毕业设计，用TensorFlow复现DeepRM.zip

import numpy as np import math import matplotlib.pyplot as plt # import theano import parameters class Env: def __init__(self, pa, nw_len_seqs=None, nw_size_seqs=None, seed=42, render=False, repre='image', end='no_new_job'): self.pa = pa self.render = render self.repre = repre # image or compact representation self.end = end # termination type, 'no_new_job' or 'all_done' self.nw_dist = pa.dist.bi_model_dist self.curr_time = 0 # set up random seed if self.pa.unseen: np.random.seed(314159) else: np.random.seed(seed) if nw_len_seqs is None or nw_size_seqs is None: # generate new work self.nw_len_seqs, self.nw_size_seqs = \ self.generate_sequence_work(self.pa.simu_len * self.pa.num_ex) self.workload = np.zeros(pa.num_res) for i in range(pa.num_res): self.workload[i] = \ np.sum(self.nw_size_seqs[:, i] * self.nw_len_seqs) / \ float(pa.res_slot) / \ float(len(self.nw_len_seqs)) print("Load on # " + str(i) + " resource dimension is " + str(self.workload[i])) self.nw_len_seqs = np.reshape(self.nw_len_seqs, [self.pa.num_ex, self.pa.simu_len]) self.nw_size_seqs = np.reshape(self.nw_size_seqs, [self.pa.num_ex, self.pa.simu_len, self.pa.num_res]) else: self.nw_len_seqs = nw_len_seqs self.nw_size_seqs = nw_size_seqs self.seq_no = 0 # which example sequence self.seq_idx = 0 # index in that sequence # initialize system self.machine = Machine(pa) self.job_slot = JobSlot(pa) self.job_backlog = JobBacklog(pa) self.job_record = JobRecord() self.extra_info = ExtraInfo(pa) def generate_sequence_work(self, simu_len): nw_len_seq = np.zeros(simu_len, dtype=int) nw_size_seq = np.zeros((simu_len, self.pa.num_res), dtype=int) for i in range(simu_len): if np.random.rand() < self.pa.new_job_rate: # a new job comes nw_len_seq[i], nw_size_seq[i, :] = self.nw_dist() return nw_len_seq, nw_size_seq def get_new_job_from_seq(self, seq_no, seq_idx): new_job = Job(res_vec=self.nw_size_seqs[seq_no, seq_idx, :], job_len=self.nw_len_seqs[seq_no, seq_idx], job_id=len(self.job_record.record), enter_time=self.curr_time) return new_job def observe(self): if self.repre == 'image': backlog_width = int(math.ceil(self.pa.backlog_size / float(self.pa.time_horizon))) image_repr = np.zeros((self.pa.network_input_height, self.pa.network_input_width)) ir_pt = 0 for i in range(self.pa.num_res): image_repr[:, ir_pt: ir_pt + self.pa.res_slot] = self.machine.canvas[i, :, :] ir_pt += self.pa.res_slot for j in range(self.pa.num_nw): if self.job_slot.slot[j] is not None: # fill in a block of work image_repr[: self.job_slot.slot[j].len, ir_pt: ir_pt + self.job_slot.slot[j].res_vec[i]] = 1 ir_pt += self.pa.max_job_size image_repr[: int(self.job_backlog.curr_size / backlog_width), ir_pt: ir_pt + backlog_width] = 1 if self.job_backlog.curr_size % backlog_width > 0: image_repr[int(self.job_backlog.curr_size / backlog_width), ir_pt: ir_pt + self.job_backlog.curr_size % backlog_width] = 1 ir_pt += backlog_width image_repr[:, ir_pt: ir_pt + 1] = self.extra_info.time_since_last_new_job / \ float(self.extra_info.max_tracking_time_since_last_job) ir_pt += 1 assert ir_pt == image_repr.shape[1] return image_repr.ravel()[np.newaxis, :] #return image_repr def plot_state(self): plt.figure("screen", figsize=(20, 5)) skip_row = 0 for i in range(self.pa.num_res): plt.subplot(self.pa.num_res, 1 + self.pa.num_nw + 1, # first +1 for current work, last +1 for backlog queue i * (self.pa.num_nw + 1) + skip_row + 1) # plot the backlog at the end, +1 to avoid 0 plt.imshow(self.machine.canvas[i, :, :], interpolation='nearest', vmax=1) for j in range(self.pa.num_nw): job_slot = np.zeros((self.pa.time_horizon, self.pa.max_job_size)) if self.job_slot.slot[j] is not None: # fill in a block of work job_slot[: self.job_slot.slot[j].len, :self.job_slot.slot[j].res_vec[i]] = 1 plt.subplot(self.pa.num_res, 1 + self.pa.num_nw + 1, # first +1 for current work, last +1 for backlog queue 1 + i * (self.pa.num_nw + 1) + j + skip_row + 1) # plot the backlog at the end, +1 to avoid 0 plt.imshow(job_slot, interpolation='nearest', vmax=1) if j == self.pa.num_nw - 1: skip_row += 1 skip_row -= 1 backlog_width = int(math.ceil(self.pa.backlog_size / float(self.pa.time_horizon))) backlog = np.zeros((self.pa.time_horizon, backlog_width)) backlog[: self.job_backlog.curr_size / backlog_width, : backlog_width] = 1 backlog[self.job_backlog.curr_size / backlog_width, : self.job_backlog.curr_size % backlog_width] = 1 plt.subplot(self.pa.num_res, 1 + self.pa.num_nw + 1, # first +1 for current work, last +1 for backlog queue self.pa.num_nw + 1 + 1) plt.imshow(backlog, interpolation='nearest', vmax=1) plt.subplot(self.pa.num_res, 1 + self.pa.num_nw + 1, # first +1 for current work, last +1 for backlog queue self.pa.num_res * (self.pa.num_nw + 1) + skip_row + 1) # plot the backlog at the end, +1 to avoid 0 extra_info = np.ones((self.pa.time_horizon, 1)) * \ self.extra_info.time_since_last_new_job / \ float(self.extra_info.max_tracking_time_since_last_job) plt.imshow(extra_info, interpolation='nearest', vmax=1) plt.show() # manual # plt.pause(0.01) # automatic def get_reward(self): reward = 0 for j in self.machine.running_job: reward += self.pa.delay_penalty / float(j.len) for j in self.job_slot.slot: if j is not None: reward += self.pa.hold_penalty / float(j.len) for j in self.job_backlog.backlog: if j is not None: reward += self.pa.dismiss_penalty / float(j.len) return reward def step(self, a, repeat=False): status = None done = False reward = 0 info = None if a == self.pa.num_nw: # explicit void action status = 'MoveOn' elif self.job_slot.slot[a] is None: # implicit void action #if self.seq_idx >= self.pa.simu_len and \ #len(self.machine.running_job) > 0 and \ #all(s is None for s in self.job_backlog.backlog): #ob, reward, done, info = self.step(a + 1, repeat=True) #return ob, reward, done, info #else: status = 'MoveOn' else: allocated = self.machine.allocate_job(self.job_slot.slot[a], self.curr_time) if not allocated: # implicit void action status = 'MoveOn' else: status = 'Allocate' if status == 'MoveOn': self.curr_time +=