molcraft 0.1.0rc10__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- molcraft/__init__.py +18 -0
- molcraft/callbacks.py +100 -0
- molcraft/chem.py +714 -0
- molcraft/datasets.py +132 -0
- molcraft/descriptors.py +149 -0
- molcraft/features.py +379 -0
- molcraft/featurizers.py +727 -0
- molcraft/layers.py +2034 -0
- molcraft/losses.py +37 -0
- molcraft/models.py +627 -0
- molcraft/ops.py +195 -0
- molcraft/records.py +187 -0
- molcraft/tensors.py +561 -0
- molcraft/trainers.py +212 -0
- molcraft-0.1.0rc10.dist-info/METADATA +118 -0
- molcraft-0.1.0rc10.dist-info/RECORD +19 -0
- molcraft-0.1.0rc10.dist-info/WHEEL +5 -0
- molcraft-0.1.0rc10.dist-info/licenses/LICENSE +21 -0
- molcraft-0.1.0rc10.dist-info/top_level.txt +1 -0
molcraft/__init__.py
ADDED
|
@@ -0,0 +1,18 @@
|
|
|
1
|
+
__version__ = '0.1.0rc10'
|
|
2
|
+
|
|
3
|
+
import os
|
|
4
|
+
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
|
|
5
|
+
|
|
6
|
+
from molcraft import chem
|
|
7
|
+
from molcraft import features
|
|
8
|
+
from molcraft import descriptors
|
|
9
|
+
from molcraft import featurizers
|
|
10
|
+
from molcraft import layers
|
|
11
|
+
from molcraft import models
|
|
12
|
+
from molcraft import ops
|
|
13
|
+
from molcraft import records
|
|
14
|
+
from molcraft import tensors
|
|
15
|
+
from molcraft import callbacks
|
|
16
|
+
from molcraft import datasets
|
|
17
|
+
from molcraft import losses
|
|
18
|
+
from molcraft import trainers
|
molcraft/callbacks.py
ADDED
|
@@ -0,0 +1,100 @@
|
|
|
1
|
+
import warnings
|
|
2
|
+
import keras
|
|
3
|
+
import numpy as np
|
|
4
|
+
|
|
5
|
+
|
|
6
|
+
class TensorBoard(keras.callbacks.TensorBoard):
|
|
7
|
+
|
|
8
|
+
def _log_weights(self, epoch):
|
|
9
|
+
with self._train_writer.as_default():
|
|
10
|
+
for layer in self.model.layers:
|
|
11
|
+
for weight in layer.weights:
|
|
12
|
+
# Use weight.path istead of weight.name to distinguish
|
|
13
|
+
# weights of different layers.
|
|
14
|
+
histogram_weight_name = weight.path + "/histogram"
|
|
15
|
+
self.summary.histogram(
|
|
16
|
+
histogram_weight_name, weight, step=epoch
|
|
17
|
+
)
|
|
18
|
+
if self.write_images:
|
|
19
|
+
image_weight_name = weight.path + "/image"
|
|
20
|
+
self._log_weight_as_image(
|
|
21
|
+
weight, image_weight_name, epoch
|
|
22
|
+
)
|
|
23
|
+
self._train_writer.flush()
|
|
24
|
+
|
|
25
|
+
|
|
26
|
+
class LearningRateDecay(keras.callbacks.LearningRateScheduler):
|
|
27
|
+
|
|
28
|
+
def __init__(self, rate: float, delay: int = 0, **kwargs):
|
|
29
|
+
|
|
30
|
+
def lr_schedule(epoch: int, lr: float):
|
|
31
|
+
if epoch < delay:
|
|
32
|
+
return float(lr)
|
|
33
|
+
return float(lr * keras.ops.exp(-rate))
|
|
34
|
+
|
|
35
|
+
super().__init__(schedule=lr_schedule, **kwargs)
|
|
36
|
+
|
|
37
|
+
|
|
38
|
+
class Rollback(keras.callbacks.Callback):
|
|
39
|
+
"""Rollback callback.
|
|
40
|
+
|
|
41
|
+
Currently, this callback simply restores the model and (optionally) the optimizer
|
|
42
|
+
variables if current loss deviates too much from the best observed loss.
|
|
43
|
+
|
|
44
|
+
This callback might be useful in situations where the loss tend to spike and put
|
|
45
|
+
the model in an undesired/problematic high-loss parameter space.
|
|
46
|
+
|
|
47
|
+
Args:
|
|
48
|
+
tolerance (float):
|
|
49
|
+
The threshold for when the restoration is triggered. The devaiation is
|
|
50
|
+
calculated as follows: (current_loss - best_loss) / best_loss.
|
|
51
|
+
"""
|
|
52
|
+
|
|
53
|
+
def __init__(
|
|
54
|
+
self,
|
|
55
|
+
tolerance: float = 0.5,
|
|
56
|
+
rollback_optimizer: bool = True,
|
|
57
|
+
):
|
|
58
|
+
super().__init__()
|
|
59
|
+
self.tolerance = tolerance
|
|
60
|
+
self.rollback_optimizer = rollback_optimizer
|
|
61
|
+
|
|
62
|
+
def on_train_begin(self, logs=None):
|
|
63
|
+
self._rollback_weights = self._get_model_vars()
|
|
64
|
+
if self.rollback_optimizer:
|
|
65
|
+
self._rollback_optimizer_vars = self._get_optimizer_vars()
|
|
66
|
+
self._rollback_loss = float('inf')
|
|
67
|
+
|
|
68
|
+
def on_epoch_end(self, epoch: int, logs: dict = None):
|
|
69
|
+
current_loss = logs.get('val_loss', logs.get('loss'))
|
|
70
|
+
deviation = (current_loss - self._rollback_loss) / self._rollback_loss
|
|
71
|
+
|
|
72
|
+
if np.isnan(current_loss) or np.isinf(current_loss):
|
|
73
|
+
self._rollback()
|
|
74
|
+
# Rolling back model because of nan or inf loss
|
|
75
|
+
return
|
|
76
|
+
|
|
77
|
+
if deviation > self.tolerance:
|
|
78
|
+
self._rollback()
|
|
79
|
+
# Rolling back model because of large loss deviation.
|
|
80
|
+
return
|
|
81
|
+
|
|
82
|
+
if current_loss < self._rollback_loss:
|
|
83
|
+
self._save_state(current_loss)
|
|
84
|
+
|
|
85
|
+
def _save_state(self, current_loss: float) -> None:
|
|
86
|
+
self._rollback_loss = current_loss
|
|
87
|
+
self._rollback_weights = self._get_model_vars()
|
|
88
|
+
if self.rollback_optimizer:
|
|
89
|
+
self._rollback_optimizer_vars = self._get_optimizer_vars()
|
|
90
|
+
|
|
91
|
+
def _rollback(self) -> None:
|
|
92
|
+
self.model.set_weights(self._rollback_weights)
|
|
93
|
+
if self.rollback_optimizer:
|
|
94
|
+
self.model.optimizer.set_weights(self._rollback_optimizer_vars)
|
|
95
|
+
|
|
96
|
+
def _get_optimizer_vars(self):
|
|
97
|
+
return [v.numpy() for v in self.model.optimizer.variables]
|
|
98
|
+
|
|
99
|
+
def _get_model_vars(self):
|
|
100
|
+
return self.model.get_weights()
|