Is it possible to remove a dimension using Reshape or any other function.
I have the following network.
import keras
from keras.layers.merge import Concatenate
from keras.models import Model
from keras.layers import Input, Dense
from keras.layers import Dropout
from keras.layers.core import Dense, Activation, Lambda, Reshape,Flatten
from keras.layers import Conv2D, MaxPooling2D, Reshape, ZeroPadding2D
import numpy as np
#Number_of_splits = ((input_width-win_dim)+1)/stride_dim
splits = ((40-5)+1)/1
print splits
train_data_1 = np.random.randint(100,size=(100,splits,45,5,3))
test_data_1 = np.random.randint(100,size=(10,splits,45,5,3))
labels_train_data =np.random.randint(145,size=(100,15))
labels_test_data =np.random.randint(145,size=(10,15))
list_of_input = [Input(shape = (45,5,3)) for i in range(splits)]
list_of_conv_output = []
list_of_max_out = []
for i in range(splits):
list_of_conv_output.append(Conv2D(filters = 145 , kernel_size = (15,3))(list_of_input[i])) #output dim: 36x(31,3,145)
list_of_max_out.append((MaxPooling2D(pool_size=(2,2))(list_of_conv_output[i]))) #output dim: 36x(15,1,145)
merge = keras.layers.concatenate(list_of_max_out) #Output dim: (15,1,5220)
#reshape = Reshape((merge.shape[0],merge.shape[3]))(merge) # expected output dim: (15,145)
dense1 = Dense(units = 1000, activation = 'relu', name = "dense_1")(merge)
dense2 = Dense(units = 1000, activation = 'relu', name = "dense_2")(dense1)
dense3 = Dense(units = 145 , activation = 'softmax', name = "dense_3")(dense2)
model = Model(inputs = list_of_input , outputs = dense3)
model.compile(loss="sparse_categorical_crossentropy", optimizer="adam")
print model.summary()
raw_input("SDasd")
hist_current = model.fit(x = [train_input[i] for i in range(100)],
y = labels_train_data,
shuffle=False,
validation_data=([test_input[i] for i in range(10)], labels_test_data),
validation_split=0.1,
epochs=150000,
batch_size = 15,
verbose=1)
The maxpooling layer creates an output with dimension (15,1,36) which i would like to remove the middle axis, so the output dimension end up being (15,36)..
If possible would I like to avoid specifying the outer dimension, or as i've tried use the prior layer dimension to reshape it.
#reshape = Reshape((merge.shape[0],merge.shape[3]))(merge) # expected output dim: (15,145)
I need my output dimension for the entire network to be (15,145), in which the middle dimension is causing some problems.
How do i remove the middle dimension?
I wanted to remove all dimensions that are equal to 1, but not specify a specific size with Reshape so that my code does not break if I change the input size or number of kernels in a convolution. This works with the functional keras API on a tensorflow backend.
from keras.layers.core import Reshape
old_layer = Conv2D(#actualArguments) (older_layer)
#old_layer yields, e.g., a (None, 15,1,36) size tensor, where None is the batch size
newdim = tuple([x for x in old_layer.shape.as_list() if x != 1 and x is not None])
#newdim is now (15, 36). Reshape does not take batch size as an input dimension.
reshape_layer = Reshape(newdim) (old_layer)
reshape = Reshape((15,145))(merge) # expected output dim: (15,145)
Related
I'm pretty new to deep learning, but I want to add Tensorboard to the following code to track loss, accuracy, average precision and so on.
Sample code from the TorchVision -2.3 Object Detection Finetuning Tutorial
http://pytorch.org/tutorials/intermediate/torchvision_tutorial.html
import os
import numpy as np
import torch
from PIL import Image
import sys
import torchvision
from torchvision.models.detection.faster_rcnn import FastRCNNPredictor
from torchvision.models.detection.mask_rcnn import MaskRCNNPredictor
from engine import train_one_epoch, evaluate
import utils
import transforms as T
#from torch.utils.tensorboard import SummaryWriter
#writer = SummaryWriter()
class PennFudanDataset(object):
def __init__(self, root, transforms):
self.root = root
self.transforms = transforms
# load all image files, sorting them to
# ensure that they are aligned
self.imgs = list(sorted(os.listdir(os.path.join(root, "PNGImages"))))
self.masks = list(sorted(os.listdir(os.path.join(root, "PedMasks"))))
def __getitem__(self, idx):
# load images and masks
img_path = os.path.join(self.root, "PNGImages", self.imgs[idx])
mask_path = os.path.join(self.root, "PedMasks", self.masks[idx])
img = Image.open(img_path).convert("RGB")
# note that we haven't converted the mask to RGB,
# because each color corresponds to a different instance
# with 0 being background
mask = Image.open(mask_path)
mask = np.array(mask)
# instances are encoded as different colors
obj_ids = np.unique(mask)
# first id is the background, so remove it
obj_ids = obj_ids[1:]
# split the color-encoded mask into a set
# of binary masks
masks = mask == obj_ids[:, None, None]
# get bounding box coordinates for each mask
num_objs = len(obj_ids)
boxes = []
for i in range(num_objs):
pos = np.where(masks[i])
xmin = np.min(pos[1])
xmax = np.max(pos[1])
ymin = np.min(pos[0])
ymax = np.max(pos[0])
boxes.append([xmin, ymin, xmax, ymax])
boxes = torch.as_tensor(boxes, dtype=torch.float32)
# there is only one class
labels = torch.ones((num_objs,), dtype=torch.int64)
masks = torch.as_tensor(masks, dtype=torch.uint8)
image_id = torch.tensor([idx])
area = (boxes[:, 3] - boxes[:, 1]) * (boxes[:, 2] - boxes[:, 0])
# suppose all instances are not crowd
iscrowd = torch.zeros((num_objs,), dtype=torch.int64)
target = {}
target["boxes"] = boxes
target["labels"] = labels
target["masks"] = masks
target["image_id"] = image_id
target["area"] = area
target["iscrowd"] = iscrowd
if self.transforms is not None:
img, target = self.transforms(img, target)
return img, target
def __len__(self):
return len(self.imgs)
def get_model_instance_segmentation(num_classes):
# load an instance segmentation model pre-trained pre-trained on COCO
model = torchvision.models.detection.maskrcnn_resnet50_fpn(pretrained=True)
# get number of input features for the classifier
in_features = model.roi_heads.box_predictor.cls_score.in_features
# replace the pre-trained head with a new one
model.roi_heads.box_predictor = FastRCNNPredictor(in_features, num_classes)
# now get the number of input features for the mask classifier
in_features_mask = model.roi_heads.mask_predictor.conv5_mask.in_channels
hidden_layer = 256
# and replace the mask predictor with a new one
model.roi_heads.mask_predictor = MaskRCNNPredictor(in_features_mask,
hidden_layer,
num_classes)
return model
def get_transform(train):
transforms = []
transforms.append(T.ToTensor())
if train:
transforms.append(T.RandomHorizontalFlip(0.5))
return T.Compose(transforms)
def main():
# train on the GPU or on the CPU, if a GPU is not available
device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
# our dataset has two classes only - background and person
num_classes = 2
# use our dataset and defined transformations
dataset = PennFudanDataset('PennFudanPed', get_transform(train=True))
dataset_test = PennFudanDataset('PennFudanPed', get_transform(train=False))
# split the dataset in train and test set
indices = torch.randperm(len(dataset)).tolist()
dataset = torch.utils.data.Subset(dataset, indices[:-50])
dataset_test = torch.utils.data.Subset(dataset_test, indices[-50:])
# define training and validation data loaders
data_loader = torch.utils.data.DataLoader(
dataset, batch_size=4, shuffle=True, num_workers=4,
collate_fn=utils.collate_fn)
data_loader_test = torch.utils.data.DataLoader(
dataset_test, batch_size=2, shuffle=False, num_workers=4,
collate_fn=utils.collate_fn)
# get the model using our helper function
model = get_model_instance_segmentation(num_classes)
# move model to the right device
model.to(device)
# construct an optimizer
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params, lr=0.005,
momentum=0.9, weight_decay=0.0005)
# and a learning rate scheduler
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=3,
gamma=0.1)
# let's train it for 10 epochs
num_epochs = 10
for epoch in range(num_epochs):
# train for one epoch, printing every 10 iterations
train_one_epoch(model, optimizer, data_loader, device, epoch, print_freq=10)
# update the learning rate
lr_scheduler.step()
# evaluate on the test dataset
evaluate(model, data_loader_test, device=device)
print("That's it!")
if __name__ == "__main__":
main()
I am trying to use GPUs to accelerate convolution and pooling operations in my neural network application(Spiking networks). I wrote a small script to see how much speedup I can get by using Tensorflow. Surprisingly, SciPy/Numpy does better. In my application, all the inputs(images) are stored on the disk but for an example, I created a randomly initialized image of size 27x27 and weights kernel of size 5x5x30, i made sure that I am not transferring anything from CPU to GPU and I also increased the input image size to 270x270 and the weights kernel to 7x7x30, still I don't see any improvement. I made sure that all the TF methods are in fact being executed on my GPUs by setting
sess =tf.Session(config=tf.ConfigProto(log_device_placement=True))
I have access to 2 GPUs(Tesla K20m) on a cluster.
Here's my code:
import tensorflow as tf
import numpy as np
from scipy import signal
import time
sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))
image_size = 27
kernel_size = 5
nofMaps = 30
def convolution(Image, weights):
in_channels = 1 # 1 because our image has 1 units in the -z direction.
out_channels = weights.shape[-1]
strides_1d = [1, 1, 1, 1]
#in_2d = tf.constant(Image, dtype=tf.float32)
in_2d = Image
#filter_3d = tf.constant(weights, dtype=tf.float32)
filter_3d =weights
in_width = int(in_2d.shape[0])
in_height = int(in_2d.shape[1])
filter_width = int(filter_3d.shape[0])
filter_height = int(filter_3d.shape[1])
input_4d = tf.reshape(in_2d, [1, in_height, in_width, in_channels])
kernel_4d = tf.reshape(filter_3d, [filter_height, filter_width, in_channels, out_channels])
inter = tf.nn.conv2d(input_4d, kernel_4d, strides=strides_1d, padding='VALID')
output_3d = tf.squeeze(inter)
output_3d= sess.run(output_3d)
return output_3d
def pooling(Image):
in_channels = Image.shape[-1]
Image_3d = tf.constant(Image, dtype = tf.float32)
in_width = int(Image.shape[0])
in_height = int(Image.shape[1])
Image_4d = tf.reshape(Image_3d,[1,in_width,in_height,in_channels])
pooled_pots4d = tf.layers.max_pooling2d(inputs=Image_4d, pool_size=[2, 2], strides=2)
pooled_pots3d = tf.squeeze(pooled_pots4d)
return sess.run(pooled_pots3d)
t1 = time.time()
#with tf.device('/device:GPU:1'):
Image = tf.random_uniform([image_size, image_size], name='Image')
weights = tf.random_uniform([kernel_size,kernel_size,nofMaps], name='Weights')
conv_result = convolution(Image,weights)
pool_result = pooling(conv_result)
print('Time taken:{}'.format(time.time()-t1))
#with tf.device('/device:CPU:0'):
print('Pool_result shape:{}'.format(pool_result.shape))
#print('first map of pool result:\n',pool_result[:,:,0])
def scipy_convolution(Image,weights):
instant_conv1_pots = np.zeros((image_size-kernel_size+1,image_size-kernel_size+1,nofMaps))
for i in range(weights.shape[-1]):
instant_conv1_pots[:,:,i]=signal.correlate(Image,weights[:,:,i],mode='valid',method='fft')
return instant_conv1_pots
def scipy_pooling(conv1_spikes):
'''
Reshape splitting each of the two axes into two each such that the
latter of the split axes is of the same length as the block size.
This would give us a 4D array. Then, perform maximum finding along those
latter axes, which would be the second and fourth axes in that 4D array.
https://stackoverflow.com/questions/41813722/numpy-array-reshaped-but-how-to-change-axis-for-pooling
'''
if(conv1_spikes.shape[0]%2!=0): #if array is odd size then omit the last row and col
conv1_spikes = conv1_spikes[0:-1,0:-1,:]
else:
conv1_spikes = conv1_spikes
m,n = conv1_spikes[:,:,0].shape
o = conv1_spikes.shape[-1]
pool1_spikes = np.zeros((m/2,n/2,o))
for i in range(o):
pool1_spikes[:,:,i]=conv1_spikes[:,:,i].reshape(m/2,2,n/2,2).max(axis=(1,3))
return pool1_spikes
t1 = time.time()
Image = np.random.rand(image_size,image_size)
weights = np.random.rand(kernel_size,kernel_size,nofMaps)
conv_result = scipy_convolution(Image,weights)
pool_result = scipy_pooling(conv_result)
print('Time taken:{}'.format(time.time()-t1))
print('Pool_result shape:{}'.format(pool_result.shape))
#print('first map of pool result:\n',pool_result[:,:,0])
~
Results are as follows:
Time taken:0.746644973755
Pool_result shape:(11, 11, 30)
Time taken:0.0127348899841
Pool_result shape:(11, 11, 30)
With suggestions from the commenter, I set image_size=270 and enclosed both convolution and pool functions in a for loop, now, TF performs better than SciPy note that I am using tf.nn.conv2d and NOT the tf.layers.conv2d. I also set the parameter use_cudnn_on_gpu=True in tf.nn.conv2d but that didn't hurt or help.
Here's the code:
import tensorflow as tf
import numpy as np
from scipy import signal
import time
sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))
image_size = 270
kernel_size = 5
nofMaps = 30
def convolution(Image, weights):
in_channels = 1 # 1 because our image has 1 units in the -z direction.
out_channels = weights.shape[-1]
strides_1d = [1, 1, 1, 1]
#in_2d = tf.constant(Image, dtype=tf.float32)
in_2d = Image
#filter_3d = tf.constant(weights, dtype=tf.float32)
filter_3d =weights
in_width = int(in_2d.shape[0])
in_height = int(in_2d.shape[1])
filter_width = int(filter_3d.shape[0])
filter_height = int(filter_3d.shape[1])
input_4d = tf.reshape(in_2d, [1, in_height, in_width, in_channels])
kernel_4d = tf.reshape(filter_3d, [filter_height, filter_width, in_channels, out_channels])
inter = tf.nn.conv2d(input_4d, kernel_4d, strides=strides_1d, padding='VALID',use_cudnn_on_gpu=True)
output_3d = tf.squeeze(inter)
#t1 = time.time()
output_3d= sess.run(output_3d)
#print('TF Time for Conv:{}'.format(time.time()-t1))
return output_3d
def pooling(Image):
in_channels = Image.shape[-1]
Image_3d = tf.constant(Image, dtype = tf.float32)
in_width = int(Image.shape[0])
in_height = int(Image.shape[1])
Image_4d = tf.reshape(Image_3d,[1,in_width,in_height,in_channels])
pooled_pots4d = tf.layers.max_pooling2d(inputs=Image_4d, pool_size=[2, 2], strides=2)
pooled_pots3d = tf.squeeze(pooled_pots4d)
#t1 = time.time()
pool_res = sess.run(pooled_pots3d)
#print('TF Time for Pool:{}'.format(time.time()-t1))
return pool_res
#with tf.device('/device:GPU:1'):
Image = tf.random_uniform([image_size, image_size], name='Image')
weights = tf.random_uniform([kernel_size,kernel_size,nofMaps], name='Weights')
#init = tf.global_variables_initializer
#sess.run(init)
t1 = time.time()
for i in range(150):
#t1 = time.time()
conv_result = convolution(Image,weights)
pool_result = pooling(conv_result)
#print('TF Time taken:{}'.format(time.time()-t1))
print('TF Time taken:{}'.format(time.time()-t1))
#with tf.device('/device:CPU:0'):
print('TF Pool_result shape:{}'.format(pool_result.shape))
#print('first map of pool result:\n',pool_result[:,:,0])
def scipy_convolution(Image,weights):
instant_conv1_pots = np.zeros((image_size-kernel_size+1,image_size-kernel_size+1,nofMaps))
for i in range(weights.shape[-1]):
instant_conv1_pots[:,:,i]=signal.correlate(Image,weights[:,:,i],mode='valid',method='fft')
return instant_conv1_pots
def scipy_pooling(conv1_spikes):
'''
Reshape splitting each of the two axes into two each such that the
latter of the split axes is of the same length as the block size.
This would give us a 4D array. Then, perform maximum finding along those
latter axes, which would be the second and fourth axes in that 4D array.
https://stackoverflow.com/questions/41813722/numpy-array-reshaped-but-how-to-change-axis-for-pooling
'''
if(conv1_spikes.shape[0]%2!=0): #if array is odd size then omit the last row and col
conv1_spikes = conv1_spikes[0:-1,0:-1,:]
else:
conv1_spikes = conv1_spikes
m,n = conv1_spikes[:,:,0].shape
o = conv1_spikes.shape[-1]
pool1_spikes = np.zeros((m/2,n/2,o))
for i in range(o):
pool1_spikes[:,:,i]=conv1_spikes[:,:,i].reshape(m/2,2,n/2,2).max(axis=(1,3))
return pool1_spikes
Image = np.random.rand(image_size,image_size)
weights = np.random.rand(kernel_size,kernel_size,nofMaps)
t1 = time.time()
for i in range(150):
conv_result = scipy_convolution(Image,weights)
pool_result = scipy_pooling(conv_result)
print('Scipy Time taken:{}'.format(time.time()-t1))
print('Scipy Pool_result shape:{}'.format(pool_result.shape))
#print('first map of pool result:\n',pool_result[:,:,0])
Here are results:
image_size = 27x27
kernel_size = 5x5x30
iterations = 150
TF Time taken:11.0800771713
TF Pool_result shape:(11, 11, 30)
Scipy Time taken:1.4141368866
Scipy Pool_result shape:(11, 11, 30)
image_size = 270x270
kernel_size = 5x5x30
iterations = 150
TF Time taken:26.2359631062
TF Pool_result shape:(133, 133, 30)
Scipy Time taken:31.6651778221
Scipy Pool_result shape:(11, 11, 30)
image_size = 500x500
kernel_size = 5x5x30
iterations = 150
TF Time taken:89.7967050076
TF Pool_result shape:(248, 248, 30)
Scipy Time taken:143.391746044
Scipy Pool_result shape:(248, 248, 30)
In the 2nd case you can see that I got about 18% reduction in time.
In the 3rd case you can see that I goto about 38% reduction in time.
I'm trying to make speech recognition system with tensorflow.
Input data is an numpy array of size 50000 X 1.
Output data (mapping data) is an numpy array of size 400 X 1.
Input and mapping data is passed in batches of 2 in a list.
I've used this tutorial to design the neural network. Following is the code snippet:
For RNN:
input_data = tf.placeholder(tf.float32, [batch_size, sound_constants.MAX_ROW_SIZE_IN_DATA, sound_constants.MAX_COLUMN_SIZE_IN_DATA], name="train_input")
target = tf.placeholder(tf.float32, [batch_size, sound_constants.MAX_ROW_SIZE_IN_TXT, sound_constants.MAX_COLUMN_SIZE_IN_TXT], name="train_output")
fwd_cell = tf.nn.rnn_cell.BasicLSTMCell(num_hidden, state_is_tuple=True, forget_bias=1.0)
# creating one backward cell
bkwd_cell = tf.nn.rnn_cell.BasicLSTMCell(num_hidden, state_is_tuple=True, forget_bias=1.0)
# creating bidirectional RNN
val, _, _ = tf.nn.static_bidirectional_rnn(fwd_cell, bkwd_cell, tf.unstack(input_data), dtype=tf.float32)
For feeding data:
feed = {g['input_data'] : trb[0], g['target'] : trb[1], g['dropout'] : 0.6}
accuracy_, _ = sess.run([g['accuracy'], g['ts']], feed_dict=feed)
accuracy += accuracy_
When I ran the code, I got this error:
Traceback (most recent call last):
File "/home/wolborg/PycharmProjects/speech-to-text-rnn/src/rnn_train_1.py", line 205, in <module>
tr_losses, te_losses = train_network(g)
File "/home/wolborg/PycharmProjects/speech-to-text-rnn/src/rnn_train_1.py", line 177, in train_network
accuracy_, _ = sess.run([g['accuracy'], g['ts']], feed_dict=feed)
File "/home/wolborg/anaconda2/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 895, in run
run_metadata_ptr)
File "/home/wolborg/anaconda2/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 1102, in _run
raise ValueError('Tensor %s may not be fed.' % subfeed_t)
ValueError: Tensor Tensor("Const:0", shape=(), dtype=float32) may not be fed.
Process finished with exit code 1
Earlier, I was facing this issue with tf.sparse_placeholder, then after some browsing, I changed input type to tf.placeholder and made related changes. Now I'm clueless on where I'm making the error.
Please suggest something as how should I feed data.
Entire code:
import tensorflow as tf
# for taking MFCC and label input
import numpy as np
import rnn_input_data_1
import sound_constants
# input constants
# Training Parameters
num_input = 10 # mfcc data input
training_data_size = 8 # determines number of files in training and testing module
testing_data_size = num_input - training_data_size
# Network Parameters
learning_rate = 0.0001 # for large training set, it can be set 0.001
num_hidden = 200 # number of hidden layers
num_classes = 28 # total alphabet classes (a-z) + extra symbols (', ' ')
epoch = 1 # number of iterations
batch_size = 2 # number of batches
mfcc_coeffs, text_data = rnn_input_data_1.mfcc_and_text_encoding()
class DataGenerator:
def __init__(self, data_size):
self.ptr = 0
self.epochs = 0
self.data_size = data_size
def next_batch(self):
self.ptr += batch_size
if self.ptr > self.data_size:
self.epochs += 1
self.ptr = 0
return mfcc_coeffs[self.ptr-batch_size : self.ptr], text_data[self.ptr-batch_size : self.ptr]
def reset_graph():
if 'sess' in globals() and sess:
sess.close()
tf.reset_default_graph()
def struct_network():
print ('Inside struct network !!')
reset_graph()
input_data = tf.placeholder(tf.float32, [batch_size, sound_constants.MAX_ROW_SIZE_IN_DATA, sound_constants.MAX_COLUMN_SIZE_IN_DATA], name="train_input")
target = tf.placeholder(tf.float32, [batch_size, sound_constants.MAX_ROW_SIZE_IN_TXT, sound_constants.MAX_COLUMN_SIZE_IN_TXT], name="train_output")
keep_prob = tf.constant(1.0)
fwd_cell = tf.nn.rnn_cell.BasicLSTMCell(num_hidden, state_is_tuple=True, forget_bias=1.0)
# creating one backward cell
bkwd_cell = tf.nn.rnn_cell.BasicLSTMCell(num_hidden, state_is_tuple=True, forget_bias=1.0)
# creating bidirectional RNN
val, _, _ = tf.nn.static_bidirectional_rnn(fwd_cell, bkwd_cell, tf.unstack(input_data), dtype=tf.float32)
# adding dropouts
val = tf.nn.dropout(val, keep_prob)
val = tf.transpose(val, [1, 0, 2])
last = tf.gather(val, int(val.get_shape()[0]) - 1)
# creating bidirectional RNN
print ('BiRNN created !!')
print ('Last Size: ', last.get_shape())
weight = tf.Variable(tf.truncated_normal([num_hidden * 2, sound_constants.MAX_ROW_SIZE_IN_TXT]))
bias = tf.Variable(tf.constant(0.1, shape=[sound_constants.MAX_ROW_SIZE_IN_TXT]))
# mapping to 28 output classes
logits = tf.matmul(last, weight) + bias
prediction = tf.nn.softmax(logits)
prediction = tf.reshape(prediction, shape = [batch_size, sound_constants.MAX_ROW_SIZE_IN_TXT, sound_constants.MAX_COLUMN_SIZE_IN_TXT])
# getting probability distribution
mat1 = tf.cast(tf.argmax(prediction,1),tf.float32)
correct = tf.equal(prediction, target)
accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
logits = tf.reshape(logits, shape=[batch_size, sound_constants.MAX_ROW_SIZE_IN_TXT, sound_constants.MAX_COLUMN_SIZE_IN_TXT])
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=target))
train_step = tf.train.AdamOptimizer(1e-4).minimize(loss)
# returning components as dictionary elements
return {'input_data' : input_data,
'target' : target,
'dropout': keep_prob,
'loss': loss,
'ts': train_step,
'preds': prediction,
'accuracy': accuracy
}
def train_network(graph):
# initialize tensorflow session and all variables
# tf_gpu_config = tf.ConfigProto(allow_soft_placement = True, log_device_placement = True)
# tf_gpu_config.gpu_options.allow_growth = True
# with tf.Session(config = tf_gpu_config) as sess:
with tf.Session() as sess:
train_instance = DataGenerator(training_data_size)
test_instance = DataGenerator(testing_data_size)
print ('Training data size: ', train_instance.data_size)
print ('Testing data size: ', test_instance.data_size)
sess.run(tf.global_variables_initializer())
print ('Starting session...')
step, accuracy = 0, 0
tr_losses, te_losses = [], []
current_epoch = 0
while current_epoch < epoch:
step += 1
trb = train_instance.next_batch()
feed = {g['input_data'] : trb[0], g['target'] : trb[1], g['dropout'] : 0.6}
accuracy_, _ = sess.run([g['accuracy'], g['ts']], feed_dict=feed)
accuracy += accuracy_
if train_instance.epochs > current_epoch:
current_epoch += 1
tr_losses.append(accuracy / step)
step, accuracy = 0, 0
#eval test set
te_epoch = test_instance.epochs
while test_instance.epochs == te_epoch:
step += 1
print ('Testing round ', step)
trc = test_instance.next_batch()
feed = {g['input_data']: trc[0], g['target']: trc[1]}
accuracy_ = sess.run([g['accuracy']], feed_dict=feed)[0]
accuracy += accuracy_
te_losses.append(accuracy / step)
step, accuracy = 0,0
print("Accuracy after epoch", current_epoch, " - tr:", tr_losses[-1], "- te:", te_losses[-1])
return tr_losses, te_losses
g = struct_network()
tr_losses, te_losses = train_network(g)
You defined keep_prob as a tf.constant, but then trying to feed the value into it. Replace keep_prob = tf.constant(1.0) with keep_prob = tf.placeholder(tf.float32,[]) or keep_prob = tf.placeholder_with_default(1.0,[])
I seem to have some problems starting my learning... I am not sure why..
the network is multi input (72 1d arrays) and output is a 1d array length 24. the 1d array output consist of numbers related to 145 different classes.
So: 72 inputs => 24 outputs
Minimal working example - without the input/output being set.
import keras
from keras.utils import np_utils
from keras import metrics
from keras.models import Sequential
from keras.layers.core import Dense, Activation, Lambda, Reshape,Flatten
from keras.layers import Conv1D,Conv2D, MaxPooling2D, MaxPooling1D, Reshape, ZeroPadding2D
from keras.utils import np_utils
from keras.layers.advanced_activations import LeakyReLU, PReLU
from keras.layers.advanced_activations import ELU
from keras.models import Model
from keras.layers import Input, Dense
from keras.layers import Dropout
from keras import backend as K
from keras.callbacks import ReduceLROnPlateau
from keras.callbacks import CSVLogger
from keras.callbacks import EarlyStopping
from keras.models import load_model
from keras.layers.merge import Concatenate
import numpy as np
def chunks(l, n):
"""Yield successive n-sized chunks from l."""
for i in range(0, len(l), n):
yield l[i:i + n]
nano_train_input = []
nano_train_output = []
nano_test_input = []
nano_test_output = []
## Creating train input:
for i in range(974):
nano_train_input.append(np.random.random((78,684,4)))
nano_train_output.append(np.random.randint(145,size=(228)).tolist())
## Creating test input:
for i in range(104):
nano_test_input.append(np.random.random((78,684,4)))
nano_test_output.append(np.random.randint(145,size=(228)).tolist())
def model(train_input, train_output, test_input, test_output, names=0):
# Paper uses dimension (40 x 45 =(15 * 3))
# Filter size 5
# Pooling size
# I use dimension (78 x 72 = (24 * 3)
# Filter size 9
print "In model"
i = 0
print_once = True
data_test_output = []
data_test_input = []
for matrix in test_input:
row,col,channel = matrix.shape
remove_output = (col/3)%24
remove_input = col%72
if remove_output > 0 :
test_output[i] = test_output[i][:-(remove_output)]
for split in chunks(test_output[i],24):
data_test_output.append(np.array(split))
if remove_input > 0:
out = np.split(matrix[:,:-(remove_input),:-1],matrix[:,:-(remove_input),:-1].shape[1]/72,axis=1)
else:
out = np.split(matrix[:,:,:-1],matrix[:,:,:-1].shape[1]/72,axis=1)
data_test_input.extend(out)
del out
i=i+1 # Increment
i=0
data_train_output = []
data_train_input = []
for matrix in train_input:
row,col,channel = matrix.shape
remove_output = (col/3)%24
remove_input = col%72
if remove_output > 0 :
train_output[i] = train_output[i][:-(remove_output)]
for split in chunks(train_output[i],24):
data_train_output.append(np.array(split))
if remove_input > 0:
out = np.split(matrix[:,:-(remove_input),:-1],matrix[:,:-(remove_input),:-1].shape[1]/72,axis=1)
else:
out = np.split(matrix[:,:,:-1],matrix[:,:,:-1].shape[1]/72,axis=1)
data_train_input.extend(out)
del out
i=i+1 # Increment
print
print "Len:"
print len(data_train_input)
print len(data_train_output)
print len(data_test_input)
print len(data_test_output)
print
print "Type[0]:"
print type(data_train_input[0])
print type(data_train_output[0])
print type(data_test_input[0])
print type(data_test_output[0])
print
print "Type:"
print type(data_train_input)
print type(data_train_output)
print type(data_test_input)
print type(data_test_output)
print
print "shape of [0]:"
print data_train_input[0].shape
print data_train_output[0].shape
print data_test_input[0].shape
print data_test_output[0].shape
list_of_input = [Input(shape = (78,3)) for i in range(72)]
list_of_conv_output = []
list_of_max_out = []
for i in range(72):
list_of_conv_output.append(Conv1D(filters = 32 , kernel_size = 6 , padding = "same", activation = 'relu')(list_of_input[i]))
list_of_max_out.append(MaxPooling1D(pool_size=3)(list_of_conv_output[i]))
merge = keras.layers.concatenate(list_of_max_out)
reshape = Flatten()(merge)
dense1 = Dense(units = 500, activation = 'relu', name = "dense_1")(reshape)
dense2 = Dense(units = 250, activation = 'relu', name = "dense_2")(dense1)
dense3 = Dense(units = 24 , activation = 'softmax', name = "dense_3")(dense2)
model = Model(inputs = list_of_input , outputs = dense3)
model.compile(loss="categorical_crossentropy", optimizer="adam" , metrics = [metrics.sparse_categorical_accuracy])
reduce_lr=ReduceLROnPlateau(monitor='val_loss', factor=0.1, patience=3, verbose=1, mode='auto', epsilon=0.01, cooldown=0, min_lr=0.000000000000000000001)
stop = EarlyStopping(monitor='val_loss', min_delta=0, patience=5, verbose=1, mode='auto')
print "Train!"
print model.summary()
hist_current = model.fit(x = ,
y = ,
shuffle=False,
validation_data=(,),
validation_split=0.1,
epochs=150000,
verbose=1,
callbacks=[reduce_lr,stop])
model(nano_train_input,nano_train_output,nano_test_input, nano_test_output)
The input and output is stored as a list of numpy.ndarrays.
This is a minimal working example.. how am I supposed to pass the input an output?
I would try:
merge = keras.layers.concatenate(list_of_max_out)
merge = Flatten()(merge) # or GlobalMaxPooling1D or GlobalAveragePooling1D
dense1 = Dense(500, activation = 'relu')(merge)
You probably want to apply something to transform your output from Convolutional layers. In order to do that - you need to squash the time / sequential dimension. In order to do that try techniques I provided.
If you take a look at your code and outputs you indeed have what you say: 24 outputs (data_train_outputs[0].shape). However, if you look at your layer output of Keras, you have this as output:
dense_3 (Dense) (None, 26, 145) 36395
I would say that this should be an array with shape (None, 24)....
I suggest you add a reshape layer to get the output you want to have!
I have a point cloud in 4 dimensions, where each point in the cloud has a location and a value (x,y,z,Value). In addition, I have a 'special' point, S0, within the 3d point cloud; I've used this example to find the closest 10 points in the cloud, relative to S0. Now, I have a numpy array for each of the 10 closest points and their values. How can I interpolate these 10 points, to find the interpolated value at point S0? Example code is shown below:
import numpy as np
import matplotlib.pyplot as plt
numpoints = 20
linexs = 320
lineys = 40
linezs = 60
linexe = 20
lineye = 20
lineze = 0
# Create vectors of points
xpts = np.linspace(linexs, linexe, numpoints)
ypts = np.linspace(lineys, lineye, numpoints)
zpts = np.linspace(linezs, lineze, numpoints)
lin = np.dstack((xpts,ypts,zpts))
# Image line of points
fig = plt.figure()
ax = fig.add_subplot(211, projection='3d')
ax.set_xlim(0,365); ax.set_ylim(-85, 85); ax.set_zlim(0, 100)
ax.plot_wireframe(xpts, ypts, zpts)
ax.view_init(elev=12, azim=78)
def randrange(n, vmin, vmax):
return (vmax - vmin)*np.random.rand(n) + vmin
n = 10
for n in range(21):
xs = randrange(n, 0, 350)
ys = randrange(n, -75, 75)
zs = randrange(n, 0, 100)
ax.scatter(xs, ys, zs)
dat = np.dstack((xs,ys,zs))
ax.set_xlabel('X Label')
ax.set_xlim(0,350)
ax.set_ylabel('Y Label')
ax.set_ylim(-75,75)
ax.set_zlabel('Z Label')
ax.set_zlim(0,100)
ax = fig.add_subplot(212, projection='3d')
ax.set_xlim(0,365); ax.set_ylim(-85, 85); ax.set_zlim(0, 100)
ax.plot_wireframe(xpts,ypts,zpts)
ax.view_init(elev=12, azim=78)
plt.show()
dist = []
# Calculate distance from first point to all other points in cloud
for l in range(len(xpts)):
aaa = lin[0][0]-dat
dist.append(np.sqrt(aaa[0][l][0]**2+aaa[0][l][1]**2+aaa[0][l][2]**2))
full = np.dstack((dat,dist))
aaa = full[0][full[0][:,3].argsort()]
print(aaa[0:10])
A basic example. Note that the meshgrid is not needed for the interpolation, but only to make a fast ufunc to generate an example function A=f(x,y,z), here A=x+y+z.
from scipy.interpolate import interpn
import numpy as np
#make up a regular 3d grid
X=np.linspace(-5,5,11)
Y=np.linspace(-5,5,11)
Z=np.linspace(-5,5,11)
xv,yv,zv = np.meshgrid(X,Y,Z)
# make up a function
# see http://docs.scipy.org/doc/numpy/reference/ufuncs.html
A = np.add(xv,np.add(yv,zv))
#this one is easy enough for us to know what to expect at (.5,.5,.5)
# usage : interpn(points, values, xi, method='linear', bounds_error=True, fill_value=nan)
interpn((X,Y,Z),A,[0.5,0.5,0.5])
Output:
array([ 1.5])
If you pass in an array of points of interest, it will give you multiple answers.