This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
def compute_loss(self, x, h, edge_indices, node_masks, edge_masks): | |
""" | |
Args: | |
x: xyz座標, shape==(B, N, 3) | |
h: OneHot encoded原子タイプ, shape==(B, N, len(settings.ATOM_MAP)) | |
edge_indices: すべての2つの原子の組み合わせ番号 shape==(B, N*N, ...) | |
node_masks: paddingされたダミー原子でないか, shape==(B, N, ...) | |
edge_masks: エッジの両端がダミー原子でないか, shape==(B, N*N, ...) | |
""" |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
import tensorflow as tf | |
import tensorflow.keras.layers as kl | |
class EquivariantGNNBlock(tf.keras.Model): | |
def __init__(self): | |
super(EquivariantGNNBlock, self).__init__() | |
self.dense_e = tf.keras.Sequential([ | |
kl.Dense(256, activation=tf.nn.silu, kernel_initializer='truncated_normal'), | |
kl.Dense(256, activation=tf.nn.silu, kernel_initializer='truncated_normal'), |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
import io | |
import tarfile | |
import urllib.request | |
GDB9_URL = "https://deepchemdata.s3-us-west-1.amazonaws.com/datasets/gdb9.tar.gz" | |
with urllib.request.urlopen(GDB9_URL) as response: | |
file = io.BytesIO(response.read()) |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
eps_t = self(x_t, t, states) | |
mu = (1.0 / tf.sqrt(1.0 - beta_t)) * (x_t - (beta_t / tf.sqrt(1.0 - alphas_cumprod_t)) * eps_t) | |
sigma = tf.sqrt(tf.reshape(tf.gather(self.variance, indices=t), (-1, 1))) | |
noise = tf.random.normal(shape=x_t.shape, mean=0., stddev=1.) | |
x_t_minus_1 = mu + sigma * noise |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
eps = tf.random.normal(shape=x_0.shape, mean=0., stddev=1.) | |
x_t = tf.sqrt(alphas_cumprod_t) * x_0 + tf.sqrt(1. - alphas_cumprod_t) * eps |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
class DiffusionPolicy(tf.keras.Model): | |
def __init__(self, action_space: int): | |
super(DiffusionPolicy, self).__init__() | |
self.n_timesteps = 5 | |
self.action_space = action_space | |
self.time_embedding = SinusoidalPositionalEmbedding(L=self.n_timesteps, D=12) | |
self.dense1 = kl.Dense(256, activation=mish) | |
self.dense2 = kl.Dense(256, activation=mish) |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
def update_policy(self, states, actions): | |
""" Advantage weighted regression | |
""" | |
q1, q2 = self.target_qnet(states, actions) | |
Q = tf.minimum(q1, q2) | |
V = self.valuenet(states) | |
exp_Adv = tf.minimum(tf.exp((Q - V) * self.temperature), 100.0) | |
with tf.GradientTape() as tape: |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
def update_value(self, states, actions): | |
""" Expectile Regression | |
""" | |
q1, q2 = self.target_qnet(states, actions) | |
target_values = tf.minimum(q1, q2) | |
with tf.GradientTape() as tape: | |
values = self.valuenet(states) | |
error = (target_values - values) |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
def update_q(self, states, actions, rewards, dones, next_states): | |
rewards = tf.clip_by_value(tf.reshape(rewards, (-1, 1)), -1.0, 1.0) | |
dones = tf.reshape(dones, (-1, 1)) | |
target_q = rewards + self.gamma * (1.0 - dones) * self.valuenet(next_states) | |
with tf.GradientTape() as tape: | |
q1, q2 = self.qnet(states, actions) |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
import numpy as np | |
import tensorflow as tf | |
import tensorflow.keras.layers as kl | |
import tensorflow_probability as tfp | |
class DecisionTransformer(tf.keras.Model): | |
def __init__(self, action_space, max_timestep, context_length=30, | |
n_blocks=6, n_heads=8, embed_dim=128): |
NewerOlder