, **kwargs)
File "/home/koala/.local/lib/python2.7/site-packages/theano/scan_module/scan_op.py", line 430, in make_node
new_inputs.append(format(outer_seq, as_var=inner_seq))
File "/home/koala/.local/lib/python2.7/site-packages/theano/scan_module/scan_op.py", line 422, in format
rval = tmp.filter_variable(rval)
File "/home/koala/.local/lib/python2.7/site-packages/theano/tensor/type.py", line 233, in filter_variable
self=self))
TypeError: Cannot convert Type TensorType(float32, 3D) (of Variable Subtensor{:int64:}.0) into Type TensorType(float32, (False, False, True)). You can try to manually convert Subtensor{:int64:}.0 into a TensorType(float32, (False, False, True)).
Is this a problem with the data format at all.
For me the problem was fixed when I went and tried it on my real dataset. The difference being that in the real dataset I have more than 1 label. So an example of dataset on which this code works is:
(...)
ols = 2 # Output layer size
(...)
m_train, n_class = 1000, ols
data = np.array(np.random.random((m_train, ilt, ils)))
labels = np.random.randint(n_class, size=(m_train, 1))
# Make labels onehot
onehot_labels = np.zeros(shape=(labels.shape[0], ols))
onehot_labels[np.arange(labels.shape[0]), labels.astype(np.int)] = 1
Using the MNIST tutorial of Tensorflow, I try to make a convolutional network for face recognition with the "Database of Faces".
The images size are 112x92, I use 3 more convolutional layer to reduce it to 6 x 5 as adviced here
I'm very new at convolutional network and most of my layer declaration is made by analogy to the Tensorflow MNIST tutorial, it may be a bit clumsy, so feel free to advice me on this.
x_image = tf.reshape(x, [-1, 112, 92, 1])
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
h_pool1 = max_pool_2x2(h_conv1)
W_conv2 = weight_variable([5, 5, 32, 64])
b_conv2 = bias_variable([64])
h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
h_pool2 = max_pool_2x2(h_conv2)
W_conv3 = weight_variable([5, 5, 64, 128])
b_conv3 = bias_variable([128])
h_conv3 = tf.nn.relu(conv2d(h_pool2, W_conv3) + b_conv3)
h_pool3 = max_pool_2x2(h_conv3)
W_conv4 = weight_variable([5, 5, 128, 256])
b_conv4 = bias_variable([256])
h_conv4 = tf.nn.relu(conv2d(h_pool3, W_conv4) + b_conv4)
h_pool4 = max_pool_2x2(h_conv4)
W_conv5 = weight_variable([5, 5, 256, 512])
b_conv5 = bias_variable([512])
h_conv5 = tf.nn.relu(conv2d(h_pool4, W_conv5) + b_conv5)
h_pool5 = max_pool_2x2(h_conv5)
W_fc1 = weight_variable([6 * 5 * 512, 1024])
b_fc1 = bias_variable([1024])
h_pool5_flat = tf.reshape(h_pool5, [-1, 6 * 5 * 512])
h_fc1 = tf.nn.relu(tf.matmul(h_pool5_flat, W_fc1) + b_fc1)
keep_prob = tf.placeholder("float")
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
print orlfaces.train.num_classes # 40
W_fc2 = weight_variable([1024, orlfaces.train.num_classes])
b_fc2 = bias_variable([orlfaces.train.num_classes])
y_conv = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)
My problem appear when the session run the "correct_prediction" op which is
tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))
At least I think given the error message:
W tensorflow/core/common_runtime/executor.cc:1027] 0x19369d0 Compute status: Invalid argument: Incompatible shapes: [8] vs. [20]
[[Node: Equal = Equal[T=DT_INT64, _device="/job:localhost/replica:0/task:0/cpu:0"](ArgMax, ArgMax_1)]]
Traceback (most recent call last):
File "./convolutional.py", line 133, in <module>
train_accuracy = accuracy.eval(feed_dict = {x: batch[0], y_: batch[1], keep_prob: 1.0})
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.py", line 405, in eval
return _eval_using_default_session(self, feed_dict, self.graph, session)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.py", line 2728, in _eval_using_default_session
return session.run(tensors, feed_dict)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.py", line 345, in run
results = self._do_run(target_list, unique_fetch_targets, feed_dict_string)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.py", line 419, in _do_run
e.code)
tensorflow.python.framework.errors.InvalidArgumentError: Incompatible shapes: [8] vs. [20]
[[Node: Equal = Equal[T=DT_INT64, _device="/job:localhost/replica:0/task:0/cpu:0"](ArgMax, ArgMax_1)]]
Caused by op u'Equal', defined at:
File "./convolutional.py", line 125, in <module>
correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/gen_math_ops.py", line 328, in equal
return _op_def_lib.apply_op("Equal", x=x, y=y, name=name)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/op_def_library.py", line 633, in apply_op
op_def=op_def)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.py", line 1710, in create_op
original_op=self._default_original_op, op_def=op_def)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.py", line 988, in __init__
self._traceback = _extract_stack()
It looks like the y_conv output a matrix of shape 8 x batch_size instead of number_of_class x batch_size
If I change the batch size from 20 to 10, the error message stay the same but instead [8] vs. [20] I get [4] vs. [10]. So from that I conclude that the problem may come from the y_conv declaration (last line of the code above).
The loss function, optimizer, training, etc declarations is the same as in the MNIST tutorial:
cross_entropy = -tf.reduce_sum(y_ * tf.log(y_conv))
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
sess.run((tf.initialize_all_variables()))
for i in xrange(1000):
batch = orlfaces.train.next_batch(20)
if i % 100 == 0:
train_accuracy = accuracy.eval(feed_dict = {x: batch[0], y_: batch[1], keep_prob: 1.0})
print "Step %d, training accuracy %g" % (i, train_accuracy)
train_step.run(feed_dict = {x: batch[0], y_: batch[1], keep_prob: 0.5})
print "Test accuracy %g" % accuracy.eval(feed_dict = {x: orlfaces.test.images, y_: orlfaces.test.labels, keep_prob: 1.0})
Thanks for reading, have a good day
Well, after a lot debugging, I found that my issue was due to a bad instantiation of the labels. Instead of creating arrays full of zeros and replace one value by one, I created them with random value! Stupid mistake. In case someone wondering what I did wrong there and how I fix it here is the change I made.
Anyway during all the debugging I made, to find this mistake, I found some useful information to debug this kind of problem:
For the cross entropy declaration, the tensorflow's MNIST tutorial use a formula that can lead to NaN value
This formula is
cross_entropy = -tf.reduce_sum(y_ * tf.log(y_conv))
Instead of this, I found two ways to declare it in a safer fashion:
cross_entropy = -tf.reduce_sum(y_ * tf.log(tf.clip_by_value(y_conv, 1e-10, 1.0)))
or also:
cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logit, y_))
As mrry says. printing the shape of the tensors can help to detect shape anomaly.
To get the shape of a tensor just call his get_shape() method like this:
print "W shape:", W.get_shape()
user1111929 in this question use a debug print that help me assert where the problem come from.
I am trying to implement another pooling function for neural network with Theano, expect of already existing maxpool, for example average pool.
Using to this source, where average pooling is already implemented, my code looks like:
Random initialization just to test:
invals = numpy.random.RandomState(1).rand(3,2,5,5)
Definition of Theano scalars and functions:
pdim = T.scalar('pool dim', dtype='float32')
pool_inp = T.tensor4('pool input', dtype='float32')
pool_sum = TSN.images2neibs(pool_inp, (pdim, pdim))
pool_out = pool_sum.mean(axis=-1)
pool_fun = theano.function([pool_inp, pdim], pool_out, name = 'pool_fun', allow_input_downcast=True)
TSN is theano.sandbox.neighbours
And the call of the function:
pool_dim = 2
temp = pool_fun(invals, pool_dim)
temp.shape = (invals.shape[0], invals.shape[1], invals.shape[2]/pool_dim,
invals.shape[3]/pool_dim)
print ('invals[1,0,:,:]=\n', invals[1,0,:,:])
print ('output[1,0,:,:]=\n',temp[1,0,:,:])
And I am getting an error:
TypeError: neib_shape[0]=2, neib_step[0]=2 and ten4.shape[2]=5 not consistent
Apply node that caused the error: Images2Neibs{valid}(pool input, MakeVector.0, MakeVector.0)
Inputs shapes: [(3, 2, 5, 5), (2,), (2,)]
Inputs strides: [(200, 100, 20, 4), (4,), (4,)]
Inputs types: [TensorType(float32, 4D), TensorType(float32, vector), TensorType(float32, vector)]
Use the Theano flag 'exception_verbosity=high' for a debugprint of this apply node.
I don't really understand this error. Would be glad to have any suggestions how to correct this error or example of other pooling techniques, programmed in Theano.
Thanks!
Edit: with the ignoring the border, it works perfectly
pool_sum = TSN.images2neibs(pool_inp, (pdim, pdim), mode='ignore_borders')
invals[1,0,:,:]=
[[ 0.01936696 0.67883553 0.21162812 0.26554666 0.49157316]
[ 0.05336255 0.57411761 0.14672857 0.58930554 0.69975836]
[ 0.10233443 0.41405599 0.69440016 0.41417927 0.04995346]
[ 0.53589641 0.66379465 0.51488911 0.94459476 0.58655504]
[ 0.90340192 0.1374747 0.13927635 0.80739129 0.39767684]]
output[1,0,:,:]=
[[ 0.33142066 0.30330223]
[ 0.42902038 0.64201581]]
invals has shape (5, 5) in the last two dimensions, however you want to pool over (2, 2) subsets. This only works if you ignore the border (i.e. the last column and the last row of invals).