def f(**kwargs):
kwargs['objective'] = 'reg:squarederror'
kwargs = Jsonize(kwargs)()
model = Model(inputs=inputs,
outputs=outputs,
lookback=1,
batches="2d",
val_data="same",
test_fraction=0.3,
model={"xgboostregressor": kwargs},
transformation=None,
data=data,
prefix='testing',
verbosity=0)
model.fit(indices="random")
t, p = model.predict(indices=model.test_indices, prefix='test')
mse = RegressionMetrics(t, p).mse()
print(f"Validation mse {mse}")
return mse
def test_random_idx_with_nan_inputs_outputs(self):
"""
Test that when nans are present in inputs and outputs and we use random indices, then x,y data is correctly made.
"""
df = get_df_with_nans(inputs=True, outputs=True, frac=0.1)
model = Model(inputs=['in1', 'in2'],
outputs=['out1'],
transformation=None,
val_data='same',
test_fraction=0.3,
epochs=1,
data=df,
input_nans={'fillna': {
'method': 'bfill'
}},
verbosity=1)
model.fit(indices='random')
x, _, y = model.train_data(indices=model.train_indices)
# for i in range(100):
# idx = model.train_indices[i]
# df_x = df[['in1', 'in2']].iloc[idx]
# if idx > model.lookback and int(df_x.isna().sum()) == 0:
# self.assertAlmostEqual(float(df['out1'].iloc[idx]), y[i], 6)
# self.assertTrue(np.allclose(df[['in1', 'in2']].iloc[idx], x[0][i, -1]))
assert np.max(model.test_indices) < (
model.data.shape[0] - int(model.data[model.out_cols].isna().sum()))
assert np.max(model.train_indices) < (
model.data.shape[0] - int(model.data[model.out_cols].isna().sum()))
return
def run_same_train_val_data(**kwargs):
model = Model(data=nasdaq_df,
val_data="same",
test_fraction=0.2,
epochs=1,
verbosity=0)
model.fit(**kwargs)
x, _, y = model.train_data(indices=model.train_indices)
return x, y
def fn(**suggestion):
model = Model(inputs=inputs,
outputs=outputs,
model={"xgboostregressor": suggestion},
data=data,
prefix=f'test_{algorithm}_xgboost_{backend}',
verbosity=0)
model.fit(indices="random")
t, p = model.predict(indices=model.test_indices, prefix='test')
mse = RegressionMetrics(t, p).mse()
return mse
def objective_fn(**suggestion):
print(suggestion, 'suggestion')
model = Model(model={
'layers': {
'lstm': {
'config': {
'units': 64,
'activation': suggestion['activation'],
'dropout': 0.2,
'recurrent_dropout': 0.2
}
}
}
},
inputs=inputs,
outputs=outputs,
lookback=int(suggestion['lookback']),
lr=float(suggestion['lr']),
batch_size=int(suggestion['batch_size']),
data=data['224206'],
verbosity=0,
epochs=500,
prefix=_suffix)
h = model.fit()
return np.min(h.history['val_loss'])
def fn(**suggestion):
model = Model(inputs=inputs,
outputs=outputs,
model={"xgboostregressor": suggestion},
data=data,
prefix='test_tpe_xgboost',
verbosity=0)
model.fit(indices="random")
t, p = model.predict(indices=model.test_indices, prefix='test')
mse = RegressionMetrics(t, p).mse()
print(f"Validation mse {mse}")
return mse
def test_random_idx_with_nan_in_outputs(self):
# testing that if output contains nans and we use random indices, then correct examples are assinged
# for training and testing given val_data is 'same'.
df = get_df_with_nans(inputs=False, outputs=True, frac=0.8)
model = Model(inputs=['in1', 'in2'],
outputs=['out1'],
transformation=None,
val_data='same',
test_fraction=0.3,
epochs=1,
data=df,
verbosity=0)
model.fit(indices='random')
idx5 = [50, 0, 72, 153, 39, 31, 170, 8] # last 8 train indices
self.assertTrue(np.allclose(idx5, model.train_indices[-8:]))
x, _, y = model.train_data(indices=model.train_indices)
eighth_non_nan_val_4m_st = df['out1'][df['out1'].notnull()].iloc[8]
# the last training index is 8, so the last y value must be 8th non-nan value
self.assertAlmostEqual(float(y[-1]), eighth_non_nan_val_4m_st)
# checking that x values are also correct
eighth_non_nan_val_4m_st = df[['in1',
'in2']][df['out1'].notnull()].iloc[8]
self.assertTrue(
np.allclose(df[['in1', 'in2']].iloc[86], eighth_non_nan_val_4m_st))
self.assertTrue(np.allclose(x[0][-1, -1], eighth_non_nan_val_4m_st))
xx, _, yy = model.test_data(indices=model.test_indices)
# the second test index is 9, so second value of yy must be 9th non-nan value
self.assertEqual(model.test_indices[2], 10)
self.assertAlmostEqual(float(yy[2]),
df['out1'][df['out1'].notnull()].iloc[10])
self.assertTrue(
np.allclose(xx[0][2, -1],
df[['in1', 'in2']][df['out1'].notnull()].iloc[10]))
assert np.max(model.test_indices) < (
model.data.shape[0] - int(model.data[model.out_cols].isna().sum()))
assert np.max(model.train_indices) < (
model.data.shape[0] - int(model.data[model.out_cols].isna().sum()))
return
def build_and_run(outputs, transformation=None, indices=None):
model = Model(model={"layers": make_layers(len(outputs['inp_1d']))},
lookback=lookback,
inputs={
"inp_1d": inp_1d,
"inp_2d": inp_2d
},
outputs=outputs,
data={
'inp_1d': make_1d(outputs['inp_1d']),
'inp_2d': data_2d
},
transformation=transformation,
epochs=2,
verbosity=0)
model.fit(indices=indices)
return model.predict(indices=model.test_indices if indices else None)
def test_ml_random_indices(self):
model = Model(inputs=data_reg['feature_names'],
outputs=["target"],
lookback=1,
batches="2d",
val_fraction=0.0,
val_data="same",
test_fraction=0.3,
category="ML",
problem="regression",
model={"xgboostregressor": {}},
transformation=None,
data=df_reg,
verbosity=0)
model.fit(indices="random")
trtt, trp = model.predict(indices=model.train_indices, prefix='train')
t, p = model.predict(indices=model.test_indices, prefix='test')
self.assertGreater(len(t), 1)
self.assertGreater(len(trtt), 1)
return
def test_datetimeindex(self):
# makes sure that using datetime_index=True during prediction, the returned values are in correct order
model = Model(data=data1,
inputs=in_cols,
outputs=out_cols,
epochs=2,
model={
'layers': {
"LSTM": {
"config": {
"units": 2
}
},
"Dense": {
"config": {
"units": 1
}
},
"Reshape": {
"config": {
"target_shape": (1, 1)
}
}
}
},
lookback=lookback,
verbosity=0)
model.fit(indices="random")
t, p = model.predict(indices=model.train_indices,
use_datetime_index=True)
# the values in t must match the corresponding indices after adding 10000, because y column starts from 100000
for i in range(100):
self.assertEqual(int(t[i]), model.train_indices[i] + 10000)
return
def test_multi_out_nans(self):
"""
Test that when multiple outputs are the target and they contain nans, then we ignore these nans during
loss calculation.
"""
if int(''.join(tf.__version__.split('.')[0:2])) < 23 or int(
tf.__version__[0]) < 2:
warnings.warn(
f"test with ignoring nan in labels can not be done in tf version {tf.__version__}"
)
else:
df = get_df_with_nans(200,
inputs=False,
outputs=True,
output_cols=['out1', 'out2'],
frac=0.5)
layers = {
"Flatten": {
"config": {}
},
"Dense": {
"config": {
"units": 2
}
},
"Reshape": {
"config": {
"target_shape": (2, 1)
}
}
}
model = Model(allow_nan_labels=True,
model={'layers': layers},
inputs=['in1', 'in2'],
outputs=['out1', 'out2'],
epochs=10,
verbosity=0,
data=df)
history = model.fit()
self.assertTrue(np.abs(np.sum(history.history['nse'])) > 0.0)
self.assertTrue(np.abs(np.sum(history.history['val_nse'])) > 0.0)
return
def test_nan_labels1(self):
if int(''.join(tf.__version__.split('.')[0:2])) < 23 or int(
tf.__version__[0]) < 2:
warnings.warn(
f"test with ignoring nan in labels can not be done in tf version {tf.__version__}"
)
else:
df = get_df_with_nans(500,
inputs=False,
outputs=True,
output_cols=['out1', 'out2'],
frac=0.9)
layers = {
"Flatten": {
"config": {}
},
"Dense": {
"config": {
"units": 2
}
},
"Reshape": {
"config": {
"target_shape": (2, 1)
}
}
}
model = Model(allow_nan_labels=1,
transformation=None,
model={'layers': layers},
inputs=['in1', 'in2'],
outputs=['out1', 'out2'],
epochs=10,
verbosity=0,
data=df.copy())
history = model.fit(indices='random')
self.assertFalse(
any(np.isin(model.train_indices, model.test_indices)))
self.assertTrue(np.abs(np.sum(history.history['val_nse'])) > 0.0)
return
def run_class_test(method):
problem = "classification" if method.lower().startswith(
"class") else "regression"
if method not in [
"STACKINGREGRESSOR",
"VOTINGREGRESSOR",
"LOGISTICREGRESSIONCV", # has convergence issues
"RIDGE_REGRESSION",
"MULTIOUTPUTREGRESSOR",
"REGRESSORCHAIN",
"REGRESSORMIXIN",
# classifications methods
"STACKINGCLASSIFIER",
"VOTINGCLASSIFIER",
"CLASSIFIERCHAIN",
"CLASSIFIERMIXIN",
"MULTIOUTPUTCLASSIFIER",
"CHECK_CLASSIFICATION_TARGETS",
"IS_CLASSIFIER"
]:
kwargs = {}
if "CATBOOST" in method:
kwargs = {'iterations': 2}
elif "TPOT" in method.upper():
kwargs = {'generations': 2, 'population_size': 2}
print(f"testing {method}")
model = Model(inputs=data_reg['feature_names'] if problem
== "regression" else data_class['feature_names'],
outputs=['target'],
val_fraction=0.2,
problem=problem,
transformation=None,
data=df_reg if problem == "regression" else data_class,
model={method: kwargs},
verbosity=0)
return model.fit()
def test_as_fns(self):
layers = {}
for idx, act_fn in enumerate([
'tanh', 'relu', 'elu', 'leakyrelu', 'crelu', 'selu', 'relu6',
'sigmoid', 'hardsigmoid', 'swish'
]):
layers["Dense_" + str(idx)] = {
'config': {
'units': 1,
'activation': act_fn
}
}
layers["reshape"] = {'config': {'target_shape': (1, 1)}}
model = Model(epochs=2,
lookback=1,
model={'layers': layers},
data=df,
verbosity=0)
history = model.fit()
val = {
'21_nt': [0.8971164431680119, 0.7911620726129243],
'23_nt': [0.10781528055667877, 0.09552989155054092],
'24_nt': [0.10781528055667877, 0.09552989155054092],
'23_posix': [0.10781528055667877, 0.09552989155054092],
'24_posix': [0.10781528055667877, 0.09552989155054092],
'21_posix': [0.10688107734841351, 0.0938945620801094],
'20_posix': [0.8971164431680119, 0.10688107734841351]
}
if int(tf.__version__.split('.')[0]) > 1:
for t, p in zip(history.history['val_loss'],
val[version + '_' + os.name]):
self.assertAlmostEqual(t, p, 2)
return
def test_as_layers(self):
layers = {}
for lyr in [
'PRELU', "RELU", "TANH", "ELU", "LEAKYRELU", "THRESHOLDRELU",
"SELU", 'sigmoid', 'hardsigmoid', 'crelu', 'relu6', 'softmax',
'softplus', 'softsign', 'swish'
]:
layers[lyr] = {'config': {}}
layers["Dense"] = {'config': {'units': 1}}
layers["reshape"] = {'config': {'target_shape': (1, 1)}}
model = Model(epochs=2,
lookback=1,
model={'layers': layers},
data=df,
verbosity=0)
val = {
'21_nt': [0.09297575600513237, 0.09400989675627566],
'23_posix': [0.0870760977268219, 0.1053781732916832],
'24_posix': [0.0870760977268219, 0.1053781732916832],
'21_posix': [0.09297575600513237, 0.095427157656984],
'20_posix': [0.09297575600513237, 0.095427157656984],
'23_nt': [0.0870760977268219, 0.1053781732916832],
'24_nt': [0.0870760977268219, 0.1053781732916832]
}
history = model.fit()
if int(tf.__version__.split('.')[0]) > 1:
for t, p in zip(history.history['val_loss'],
val[version + '_' + os.name]):
self.assertAlmostEqual(t, p, 2)
return
#How to use AI4Water for regression problems
import pandas as pd
import numpy as np
from sklearn.datasets import load_diabetes
from AI4Water import Model
data_class = load_diabetes()
cols = data_class['feature_names'] + ['target']
df = pd.DataFrame(np.concatenate(
[data_class['data'], data_class['target'].reshape(-1, 1)], axis=1),
columns=cols)
model = Model(
data=df,
inputs=data_class['feature_names'],
outputs=['target'],
lookback=1,
batches="2d",
val_fraction=0.0,
model={'DecisionTreeRegressor': {
"max_depth": 3,
"criterion": "mae"
}},
transformation=None)
h = model.fit()
x, _, y = model.train_data()
# Put channel and spatial attention of CBAM model for time-series prediction
from AI4Water import Model
from AI4Water.utils.datasets import arg_beach
layers = {
"Conv1D": {"config": {"filters": 64, "kernel_size": 7}},
"MaxPool1D": {"config": {}},
"ChannelAttention": {"config": {"conv_dim": "1d", "in_planes": 32}},
"SpatialAttention": {"config": {"conv_dim": "1d"}},
"Flatten": {"config": {}},
"Dense": {"config": {"units": 1}},
"Reshape": {"config": {"target_shape": (1,1)}}
}
model = Model(
model={'layers':layers},
lookback=10,
data=arg_beach())
history = model.fit(indices="random")
"Dense_0": {
'units': 64,
'activation': 'relu'
},
"Flatten": {},
"Dense_3": {
'units': 1
},
}
}
df = arg_beach()
input_features = list(df.columns)[0:-1]
# column in dataframe to bse used as output/target
outputs = list(df.columns)[-1]
model = Model(data=df,
batch_size=16,
lookback=1,
model=mlp_model,
inputs=input_features,
outputs=[outputs],
lr=0.0001)
history = model.fit(indices='random')
y, obs = model.predict()
model.view_model(st=0)
"config": {
"units": units,
"return_sequences": True
}
},
"Flatten": {
"config": {}
},
"Dense": {
"config": {
"units": outs
}
},
}
model = Model(model={'layers': layers},
lookback=lookback,
epochs=epochs,
batch_size=batch_size,
inputs=[f'in_{i}' for i in range(ins)],
outputs=['out'],
data=None)
x = np.random.random((examples, lookback, ins))
y = np.random.random((examples, outs, 1))
model.fit(data=(x, y))
model.plot_layer_outputs(data=(x, y))
model.plot_act_grads(data=(x, y))
model.plot_weights()
model.plot_weight_grads(data=(x, y))