Skip to content

Latest commit

 

History

History
194 lines (140 loc) · 3.94 KB

tensorflow.md

File metadata and controls

194 lines (140 loc) · 3.94 KB

tensorflow

install

# create env
conda create -n tensorflow python=3.6

# activate
source activate tensorflow

# install dep
pip install --ignore-installed --upgrade tensorflow
conda install -c anaconda protobuf libgcc
pip install pillow lxml Cython jupyter matplotlib pandas opencv-python
conda install gxx_linux-64
pip install packaging fastparquet

# install with tsinghua
pip install --ignore-installed --upgrade tensorflow -i https://pypi.tuna.tsinghua.edu.cn/simple

# install 1.14
pip install tensorflow-gpu==1.14 -i https://pypi.tuna.tsinghua.edu.cn/simple
# in mac
# use pythonw instead of python
conda install python.app

# export
export PYTHONPATH=$PYTHONPATH:my_models/:my_slim/
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/root/anaconda3/envs/tensorflow/lib/:/root/anaconda3/lib/:/data/anaconda3/lib/

tensorboard

nohup tensorboard --logdir=/path/to/train/dir --host=0.0.0.0 >/dev/null 2>&1&

saved model

# inspect
saved_model_cli show --dir /path/to/model/dir
saved_model_cli show --dir /path/to/model/dir --tag_set serve --signature_def serving_default

# when error
# __new__() got an unexpected keyword argument 'serialized_options'
pip install -U protobuf

serving

# deploy
docker pull tensorflow/serving
docker run -t --rm -p [port]:8501 -v "/path/to/model/with/version/subfolder:/models/[my_model_name]" -e MODEL_NAME=[my_model_name] tensorflow/serving -n [my_model_name]

# rest api
# metadata
curl -X GET http://[ip]:[port]/v1/models/my_model_name/metadata

# predict
curl -X POST \
  http://[ip]:[port]/v1/models/my_model_name/versions/[version]:predict \
  -H 'Content-Type: application/json' \
  -d '{
	"instances":[[0.5026455,0.48942696,0.67499523,0.3115942,...]]
}'

jupyter

# show device
from tensorflow.python.client import device_lib
device_lib.list_local_devices()
# or
import tensorflow as tf
tf.config.experimental.list_physical_devices()
# or
import tensorflow as tf
tf.test.is_gpu_available()

# error
# FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.
pip install "numpy<1.17"

# check gpu

cuda

# version
nvcc --version

# check status
nvidia-smi

# when gpu is not listed, check env
export PATH=/usr/local/cuda-10.0:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda-10.0/lib64:$LD_LIBRARY_PATH

tf.op

tf.unstack

把 Tensor value 按 axis 拆成 list of Tensors

tf.unstack(
    value, num=None, axis=0, name='unstack'
)

tf.transpose

把 Tensor 按 perm 转置

tf.transpose(
    a, perm=None, conjugate=False, name='transpose'
)

tf.scan

遍历 elems 每个 item 的第一维，apply fn which take (acc,cur) as arg，initializer 是 acc 的初始值

tf.scan(
    fn,
    elems,
    initializer
  ...

tf.einsum

element-wise computation

tf.einsum(
    equation,...
)

# Matrix multiplication
einsum('ij,jk->ik', m0, m1)  # output[i,k] = sum_j m0[i,j] * m1[j, k]

# Dot product
einsum('i,i->', u, v)  # output = sum_i u[i]*v[i]

# Outer product
einsum('i,j->ij', u, v)  # output[i,j] = u[i]*v[j]

# Transpose
einsum('ij->ji', m)  # output[j,i] = m[i,j]

# Trace
einsum('ii', m)  # output[j,i] = trace(m) = sum_i m[i, i]

# Batch matrix multiplication
einsum('aij,ajk->aik', s, t)  # out[a,i,k] = sum_j s[a,i,j] * t[a, j, k]

tf.broadcast_to

把 input 按 shape 复制

tf.broadcast_to(
    input, shape, name=None
)

tf.random.truncated_normal

只在 mean +/- 2*stddev 内采样

tf.random.truncated_normal(
    shape, mean=0.0, stddev=1.0, dtype=tf.dtypes.float32, seed=None, name=None
)

tfp

tfp.distributions.RelaxedOneHotCategorical

即 Gumbel-Softmax (Jang et al., 2016) and Concrete (Maddison et al., 2016) distributions

https://github.com/tensorflow/probability/blob/v0.11.0/tensorflow_probability/python/distributions/relaxed_onehot_categorical.py#L398-L547