script python to plot some outpus of the model
@author: jbrlod
"""
import os
import numpy as np
import matplotlib.pyplot as plt
import shutil
import xarray as xr
from baseutil import dataset
#change the trainingset name if necessary (can be a testing dataset)
trainingname = 'training-small.nc'
#data directory
datadir = '../data'
ds = dataset(basename=os.path.join(datadir, trainingname))
from keras.models import load_model
from modelutil import masked_mse
# name of the neural networks
name = 'model_4layers'
# outputname
outname = 'dataset_nn.nc'
model = load_model(os.path.join(datadir, name),
custom_objects={'masked_mse': masked_mse})
ypredict = xr.DataArray(model.predict(ds.X), coords=ds.yt.coords)
] #,'temporal','bintemp'] # list of models directories based on the inputs data models
# TRAINING_SET_NAME & WEIGHTS BASENAME & DATA DIRECTORY
count = 0
#for i, _ in enumerate(appPattern):
# input files for training
trainingName = 'training_' + appPattern[0] + '.nc'
weightsName = 'weights_' + appPattern[0] + '.nc'
# Data Directory
trainingdir = '/net/argos/data/parvati/mkouassi/share/data/' + appPattern[0]
# Loading input dataset (training dataset)
trainingds = os.path.join(trainingdir, trainingName)
# CALL THE DATASET CLASS
ds = dataset(basename=trainingds)
# BUILDING of Binary input mask and loss function weight masks
weightdsName = os.path.join(trainingdir, weightsName)
wds = weights_mask(trainingds, weightdsName, weight_c=0.1, weight_n=1)
weight = wds.weights.expand_dims('canal', 3)
# building of binary mask for classic or binary entry
Xbinary = wds.bmask.expand_dims('canal', 3)
#f Choix du type d'entree : Binaire ou classic
if EntryData[0]: # Classic
Xbinary.values = np.ones_like(Xbinary.values)
elif EntryData[1]: # Binary
Xbinary.values = np.array(Xbinary.values, dtype='int')
## Concatenation des bases des dataArray 'yt',weights
yt_2dim = xr.concat((ds.yt, weight), dim='canal')
ddir = alldir[2]
datadir = '../data/square/'+ddir
pattern = allPattern[2]
testname = alltestname[2]
weightsName = 'Weights_Square1.nc'
elif (inputds == AllInputDs[3]):
ddir = alldir[3]
datadir = '../data/square/'+ddir
pattern = allPattern[3]
testname = alltestname[3]
weightsName = 'Weights_Squares.nc'
# CALL THE DATASET CLASS
ds = dataset(basename=os.path.join(datadir,testname))
# BUILDING of Binary input mask and loss function weight masks
Inputds = os.path.join(datadir,testname)
weightdsName = os.path.join(datadir,weightsName)
wds = weights_mask(Inputds, weightdsName, weight_c=0.1, weight_n=1)
weight = wds.weights.expand_dims('canal',3)
Xbinary = wds.bmask.expand_dims('canal',3)
## Concatenation des bases des dataArray 'yt',weights
yt_2dim = xr.concat((ds.yt, weight),dim='canal')
## MODEL DIRECTORY
from keras.models import load_model
from modelutil import context_mse, masked_mse
allLoss = ['mmse','cmse'];
] #,'temporal','bintemp'] # list of models directories based on the inputs data models
# TRAINING_SET_NAME & WEIGHTS BASENAME & DATA DIRECTORY
count = 0
for i, _ in enumerate(appPattern):
# input files for training
trainingName = 'training_' + appPattern[i] + '.nc'
weightsName = 'weights_' + appPattern[i] + '.nc'
# Data Directory
trainingdir = '../data/' + appPattern[i]
# Loading input dataset (training dataset)
trainingds = os.path.join(trainingdir, trainingName)
# CALL THE DATASET CLASS
ds = dataset(basename=trainingds)
# # BUILDING of Binary input mask and loss function weight masks
weightdsName = os.path.join(trainingdir, weightsName)
wds = weights_mask(trainingds, weightdsName, weight_c=0.1, weight_n=1)
weight = wds.weights.expand_dims('canal', 3)
# #building of binary mask for classic or binary entry
Xbinary = wds.bmask.expand_dims('canal', 3)
for _, bincl in enumerate(EntryData):
if (bincl == EntryData[0]):
Xbinary.values = np.ones_like(Xbinary.values)
elif (bincl == EntryData[1]):
Xbinary.values = np.array(Xbinary.values, dtype='int')
## Concatenation des bases des dataArray 'yt',weights
#YT_2dim = np.stack([ds.yt, weight], axis = 3); YT_2dim = YT_2dim.squeeze()
# name of the input data
basename = 'medchl-small.nc'
# name of the output training dataset
trainingname = 'training-small.nc'
# option for masking
mfun = make_mask #masking function (in baseutil)
margs = \
{'msize':8, #size of the mask
'nmask':1} #number of mask per image
#Make the dateset
ds = dataset(srcname = os.path.join(datadir,basename) , overwrite = True)
ds.masking(mfun = mfun, **margs)
ds.savebase(os.path.join(datadir,trainingname))
#plot some random images
PLOT = True
#save the images
SAVE = True
#Plot some images
if PLOT:
# example dir
exampledir = os.path.join('../figures/examples/',os.path.splitext(trainingname)[0])
shutil.rmtree(exampledir,ignore_errors=True)
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
from baseutil import dataset
#change le path of the training set if necessary
ds = dataset(basename='../data/trainingset.nc')
#for some reasons, it seems to work better if dataset is instaciated
#before importing keras
from keras.models import Model
from keras.layers.convolutional import Conv2DTranspose, Conv2D
from keras.layers.core import Activation
from keras.layers import MaxPooling2D, concatenate, Input
from modelutil import masked_mse
#import xarray as xr
name = 'model_7layers'
#ds = xr.open_dataset('../data/training.nc')
#X = ds['X'].values #.values facultatif, c'est pour avoir un np.array
#yt = ds['yt'].values
#dimension des données d'entrée
img_rows, img_cols = 64, 64
n_feat_in, filter_size, filter_size = 25, 3, 3
##tester avec 1 puis implermenter avec plus de phases internes de conv et deconv
def get_model():
#mettre kes inputs
inputs = Input(shape=(64, 64, 1))
#convolution classique
conv_1 = Conv2D(20, (3, 3), strides=(1, 1), padding='same')(inputs)