- 在公司windows上运行代码，要在GPU上跑了。之前在mac上运行，从来没了解过GPU。。
- 同一时间，发现tflearn很好用，节省冗长的代码时间。所以查资料tflearn怎么选择在GPU上跑，因为tensorflow上要添写好几句话。但是，好想没找到怎么配置tflearn的GPU。
- 最有价值的也只有这个了：
TFLearn
- Easy-to-use and understand high-level API for implementing deep neural networks, with tutorial and examples.
(容易使用和易于理解的高层次 API 用于实现深度神经网络，附带教程和例子)- Fast prototyping through highly modular built-in neural network layers, regularizers, optimizers, metrics…
(通过高度模块化的内置神经网络层、正则化器、优化器等进行快速原型设计)- Full transparency over Tensorflow. All functions are built over tensors and can be used independently of TFLearn.
(对 TensorFlow 完全透明，所有函数都是基于 tensor，可以独立于 TFLearn 使用)- Powerful helper functions to train any TensorFlow graph, with support of multiple inputs, outputs and optimizers.
(强大的辅助函数，可训练任意 TensorFlow 图，支持多输入、多输出和优化器)- Easy and beautiful graph visualization, with details about weights, gradients, activations and more…
(简单而美观的图可视化，可以查看关于权值、梯度、特征图等细节)
Effortless device placement for using multiple CPU/GPU.
(无需人工干预，可使用多 CPU、多 GPU)
所以大概理解的就是不需要设置什么。
给一个示例吧：
tflearn.init_graph(num_cores=8, gpu_memory_fraction=0.5)
net = tflearn.input_data(shape=[None, 784])
net = tflearn.fully_connected(net, 64)
net = tflearn.dropout(net, 0.5)
net = tflearn.fully_connected(net, 10, activation='softmax')
net = tflearn.regression(net, optimizer='adam', loss='categorical_crossentropy')
model = tflearn.DNN(net)
model.fit(X, Y)
- 有时我们想要对计算资源进行限制，比如分配更多或更少的GPU RAM memory。为了达到这一目的，TFLearn提供了一个graph initializer在运行之前配置一个graph：
tflearn.init_graph(set_seed=8888, num_cores=16, gpu_memory_fraction=0.5)
查源代码发现，文件/Users/apple/anaconda3/lib/python3.6/site-packages/tflearn/config.py 里面这样写：
def init_graph(seed=None, log_device=False, num_cores=0, gpu_memory_fraction=0,
soft_placement=True):
""" init_graph.
Initialize a graph with specific parameters.
Arguments:
seed: `int`. Set the graph random seed.
log_device: `bool`. Log device placement or not.
num_cores: Number of CPU cores to be used. Default: All.
gpu_memory_fraction: A value between 0 and 1 that indicates what
fraction of the available GPU memory to pre-allocate for each
process. 1 means to pre-allocate all of the GPU memory,
0.5 means the process allocates ~50% of the available GPU
memory. Default: Use all GPU's available memory.
soft_placement: `bool`. Whether soft placement is allowed. If true,
an op will be placed on CPU if:
1. there's no GPU implementation for the OP
or
2. no GPU devices are known or registered
or
3. need to co-locate with reftype input(s) which are from CPU.
"""
if seed: tf.set_random_seed(seed)
gs = tf.GPUOptions(per_process_gpu_memory_fraction=gpu_memory_fraction)
config = tf.ConfigProto(log_device_placement=log_device,
inter_op_parallelism_threads=num_cores,
intra_op_parallelism_threads=num_cores,
gpu_options=gs,
allow_soft_placement=soft_placement)
tf.add_to_collection(tf.GraphKeys.GRAPH_CONFIG, config)
return config
