[python]调参自动机

多卡并行调参美滋滋，以后终于可以大力调参了

import os
import threading

model_list = ['sst_adv_cnn','sst_adv_rnn','sst5_adv_cnn','sst5_adv_rnn']
action_list = ['no_up','no_down','no_same','no_updownsame']
run_list = []
for model in model_list:
    for action in action_list:
        run_list.append(model+'_'+action+'.sh')
command_list = []
for idx,run in enumerate(run_list):
    command = 'CUDA_VISIBLE_DEVICES='+str(idx%4)+' bash '+run
    print(command)
    command_list.append(command)


class myThread (threading.Thread):
    def __init__(self, command):
        threading.Thread.__init__(self)
        self.cmd = command

    def run(self):
        print("Starting " + self.cmd)
        os.system(self.cmd)
        print("Exiting " + self.cmd)

for idx,command in enumerate(command_list):
    myThread(command).start()

组间并行，组外串行，以16为一组，在4个GPU上执行，每隔三秒检查一次是否这一组的所有程序执行完，如果是的那么就执行下一组。

import os
import threading
import random
import time 
## {0} cnn rnn {1} un down same updownsame
set_num = 16
gpu_num = 4
sleep_time = 3 #second
command_raw = r"""
#!/usr/bin/env sh
set -e

NAME=sst5_adv_cnn_{1}_{2}_{3}_{4}
TRAIN_DIR=./tmp/$NAME
INPUT_DIR=./data/sst5
mkdir $TRAIN_DIR || true

python ./src/sst_action.py \
    --mode=train_adv \
    --action_type={4} \
    --model_type=cnn \
    --action={1} \
    --input_dir=$INPUT_DIR \
    --train_dir=$TRAIN_DIR \
    --save_steps=30 \
    --test_steps=50 \
    --least_freq=2 \
    --num_classes=2 \
    --every=1 \
    --dis_warm_up_step={2} \
    --gene_warm_up_step={3} \
    --max_vocab_size=100000 \
    --embedding_dims=300 \
    --rnn_cell_size=300 \
    --batch_size=64 \
    --learning_rate=0.0001 \
    --generator_learning_rate=0.001 \
    --max_steps=25000 \
    --max_grad_norm=1.0 \
    --num_timesteps=100 \
    --keep_prob_emb=0.6 \
    --keep_prob_dense=0.9 \

python ./src/sst_action.py \
    --mode=test \
    --model_type=cnn \
    --action_type={4} \
    --action={1} \
    --input_dir=$INPUT_DIR \
    --train_dir=$TRAIN_DIR \
    --save_steps=30 \
    --test_steps=50 \
    --least_freq=2 \
    --num_classes=2 \
    --every=1 \
    --dis_warm_up_step={2} \
    --gene_warm_up_step={3} \
    --max_vocab_size=100000 \
    --embedding_dims=300 \
    --rnn_cell_size=300 \
    --batch_size=64 \
    --learning_rate=0.0001 \
    --generator_learning_rate=0.001 \
    --max_steps=25000 \
    --max_grad_norm=1.0 \
    --num_timesteps=100 \
    --keep_prob_emb=0.6 \
    --keep_prob_dense=0.9 \
"""
dis_warm_up_list = [500,800,1000,1200]
gene_warm_up_list = [500,800,1000,1200]
model_list = ['cnn','rnn']
action_list = ['no_up','no_down','no_same','no_updownsame','all']
cuda_list = [0,1,2]
command_list = []

for dis_warm_up in dis_warm_up_list:
    for gene_warm_up in gene_warm_up_list:
            for action in action_list:
                action_type=5 if action=='all' else 4
                command = 'sst5_adv'+'_cnn'+'_'+str(action_type)+'_'+action+'_'+str(dis_warm_up)+'_'+str(gene_warm_up)+'.sh'
                with open(command,'w') as f:
                    f.write(command_raw.format('cnn',action,dis_warm_up,gene_warm_up,action_type))
                command_list.append(command)
run_list = []
print(command_list)
print(len(command_list))
input('check out!')
for idx,command in enumerate(command_list):
    cuda_idx = idx%gpu_num
    run = 'CUDA_VISIBLE_DEVICES='+str(cuda_idx)+' bash '+command
    run_list.append(run)

run_list = [run_list[i:i+set_num] for i in range(0,len(run_list),set_num)]

class myThread (threading.Thread):
    def __init__(self, command):
        threading.Thread.__init__(self)
        self.cmd = command

    def run(self):
        print("Starting " + self.cmd)
        os.system(self.cmd)
        print("Exiting " + self.cmd)
def alldown(thread_list):
    for thread in thread_list:
        if thread.isAlive():
            return False
    return True
for tmp_list in run_list:
    thread_list=[]
    for idx,run in enumerate(tmp_list):
        thread = myThread(run)
        thread_list.append(thread)
        thread.start()
    while True:
        if alldown(thread_list):
            break
        time.sleep(sleep_time)