| fcn32model = fcn32_blank() | |
| fcn32shape = fcn32model.layers[-1].output_shape | |
| fcn32size = fcn32shape[2] # INFO: =32 when images are 512x512 | |
| sp4 = Convolution2D(21, 1, 1, | |
| border_mode='same', # INFO : border_mode does not matter for 1x1 | |
| activation=None, | |
| name='score_pool4') |
| if pretrained: | |
| EMBEDDING = Embedding(vocab_size, | |
| embedding_dimension, | |
| weights=[glove_embedding_matrix]) | |
| else: | |
| EMBEDDING = Embedding(vocab_size, | |
| embedding_dimension) | |
| modelWRD = Sequential() | |
| # 1st layer is a dummy-permutation=identity to specify input shape | |
| modelWRD.add( Permute((1,), input_shape=(n_words,)) ) | |
| modelWRD.add( EMBEDDING ) | |
| modelWRD.add( Lambda( | |
| lambda x : K.sum(x,axis=1), # sum over words |
WARNING : Python 2.7 is much easier to run with Caffe than Python 3.5
https://iqbalnaved.wordpress.com/2016/05/08/installing-caffe-and-pycaffe-on-ubuntu-14-04/ WARNING : Command line at step 2 is actually 2 lines WARNING : You dont have to use cmake for compiling, see berkeleyvision
http://caffe.berkeleyvision.org/install_apt.html WARNING : you probably dont need all the dependencies, especially the "manual" ones
| addToIndex = function (bucketName, docName, docContent, context) { | |
| var csd = new AWS.CloudSearchDomain({ | |
| endpoint: CS_NAME+'.'+SERVICES_REGION+'.cloudsearch.amazonaws.com', | |
| apiVersion: '2013-01-01' | |
| }); | |
| // see documentation at : | |
| // http://docs.aws.amazon.com/cloudsearch/latest/developerguide/preparing-data.html#creating-document-batches | |
| var jbatch = [ {"type": "add", |
| # before filling the null keep track of them | |
| dfX['construction_year_missing'] = (dfX['construction_year']==0)*1 | |
| dates.append( 'construction_year_missing' ) # list of dates related fields | |
| # to fill missing dates, can use : mean, median or oldest | |
| mean_year = dfX[dfX['construction_year']>0]['construction_year'].mean() | |
| dfX.loc[dfX['construction_year']==0, 'construction_year'] = int(mean_year) |
| # Split date_recorded stamp (string) into year, month, day of month, day of week features | |
| from dateutil import parser | |
| dfX['date_recorded_year'] = dfX['date_recorded'].apply(lambda x: int(x.split('-')[0])) | |
| dates.append('date_recorded_year') | |
| dfX['date_recorded_month'] = dfX['date_recorded'].apply(lambda x: int(x.split('-')[1])) | |
| dates.append('date_recorded_month') | |
| # WARNING : probably not usefull for this dataset |
| import pandas as pd | |
| dfX = pd.read_csv('PUMP_training_set_values.csv') # predictive variables | |
| dfY = pd.read_csv('PUMP_training_set_labels.csv') # target variable |
| dfOHE = None | |
| for col in categories: # encode 1 category at a time | |
| one_hot = pd.get_dummies(df[col], prefix=col) | |
| # drop column as it is now encoded | |
| if dfOHE is None: | |
| dfOHE = df.drop(col, axis=1) | |
| else: | |
| dfOHE = dfOHE.drop(col, axis=1) |
| from sklearn.linear_model import LogisticRegression | |
| dfLR = dfOHE.sample(frac=1) # shuffle the dataset before spliting it in 2 parts | |
| dfLR_trn = dfLR[0:45000] # training set | |
| dfLR_tst = dfLR[45000:] # testing set | |
| LR = LogisticRegression(multi_class='ovr') # ovr = one (class) versus rest (of classes) | |
| LR.fit(dfLR_trn[predictors].values, dfLR_trn['status_group_enc'].values) | |
| # model accuracy score between 0% and 100% |