Related
I have a convolutional neural network. And I wanted to train it on images from the training set but first they should be wrapped with my function change(tensor, float) that takes in a tensor/image of the form [hight,width,3] and a float.
Batch size =4
loading data
trainset = torchvision.datasets.CIFAR10(root='./data', train=True,
download=True, transform=transform)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=batch_size,
shuffle=True, num_workers=2)
Cnn architecture
for epoch in range(2): # loop over the dataset multiple times
running_loss = 0.0
for i, data in enumerate(trainloader, 0):
# get the inputs; data is a list of [inputs, labels]
inputs, labels = data
#size of inputs [4,3,32,32]
#size of labels [4]
inputs = change(inputs,0.1) <----------------------------
# zero the parameter gradients
optimizer.zero_grad()
# forward + backward + optimize
outputs = net(inputs) #[4, 10]
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
# print statistics
running_loss += loss.item()
if i % 2000 == 1999: # print every 2000 mini-batches
print(f'[{epoch + 1}, {i + 1:5d}] loss: {running_loss / 2000:.3f}')
running_loss = 0.0
print('Finished Training')
I am trying to apply the image function change but it gives an object error.
it there a quick way to fix it?
I am using a Julia function but it works completely fine with other objects. Error message:
JULIA: MethodError: no method matching copy(::PyObject)
Closest candidates are:
copy(!Matched::T) where T<:SHA.SHA3_CTX at /opt/julia-1.7.2/share/julia/stdlib/v1.7/SHA/src/types.jl:213
copy(!Matched::T) where T<:SHA.SHA2_CTX at /opt/julia-1.7.2/share/julia/stdlib/v1.7/SHA/src/types.jl:212
copy(!Matched::Number) at /opt/julia-1.7.2/share/julia/base/number.jl:113
I would recommend to put change function to transforms list, so you do data changes on transformation stage.
partial from functools will help you to fix number of arguments, like this:
from functools import partial
def change(input, float):
pass
# Use partial to fix number of params, such that change accepts only input
change_partial = partial(change, float=pass_float_value_here)
# Add change_partial to a list of transforms before or after converting to tensors
transforms = Compose([
RandomResizedCrop(img_size), # example
# Add change_partial here if it operates on PIL Image
change_partial,
ToTensor(), # convert to tensor
# Add change_partial here if it operates on torch tensors
change_partial,
])
I'm making a model for image denoising and use ImageDataGenerator.flow_from_directory to load the dataset. It is structured in two folders, one with noisy input images and one with the corresponding clean images. I want the generator to use the images in the first folder as inputs and the other folder as "labels"/ground truth.
With the method I'm using right now all images in both folders are treated as inputs with the folder name as label. I can extract the images manually by selecting specific batches and train on that, but it's inconvenient and probably wasn't intended to be used that way.
What is the proper way of doing this? There probably is a function for this but I can't find it.
Had similar problem. Found it necessary to create a custom generator to feed the images into model.fit. Code (rather lengthy) is posted below.
import os
import pandas as pd
import numpy as np
import glob
import cv2
from sklearn.model_selection import train_test_split
def create_df(image_dir, label_dir, shuffle=True):
path=image_dir + '/**/*'
image_file_paths=glob.glob(path,recursive=True)
path=label_dir + '/**/*'
label_file_paths=glob.glob(path,recursive=True)
# run a check and make sure filename without extensions match
df=pd.DataFrame({'image': image_file_paths, 'label':label_file_paths}).astype(str)
if shuffle:
df=df.sample(frac=1.0, replace=False, weights=None, random_state=123, axis=0).reset_index(drop=True)
return df
class jpgen():
batch_index=0 #tracks the number of batches generated
def __init__(self, df, train_split=None, test_split=None):
self.train_split=train_split # float between 0 and 1 indicating the percentage of images to use for training
self.test_split=test_split
self.df=df.copy() # create a copy of the data frame
if self.train_split != None: # split the df to create a training df
self.train_df, dummy_df=train_test_split(self.df, train_size=self.train_split, shuffle=False)
if self.test_split !=None: # create as test set and a validation set
t_split=self.test_split/(1.0-self.train_split)
self.test_df, self.valid_df=train_test_split(dummy_df, train_size=t_split, shuffle=False)
self.valid_gen_len=len(self.valid_df['image'].unique())# create var to return no of samples in valid generator
self.valid_gen_filenames=list(self.valid_df['image'])# create list ofjpg file names in valid generator
else: self.test_df=dummy_df
self.test_gen_len=len(self.test_df['image'].unique())#create var to return no of test samples
self.test_gen_filenames=list(self.test_df['image']) # create list to return jpg file paths in test_gen
else:
self.train_df=self.df
self.tr_gen_len=len(self.train_df['image'].unique()) # crete variable to return no of samples in train generator
def flow(self, batch_size=32, image_shape=None,rescale=None,shuffle=True, subset=None ):
# flows batches of jpg images and png masks to model.fit
self.batch_size=batch_size
self.image_shape=image_shape
self.shuffle=shuffle
self.subset=subset
self.rescale=rescale
image_batch_list=[] # initialize list to hold a batch of jpg images
label_batch_list=[] # initialize list to hold batches of png masks
if self.subset=='training' or self.train_split ==None:
op_df=self.train_df
elif self.subset=='test':
op_df=self.test_df
else:
op_df=self.valid_df
if self.shuffle : # shuffle the op_df then rest the index
op_df=op_df.sample(frac=1.0, replace=False, weights=None, random_state=123, axis=0).reset_index(drop=True)
#op_df will be either train, test or valid depending on subset
# develop the batch of data
while True:
label_batch_list=[]
image_batch_list=[]
start=jpgen.batch_index * self.batch_size # set start value of iteration
end=start + self.batch_size # set end value of iteration to yield 1 batch of data of length batch_size
sample_count=len(op_df['image'])
for i in range(start, end): # iterate over one batch size of data
j=i % sample_count # used to roll the images back to the front if the end is reached
k=j % self.batch_size
path_to_image= op_df.iloc[j]['image']
path_to_label= op_df.iloc[j] ['label']
label_image=cv2.imread(path_to_label, -1) # read unchanged to preserve 4 th channel print (png_image.)
label_image= cv2.cvtColor(label_image, cv2.COLOR_BGR2RGB)
image_image=cv2.imread(path_to_image)
image_image= cv2.cvtColor(image_image, cv2.COLOR_BGR2RGB)
label_image=cv2.resize(label_image, self.image_shape)
image_image=cv2.resize(image_image, self.image_shape )
if rescale !=None:
label_image=label_image * self.rescale
image_image=image_image * self.rescale
label_batch_list.append(label_image)
image_batch_list.append(image_image)
image_array=np.array(image_batch_list)
label_array=np.array(label_batch_list)
jpgen.batch_index +=1
yield (image_array, label_array)
Code below shows how to use the functions above to make generators for model.fit
image_dir=r'C:\Temp\gen_test\images'# directory with clean images
label_dir=r'C:\Temp\gen_test\labels' # directory with noisy images file names same as filenames in clean dir
shuffle=False # if True shuffles the dataframe
df=create_df(image_dir, label_dir ,shuffle) # create a dataframe with columns 'images' , 'labels'
# where labels are the noisy images
train_split=.8 # use 80% of files for training
test_split=.1 # use 10% for test, automatically sets validation split at 1-train_split-test_split
batch_size=32 # set batch_size
height=224 # set image height for generator output images and labels
width=224 # set image width for generator output images and labels
channels=3 # set number of channel in images
image_shape=(height, width)
rescale=1/255 # set value to rescale image pixels
gen=jpgen(df, train_split=train_split, test_split=test_split) # create instance of generator class
tr_gen_len=gen.tr_gen_len
test_gen_len= gen.test_gen_len
valid_gen_len=gen.valid_gen_len
test_filenames=gen.test_gen_filenames # names of test file paths used for training
train_steps=tr_gen_len//batch_size # use this value in for steps_per_epoch in model.fit
valid_steps=valid_gen_len//batch_size # use this value for validation_steps in model.fit
test_steps=test_gen_len//batch_size # use this value for steps in model.predict
# instantiate generators
train_gen=gen.flow(batch_size=batch_size, image_shape=image_shape, rescale=rescale, shuffle=False, subset='training')
valid_gen=gen.flow(batch_size=batch_size, image_shape=image_shape, rescale=rescale, shuffle=False, subset='valid')
test_gen=gen.flow(batch_size=batch_size, image_shape=image_shape, rescale=rescale, shuffle=False, subset='test')
Build your model then use
history=model.fit(train_gen, epochs=epochs, steps_per_epoch=train_steps,validation_data=valid_gen,
validation_steps=valid_steps, verbose=1, shuffle=True)
predictions=model.predict(test_gen, steps=test_steps)
I am trying to replicate the resnet18 paper. Before running this on the full Image Net dataset on disk, I'm doing some evaluation runs with the publicly available imagenette/320px dataset from TFDS (much much smaller subset of imagenet with 10 classes, already in .tfrecord format._
Note: the full notebook to do training and tracing is available here: resnet18_baseline.ipynb Just switch to a GPU runtime and run all the cells. It's already set-up with tensorboard profiling on the second batch. (You can use TPU as well, but some keras.layers.experimental.preprocessing layers do not support TPU ops yet and you have to enable soft device placement. Please use a GPU).
Input Operations
read images from the input dataset. These images usually have got different dimensions and we need some crop function because input tensors can not have different dimensions for batching. Therefore, for training I use random crop and for testing/validation datasets a center crop.
random_crop_layer = keras.layers.experimental.preprocessing.RandomCrop(224, 224)
center_crop_layer = keras.layers.experimental.preprocessing.CenterCrop(224, 224)
#tf.function(experimental_relax_shapes=True) # avoid retracing
def train_crop_fn(x, y):
return random_crop_layer(x), y
#tf.function(experimental_relax_shapes=True)
def eval_crop_fn(x, y):
return center_crop_layer(x), y
Perform some simple preprocessing/augmentations to the input data. These include rescaling to 0-1 and also scaling based on mean and stdev of the rgb colours on imagenet. Also, random
rescaling_layer = keras.layers.experimental.preprocessing.Rescaling(1./255)
train_preproc = keras.Sequential([
rescaling_layer
])
# from https://github.com/tensorflow/models/blob/master/official/vision/image_classification/preprocessing.py
# Calculated from the ImageNet training set
MEAN_RGB = (0.485 , 0.456, 0.406)
STDDEV_RGB = (0.229, 0.224, 0.225)
#tf.function
def z_score_scale(x):
return (x - MEAN_RGB) / STDDEV_RGB
#tf.function
def train_preproc_fn(x, y):
return z_score_scale(train_preproc(x)), y
#tf.function
def eval_preproc_fn(x, y):
return z_score_scale(eval_preproc(x)), y
Input Pipeline
def get_input_pipeline(input_ds, bs, crop_fn, augmentation_fn):
ret_ds = (
input_ds
.batch(1) # pre-crop are different dimensions and can't be batched
.map(crop_fn,
num_parallel_calls=tf.data.experimental.AUTOTUNE)
.unbatch()
.batch(bs)
.map(augmentation_fn, # augmentations can be batched though.
num_parallel_calls=tf.data.experimental.AUTOTUNE)
)
return ret_ds
# dataset loading
def load_imagenette():
train_ds, ds_info = tfds.load('imagenette/320px', split='train', as_supervised=True, with_info=True)
valid_ds = tfds.load('imagenette/320px', split='validation', as_supervised=True)
return train_ds, valid_ds, valid_ds, ds_info.features['label'].num_classes
# pipeline construction
train_ds, valid_ds, test_ds, num_classes = load_imagenette()
# datasets used for training (notice that I use prefetch here)
train_samples = get_input_pipeline(train_ds, BS, train_crop_fn, train_preproc_fn).prefetch(tf.data.experimental.AUTOTUNE)
valid_samples = get_input_pipeline(valid_ds, BS, eval_crop_fn, eval_preproc_fn).prefetch(tf.data.experimental.AUTOTUNE)
test_samples = get_input_pipeline(test_ds, BS, eval_crop_fn, eval_preproc_fn).prefetch(tf.data.experimental.AUTOTUNE)
Training and Profiling
I use tensorboard profiler to check the second batch size and I get a warning that this is highly input bound, with about 40% of processing wasted on inputs.
For a classic resnet18 model, you can drive the batch size up to 768 without getting a OOM error, which is what I use. A single step with bs 256 takes about 2-3 seconds.
I also get a warning that on_train_batch_size_end is slow, at around ~1.5 seconds, compared to the 1s batch time.
The model training code is very simple keras:
model.fit(
train_samples,
validation_data=valid_samples,
epochs=100,
batch_size=BS,
use_multiprocessing=True
callbacks=[tensorboard_callback, model_checkpoint_callback, early_stop_callback, reduce_lr_callback]
)
and the callbacks are specified as:
log_dir = os.path.join(os.getcwd(), 'logs')
tensorboard_callback = TensorBoard(log_dir=log_dir, update_freq="epoch", profile_batch=2)
reduce_lr_callback = tf.keras.callbacks.ReduceLROnPlateau(monitor='val_loss', factor=0.1, patience=5, min_lr=0.001, verbose=1)
model_checkpoint_callback = tf.keras.callbacks.ModelCheckpoint(filepath='model.{epoch:02d}-{val_loss:.4f}.h5',
monitor='val_loss',
verbose=1,
save_best_only=True)
early_stop_callback = keras.callbacks.EarlyStopping(monitor='val_loss', patience=15)
Lastly, here are some sample tensorboard profiling screenshots. I can't figure out how to make this run faster:
I have a netcdf file with data as a function of lon,lat and time. I would like to calculate the total number of missing entries in each grid cell summed over the time dimension, preferably with CDO or NCO so I do not need to invoke R, python etc.
I know how to get the total number of missing values
ncap2 -s "nmiss=var.number_miss()" in.nc out.nc
as I answered to this related question:
count number of missing values in netcdf file - R
and CDO can tell me the total summed over space with
cdo info in.nc
but I can't work out how to sum over time. Is there a way for example of specifying the dimension to sum over with number_miss in ncap2?
We added the missing() function to ncap2 to solve this problem elegantly as of NCO 4.6.7 (May, 2017). To count missing values through time:
ncap2 -s 'mss_val=three_dmn_var_dbl.missing().ttl($time)' in.nc out.nc
Here ncap2 chains two methods together, missing(), followed by a total over the time dimension. The 2D variable mss_val is in out.nc. The response below does the same but averages over space and reports through time (because I misinterpreted the OP).
Old/obsolete answer:
There are two ways to do this with NCO/ncap2, though neither is as elegant as I would like. Either call assemble the answer one record at a time by calling num_miss() with one record at a time, or (my preference) use the boolean comparison function followed by the total operator along the axes of choice:
zender#aerosol:~$ ncap2 -O -s 'tmp=three_dmn_var_dbl;mss_val=tmp.get_miss();tmp.delete_miss();tmp_bool=(tmp==mss_val);tmp_bool_ttl=tmp_bool.ttl($lon,$lat);print(tmp_bool_ttl);' ~/nco/data/in.nc ~/foo.nc
tmp_bool_ttl[0]=0
tmp_bool_ttl[1]=0
tmp_bool_ttl[2]=0
tmp_bool_ttl[3]=8
tmp_bool_ttl[4]=0
tmp_bool_ttl[5]=0
tmp_bool_ttl[6]=0
tmp_bool_ttl[7]=1
tmp_bool_ttl[8]=0
tmp_bool_ttl[9]=2
or
zender#aerosol:~$ ncap2 -O -s 'for(rec=0;rec<time.size();rec++){nmiss=three_dmn_var_int(rec,:,:).number_miss();print(nmiss);}' ~/nco/data/in.nc ~/foo.nc
nmiss = 0
nmiss = 0
nmiss = 8
nmiss = 0
nmiss = 0
nmiss = 1
nmiss = 0
nmiss = 2
nmiss = 1
nmiss = 2
Even though you are asking for another solution, I would like to show you that it takes only one very short line to find the answer with the help of Python. The variable m_data has exactly the same shape as a variable with missing values read using the netCDF4 package. With the execution of only one np.sum command with the correct axis specified, you have your answer.
import numpy as np
import matplotlib.pyplot as plt
import netCDF4 as nc4
# Generate random data for this experiment.
data = np.random.rand(365, 64, 128)
# Masked data, this is how the data is read from NetCDF by the netCDF4 package.
# For this example, I mask all values less than 0.1.
m_data = np.ma.masked_array(data, mask=data<0.1)
# It only takes one operation to find the answer.
n_values_missing = np.sum(m_data.mask, axis=0)
# Just a plot of the result.
plt.figure()
plt.pcolormesh(n_values_missing)
plt.colorbar()
plt.xlabel('lon')
plt.ylabel('lat')
plt.show()
# Save a netCDF file of the results.
f = nc4.Dataset('test.nc', 'w', format='NETCDF4')
f.createDimension('lon', 128)
f.createDimension('lat', 64 )
n_values_missing_nc = f.createVariable('n_values_missing', 'i4', ('lat', 'lon'))
n_values_missing_nc[:,:] = n_values_missing[:,:]
f.close()
I am using the dprint package with knitr , mainly so that I can highlight rows from a table, which I have got working, but the output image leaves a fairly large space for a footnote, and it is taking up unnecessary space.
Is there away to get rid of it?
Also since I am fairly new to dprint, if anybody has better ideas/suggestions as to how to highlight tables and make them look pretty without any footnotes... or ways to tidy up my code that would be great!
An example of the Rmd file code is below...
```{r fig.height=10, fig.width=10, dev='jpeg'}
library("dprint")
k <- data.frame(matrix(1:100, 10,10))
CBs <- style(frmt.bdy=frmt(fontfamily="HersheySans"), frmt.tbl=frmt(bty="o", lwd=1),
frmt.col=frmt(fontfamily="HersheySans", bg="khaki", fontface="bold", lwd=2, bty="_"),
frmt.grp=frmt(fontfamily="HersheySans",bg="khaki", fontface="bold"),
frmt.main=frmt(fontfamily="HersheySans", fontface="bold", fontsize=12),
frmt.ftn=frmt(fontfamily="HersheySans"),
justify="right", tbl.buf=0)
x <- dprint(~., data=k,footnote=NA, pg.dim=c(10,10), margins=c(0.2,0.2,0.2,0.2),
style=CBs, row.hl=row.hl(which(k[,1]==5), col='red'),
fit.width=TRUE, fit.height=TRUE,
showmargins=TRUE, newpage=TRUE, main="TABLE TITLE")
```
Thanks in advance!
I haven't used dprint before, but I see a couple of different things that might be causing problems:
The start of your code chunk has defined the image width and height, which dprint seems to be trying to use.
You are setting both fit.height and fit.width. I think only one of those is used (in other words, the resulting image isn't stretched to fit both height and width, but only the one that seems to make most sense, in this case, width).
After tinkering around for a minute, here's what I did that minimizes the footnote. However, I don't know if there is a more efficient way to do this.
```{r dev='jpeg'}
library("dprint")
k <- data.frame(matrix(1:100, 10,10))
CBs <- style(frmt.bdy=frmt(fontfamily="HersheySans"),
frmt.tbl=frmt(bty="o", lwd=1),
frmt.col=frmt(fontfamily="HersheySans", bg="khaki",
fontface="bold", lwd=2, bty="_"),
frmt.grp=frmt(fontfamily="HersheySans",bg="khaki",
fontface="bold"),
frmt.main=frmt(fontfamily="HersheySans", fontface="bold",
fontsize=12),
frmt.ftn=frmt(fontfamily="HersheySans"),
justify="right", tbl.buf=0)
x <- dprint(~., data=k, style=CBs, pg.dim = c(7, 4.5),
showmargins=TRUE, newpage=TRUE,
main="TABLE TITLE", fit.width=TRUE)
```
Update
Playing around to determine the sizes of the images is a total drag. But, if you run the code in R and look at the structure of x, you'll find the following:
str(x)
# List of 3
# $ cord1 : num [1:2] 0.2 6.8
# $ cord2 : Named num [1:2] 3.42 4.78
# ..- attr(*, "names")= chr [1:2] "" ""
# $ pagenum: num 2
Or, simply:
x$cord2
# 3.420247 4.782485
These are the dimensions of your resulting image, and this information can probably easily be plugged into a function to make your plots better.
Good luck!
So here's my solution...with some examples...
I've just copied and pasted my Rmd file to demonstrate how to use it.
you should be able to just copy and paste it into a blank Rmd file and then knit to HTML to see the results...
Ideally what I would have liked would have been to make it all one nice neat function rather than splitting it up into two (i.e. setup.table & print.table) but since chunk options can't be changed mid chunk as suggested by Yihui, it had to be split up into two functions...
`dprint` + `knitr` Examples to create table images
===========
```{r}
library(dprint)
# creating the sytle object to be used
CBs <- style(frmt.bdy=frmt(fontfamily="HersheySans"),
frmt.tbl=frmt(bty="o", lwd=1),
frmt.col=frmt(fontfamily="HersheySans", bg="khaki",
fontface="bold", lwd=2, bty="_"),
frmt.grp=frmt(fontfamily="HersheySans",bg="khaki",
fontface="bold"),
frmt.main=frmt(fontfamily="HersheySans", fontface="bold",
fontsize=12),
frmt.ftn=frmt(fontfamily="HersheySans"),
justify="right", tbl.buf=0)
# creating a setup function to setup printing a table (will probably put this function into my .Rprofile file)
setup.table <- function(df,width=10, style.obj='CBs'){
require(dprint)
table.style <- get(style.obj)
a <- tbl.struct(~., df)
b <- char.dim(a, style=table.style)
p <- pagelayout(dtype = "rgraphics", pg.dim = NULL, margins = NULL)
f <- size.simp(a[[1]], char.dim.obj=b, loc.y=0, pagelayout=p)
# now to work out the natural table width to height ratio (w.2.h.r) GIVEN the style
w.2.h.r <- as.numeric(f$tbl.width/(f$tbl.height +b$linespace.col+ b$linespace.main))
height <- width/w.2.h.r
table.width <- width
table.height <- height
# Setting chunk options to have right fig dimensions for the next chunk
opts_chunk$set('fig.width'=as.numeric(width+0.1))
opts_chunk$set('fig.height'=as.numeric(height+0.1))
# assigning relevant variables to be used when printing
assign("table.width",table.width, envir=.GlobalEnv)
assign("table.height",table.height, envir=.GlobalEnv)
assign("table.style", table.style, envir=.GlobalEnv)
}
# function to print the table (will probably put this function into my .Rprofile file as well)
print.table <- function(df, row.2.hl='2012-04-30', colour='lightblue',...) {
x <-dprint(~., data=df, style=table.style, pg.dim=c(table.width,table.height), ..., newpage=TRUE,fit.width=TRUE, row.hl=row.hl(which(df[,1]==row.2.hl), col=colour))
}
```
```{r}
# Giving it a go!
# Setting up two differnt size tables
small.df <- data.frame(matrix(1:100, 10,10))
big.df <- data.frame(matrix(1:800,40,20))
```
```{r}
# Using the created setup.table function
setup.table(df=small.df, width=10, style.obj='CBs')
```
```{r}
# Using the print.table function
print.table(small.df,4,'lightblue',main='table title string') # highlighting row 4
```
```{r}
setup.table(big.df,13,'CBs') # now setting up a large table
```
```{r}
print.table(big.df,38,'orange', main='the big table!') # highlighting row 38 in orange
```
```{r}
d <- style() # the default style this time will be used
setup.table(big.df,15,'d')
```
```{r}
print.table(big.df, 23, 'indianred1') # this time higlihting row 23
```