Skip to content

GitLab

Commit efa263c1 authored by Martino Bertoni's avatar Martino Bertoni 🌋
Browse files

fixed adanet

parent c5651714
Pipeline #2413 passed with stages
in 4 minutes and 59 seconds
Show whitespace changes
Inline Side-by-side
package/chemicalchecker/tool/adanet/adanet_wrap.py
View file @ efa263c1
......@@ -4,16 +4,15 @@ import shutil
import pickle
import numpy as np
import pandas as pd
from tqdm import tqdm
from time import time
from scipy.stats import pearsonr
from sklearn.metrics import r2_score, mean_squared_error
from sklearn.metrics import explained_variance_score
from sklearn.linear_model import LinearRegression
try:
import tensorflow as tf
import tensorflow.contrib.slim as slim
from tensorflow.contrib import predictor
import tensorflow.compat.v1 as tf
import tensorflow as tf2
#import tensorflow.contrib.slim as slim
#from tensorflow.contrib import predictor
except ImportError:
raise ImportError("requires tensorflow " +
"https://www.tensorflow.org/")
......@@ -93,23 +92,18 @@ class AdaNetWrapper(object):
# check the prediction task at hand
self.prediction_task = kwargs.get("prediction_task", "regression")
if self.prediction_task == "regression":
self._estimator_head = tf.contrib.estimator.regression_head(
self._estimator_head = tf.estimator.RegressionHead(
label_dimension=self.label_dimension)
elif self.prediction_task == "classification":
self._estimator_head = \
tf.contrib.estimator.binary_classification_head()
tf.estimator.BinaryClassHead()
if self.n_classes > 2:
self._estimator_head = tf.contrib.estimator.multi_class_head(
self._estimator_head = tf.estimator.MultiClassHead(
n_classes=self.n_classes)
else:
raise Exception("Prediction task '%s' not recognized.",
self.prediction_task)
# tensorflow session_config
self.session_config = tf.ConfigProto(
intra_op_parallelism_threads=self.cpu,
inter_op_parallelism_threads=self.cpu,
allow_soft_placement=True,
device_count={'CPU': self.cpu})
# log parameters
self.__log.info("**** AdaNet Parameters: ***")
......@@ -211,8 +205,7 @@ class AdaNetWrapper(object):
save_checkpoints_secs=18000, # save checkpoints every 5 hours
save_summary_steps=50000,
tf_random_seed=self.random_seed,
model_dir=self.model_dir,
session_config=self.session_config),
model_dir=self.model_dir),
model_dir=self.model_dir
)
# Train and evaluate using using the tf.estimator tooling.
......@@ -294,7 +287,7 @@ class AdaNetWrapper(object):
tf.py_function(augmentation, [x, y],
[x.dtype, y.dtype])),
num_parallel_calls=self.cpu)
iterator = dataset.make_one_shot_iterator()
iterator = tf.compat.v1.data.make_one_shot_iterator(dataset)
features, labels = iterator.get_next()
return {'x': features}, labels
......@@ -314,20 +307,22 @@ class AdaNetWrapper(object):
(regression only).
"""
if predict_fn is None:
predict_fn = predictor.from_saved_model(
model_dir, signature_def_key='predict')
imported = tf2.saved_model.load(model_dir)
predict_fn = imported.signatures["predict"]
#predict_fn = predictor.from_saved_model(
# model_dir, signature_def_key='predict')
if mask_fn is None:
# TODO if no subsampling is provided we can apply some noise
def mask_fn(data):
return data
pred = predict_fn({'x': features[:]})
pred = predict_fn(tf2.convert_to_tensor(features[:]))
if 'predictions' in pred:
if consensus:
pred_shape = pred['predictions'].shape
# axis are 0=molecules, 1=samples, 2=components
repeat = features[:].repeat(samples, axis=0)
sampling = predict_fn({'x': mask_fn(repeat)})['predictions']
sampling = predict_fn(tf2.convert_to_tensor(mask_fn(repeat)))['predictions']
sampling = sampling.reshape(
pred_shape[0], samples, pred_shape[1])
return pred['predictions'], sampling
......@@ -346,8 +341,10 @@ class AdaNetWrapper(object):
Args:
model_dir(str): path where to save the model.
"""
predict_fn = predictor.from_saved_model(
model_dir, signature_def_key='predict')
imported = tf2.saved_model.load(model_dir)
predict_fn = imported.signatures["predict"]
#predict_fn = predictor.from_saved_model(
# model_dir, signature_def_key='predict')
return predict_fn
@staticmethod
......@@ -367,8 +364,10 @@ class AdaNetWrapper(object):
probs(bool): if this is a classifier return the probabilities.
"""
if predict_fn is None:
predict_fn = predictor.from_saved_model(
model_dir, signature_def_key='predict')
imported = tf2.saved_model.load(model_dir)
predict_fn = imported.signatures["predict"]
#predict_fn = predictor.from_saved_model(
# model_dir, signature_def_key='predict')
shapes, dtypes, fn = Traintest.generator_fn(
h5_file, split, batch_size, only_x=False, return_on_epoch=True)
x_shape, y_shape = shapes
......@@ -427,7 +426,7 @@ class AdaNetWrapper(object):
with tf.Session(graph=tf.Graph()) as sess:
tf.saved_model.loader.load(sess, ["serve"], model_dir)
model_vars = tf.trainable_variables()
slim.model_analyzer.analyze_vars(model_vars, print_info=True)
#slim.model_analyzer.analyze_vars(model_vars, print_info=True)
@staticmethod
def get_trainable_variables(model_dir):
......
package/chemicalchecker/tool/adanet/cnn_generator.py
View file @ efa263c1
import functools
import adanet
import tensorflow as tf
import tensorflow.compat.v1 as tf
class SimpleCNNBuilder(adanet.subnetwork.Builder):
......
package/chemicalchecker/tool/adanet/dnn_extend_generator.py
View file @ efa263c1
import functools
import adanet
import numpy as np
import tensorflow as tf
from tensorflow import layers
from tensorflow.layers import Dense, Dropout
import tensorflow.compat.v1 as tf
from tensorflow.keras import layers
from tensorflow.keras.layers import Dense, Dropout
from chemicalchecker.util import logged
......@@ -14,7 +14,7 @@ class NanMaskingLayer(layers.Layer):
self.mask_value = mask_value
def call(self, input):
nan_idxs = tf.is_nan(input)
nan_idxs = tf.math.is_nan(input)
replace = tf.ones_like(input) * self.mask_value
return tf.where(nan_idxs, replace, input)
......@@ -92,11 +92,7 @@ class ExtendDNNBuilder(adanet.subnetwork.Builder):
def _measure_complexity(self):
"""Approximates Rademacher complexity as square-root of the depth."""
# depth_cmpl = np.sqrt(float(self._num_layers))
# max_width_cmpl = np.sqrt(float(max(self._layer_sizes)))
total_blocks_cmpl = np.sqrt(float(sum(self._layer_sizes)))
# self.__log.debug("\n\n***** COMPLEXITY\ndepth_cmpl: %s\max_width_cmpl %s\total_blocks_cmpl %s\n\n", depth_cmpl, max_width_cmpl, total_blocks_cmpl)
return total_blocks_cmpl
return tf.sqrt(tf.cast(tf.math.reduce_sum(self._layer_sizes), dtype=tf.float32))
def build_subnetwork_train_op(self, subnetwork, loss, var_list, labels,
iteration_step, summary, previous_ensemble):
......@@ -192,9 +188,9 @@ class ExtendDNNGenerator(adanet.subnetwork.Generator):
last_subnetwork = previous_ensemble.weighted_subnetworks[
-1].subnetwork
shared_tensors = last_subnetwork.shared
num_layers = tf.contrib.util.constant_value(
num_layers = tf.get_static_value(
shared_tensors["num_layers"])
layer_sizes = list(tf.contrib.util.constant_value(
layer_sizes = list(tf.get_static_value(
shared_tensors["layer_sizes"]))
# at each iteration try exdending any of the existing layers (width)
candidates = list()
......
package/chemicalchecker/tool/adanet/dnn_stack_generator.py
View file @ efa263c1
import functools
import adanet
import tensorflow as tf
from tensorflow import layers
from tensorflow.layers import Dense, Dropout
import tensorflow.compat.v1 as tf
from tensorflow.keras import layers
from tensorflow.keras.layers import Dense, Dropout
from chemicalchecker.util import logged
......@@ -15,7 +13,7 @@ class NanMaskingLayer(layers.Layer):
self.mask_value = mask_value
def call(self, input):
nan_idxs = tf.is_nan(input)
nan_idxs = tf.math.is_nan(input)
replace = tf.ones_like(input) * self.mask_value
return tf.where(nan_idxs, replace, input)
......@@ -91,7 +89,7 @@ class StackDNNBuilder(adanet.subnetwork.Builder):
def _measure_complexity(self):
"""Approximates Rademacher complexity as square-root of the depth."""
return tf.sqrt(tf.to_float(self._num_layers))
return tf.sqrt(tf.cast(self._num_layers, dtype=tf.float32))
def build_subnetwork_train_op(self, subnetwork, loss, var_list, labels,
iteration_step, summary, previous_ensemble):
......@@ -175,7 +173,7 @@ class StackDNNGenerator(adanet.subnetwork.Generator):
num_layers = 0
seed = self._seed
if previous_ensemble:
num_layers = tf.contrib.util.constant_value(
num_layers = tf.get_static_value(
previous_ensemble.weighted_subnetworks[
-1].subnetwork.persisted_tensors["num_layers"])
if seed is not None:
......
package/tests/test_adanet.py
View file @ efa263c1
......@@ -27,6 +27,7 @@ class TestAdanet(unittest.TestCase):
shutil.rmtree(self.adanet_path, ignore_errors=True)
def tearDown(self):
return
if os.path.exists(self.adanet_path):
shutil.rmtree(self.adanet_path, ignore_errors=True)
......@@ -54,7 +55,7 @@ class TestAdanet(unittest.TestCase):
self.assertAlmostEqual(res['accuracy'], 0.981, 2)
self.assertAlmostEqual(res['auc'], 0.994, 3)
self.assertAlmostEqual(res['precision'], 0.981, 2)
self.assertAlmostEqual(res['recall'], 0.9822178, 2)
self.assertAlmostEqual(res['recall'], 0.98761237, 2)
# check persistency and predict
predict_fn = AdaNet.predict_fn(ada.save_dir)
y_pred, y_true = AdaNet.predict_online(file_path, 'test', predict_fn)
......@@ -88,7 +89,7 @@ class TestAdanet(unittest.TestCase):
# check results
_, (res, _) = ada.train_and_evaluate()
self.assertAlmostEqual(res['accuracy'], 0.9668, 2)
self.assertAlmostEqual(res['loss'], 0.12861905, 2)
self.assertAlmostEqual(res['loss'], 0.13678956, 2)
# check persistency
predict_fn = AdaNet.predict_fn(ada.save_dir)
y_pred, y_true = AdaNet.predict_online(file_path, 'test', predict_fn)
......@@ -120,7 +121,7 @@ class TestAdanet(unittest.TestCase):
self.assertEqual(ada.model_dir, self.adanet_path)
# check results
_, (res, _) = ada.train_and_evaluate()
self.assertAlmostEqual(res['loss'], 7.290054, 2)
self.assertAlmostEqual(res['loss'], 3.0210078, 2)
# check persistency and predict
predict_fn = AdaNet.predict_fn(ada.save_dir)
y_pred, y_true = AdaNet.predict_online(file_path, 'test', predict_fn)
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment