From 8c07eb1eec7fa25cac4d06db97255cc85ed9dbe1 Mon Sep 17 00:00:00 2001
From: Neta Zmora <neta.zmora@intel.com>
Date: Sun, 17 Feb 2019 10:56:55 +0200
Subject: [PATCH] AMC: added configuration option to set the frequency of
computing a reward
--amc-reward-frequency
Computing the reward requires running the evaluated network on the Test
dataset (or parts of it) and may involve short-term fine-tuning before
the evaluation (depending on the configuration).
Use this new argument to configure the number of steps/iterations between
reward computation.
---
examples/automated_deep_compression/ADC.py | 25 ++++++++++---------
.../automated_deep_compression/automl_args.py | 4 +--
2 files changed, 15 insertions(+), 14 deletions(-)
diff --git a/examples/automated_deep_compression/ADC.py b/examples/automated_deep_compression/ADC.py
index 590acf4..e92f03a 100755
--- a/examples/automated_deep_compression/ADC.py
+++ b/examples/automated_deep_compression/ADC.py
@@ -214,11 +214,11 @@ def do_adc_internal(model, args, optimizer_data, validate_fn, save_checkpoint_fn
amc_cfg = distiller.utils.MutableNamedTuple({
'protocol': args.amc_protocol,
'agent_algo': args.amc_agent_algo,
- 'compute_reward_every_step': args.amc_reward_every_step,
'perform_thinning': perform_thinning,
'num_ft_epochs': num_ft_epochs,
'action_range': action_range,
- 'conv_cnt': conv_cnt})
+ 'conv_cnt': conv_cnt,
+ 'reward_frequency': args.amc_reward_frequency})
#net_wrapper = NetworkWrapper(model, app_args, services)
#return sample_networks(net_wrapper, services)
@@ -239,13 +239,14 @@ def do_adc_internal(model, args, optimizer_data, validate_fn, save_checkpoint_fn
raise ValueError("{} is not supported currently".format(args.amc_protocol))
steps_per_episode = conv_cnt
- amc_cfg.heatup_noise = 0.5
- amc_cfg.initial_training_noise = 0.5
- amc_cfg.training_noise_decay = 0.996 # 0.998
- amc_cfg.num_heatup_epochs = args.amc_heatup_epochs
- amc_cfg.num_training_epochs = args.amc_training_epochs
- training_noise_duration = amc_cfg.num_training_epochs * steps_per_episode
- heatup_duration = amc_cfg.num_heatup_epochs * steps_per_episode
+ if args.amc_agent_algo == "DDPG":
+ amc_cfg.heatup_noise = 0.5
+ amc_cfg.initial_training_noise = 0.5
+ amc_cfg.training_noise_decay = 0.996 # 0.998
+ amc_cfg.num_heatup_epochs = args.amc_heatup_epochs
+ amc_cfg.num_training_epochs = args.amc_training_epochs
+ training_noise_duration = amc_cfg.num_training_epochs * steps_per_episode
+ heatup_duration = amc_cfg.num_heatup_epochs * steps_per_episode
if amc_cfg.agent_algo == "Random-policy":
return random_agent(DistillerWrapperEnvironment(model, app_args, amc_cfg, services))
@@ -603,10 +604,10 @@ class DistillerWrapperEnvironment(gym.Env):
self.episode += 1
else:
observation = self.get_obs()
- reward = 0
- if self.amc_cfg.compute_reward_every_step:
+ if self.amc_cfg.reward_frequency > 0 and self.current_layer_id % self.amc_cfg.reward_frequency == 0:
reward, top1, total_macs, total_nnz = self.compute_reward(False)
-
+ else:
+ reward = 0
self.prev_action = pruning_action
info = {}
return observation, reward, self.episode_is_done(), info
diff --git a/examples/automated_deep_compression/automl_args.py b/examples/automated_deep_compression/automl_args.py
index e15907c..a598189 100755
--- a/examples/automated_deep_compression/automl_args.py
+++ b/examples/automated_deep_compression/automl_args.py
@@ -22,8 +22,8 @@ def add_automl_args(argparser, arch_choices=None, enable_pretrained=False):
help='The number of epochs for heatup/exploration')
group.add_argument('--amc-training-epochs', type=int, default=300,
help='The number of epochs for training/exploitation')
- group.add_argument('--amc-reward-every-step', action='store_true', default=False,
- help='Compute the reward at every step')
+ group.add_argument('--amc-reward-frequency', type=int, default=-1,
+ help='Reward computation frequency (measured in agent steps)')
group.add_argument('--amc-target-density', type=float,
help='Target density of the network we are seeking')
group.add_argument('--amc-agent-algo', choices=["ClippedPPO-continuous",
--
GitLab