From 5059419b8e436e773299031beedfd968cb84bff8 Mon Sep 17 00:00:00 2001
From: Neta Zmora <neta.zmora@intel.com>
Date: Wed, 23 Oct 2019 16:42:57 +0300
Subject: [PATCH] Fix AMC notebooks' sample commnand-line examples
As documented in issue #395, some of the command-line examples in the
AMC notebooks are incorrect.
Also, fix some bugs that were introduced with the refactoring of the
low-level pruning API
---
distiller/norms.py | 6 +++--
distiller/pruning/ranked_structures_pruner.py | 24 ++++++++++---------
.../amc/jupyter/amc_plain20.ipynb | 18 +++++++-------
.../amc/jupyter/amc_random.ipynb | 22 ++++++++---------
.../amc/jupyter/amc_resnet20.ipynb | 4 ++--
5 files changed, 39 insertions(+), 35 deletions(-)
diff --git a/distiller/norms.py b/distiller/norms.py
index 46cca86..bafda4e 100644
--- a/distiller/norms.py
+++ b/distiller/norms.py
@@ -152,7 +152,9 @@ def channels_norm(param, norm_fn, group_len=1, length_normalized=False):
Returns:
1D tensor with lp-norms of the groups
"""
- assert param.dim() == 4, "param has invalid dimensions"
+ assert param.dim() in (2, 4), "param has invalid dimensions"
+ if param.dim() == 2:
+ return cols_norm(param, norm_fn, group_len, length_normalized)
param = param.transpose(0, 1).contiguous()
group_size = group_len * np.prod(param.shape[1:])
return generic_norm(param.view(-1, group_size), norm_fn, group_size, length_normalized, dim=1)
@@ -296,7 +298,7 @@ def k_smallest_elems(mags, k, noise):
def rank_channels(param, group_len, magnitude_fn, fraction_to_partition, rounding_fn, noise):
- assert param.dim() == 4, "This ranking is only supported for 4D tensors"
+ assert param.dim() in (2, 4), "This ranking is only supported for 2D and 4D tensors"
n_channels = param.size(1)
n_ch_to_prune = num_structs_to_prune(n_channels, group_len, fraction_to_partition, rounding_fn)
if n_ch_to_prune == 0:
diff --git a/distiller/pruning/ranked_structures_pruner.py b/distiller/pruning/ranked_structures_pruner.py
index 92336dc..1453031 100755
--- a/distiller/pruning/ranked_structures_pruner.py
+++ b/distiller/pruning/ranked_structures_pruner.py
@@ -147,10 +147,14 @@ class LpRankedStructureParameterPruner(_RankedStructureParameterPruner):
model=None, binary_map=None, magnitude_fn=distiller.norms.l1_norm,
noise=0.0, group_size=1, rounding_fn=math.floor):
if binary_map is None:
+
+ if param.dim() == 2:
+ # 2D Linear parameters
+ return LpRankedStructureParameterPruner.rank_and_prune_rows(fraction_to_prune, param, param_name,
+ zeros_mask_dict, model, binary_map,
+ magnitude_fn, group_size)
bottomk_channels, channel_mags = distiller.norms.rank_channels(param, group_size, magnitude_fn,
fraction_to_prune, rounding_fn, noise)
- # bottomk_channels, channel_mags = LpRankedStructureParameterPruner.rank_channels(
- # magnitude_fn, fraction_to_prune, param, group_size, rounding_fn, noise)
if bottomk_channels is None:
# Empty list means that fraction_to_prune is too low to prune anything
return
@@ -689,13 +693,12 @@ class FMReconstructionChannelPruner(_RankedStructureParameterPruner):
if binary_map is None:
op_type = 'conv' if param.dim() == 4 else 'fc'
if op_type == 'conv':
- bottomk_channels, channel_mags = LpRankedStructureParameterPruner.rank_channels(
- magnitude_fn, fraction_to_prune, param, group_size, rounding_fn, noise)
-
+ bottomk_channels, channel_mags = distiller.norms.rank_channels(param, group_size, magnitude_fn,
+ fraction_to_prune, rounding_fn, noise)
else:
- bottomk_channels, channel_mags = LpRankedStructureParameterPruner.rank_cols(
- magnitude_fn, fraction_to_prune, param)
-
+ bottomk_channels, channel_mags = distiller.norms.rank_cols(param, group_size, magnitude_fn,
+ fraction_to_prune, rounding_fn=math.floor,
+ noise=None)
# Todo: this little piece of code can be refactored
if bottomk_channels is None:
# Empty list means that fraction_to_prune is too low to prune anything
@@ -704,7 +707,6 @@ class FMReconstructionChannelPruner(_RankedStructureParameterPruner):
threshold = bottomk_channels[-1]
binary_map = channel_mags.gt(threshold)
-
# These are the indices of channels we want to keep
indices = binary_map.nonzero().squeeze()
if len(indices.shape) == 0:
@@ -757,8 +759,8 @@ class FMReconstructionChannelPruner(_RankedStructureParameterPruner):
if zeros_mask_dict is not None:
binary_map = binary_map.type(param.type())
if op_type == 'conv':
- zeros_mask_dict[param_name].mask = _ = distiller.thresholding.expand_binary_map(param,
- 'Channels', binary_map)
+ zeros_mask_dict[param_name].mask, _ = distiller.thresholding.expand_binary_map(param,
+ 'Channels', binary_map)
msglogger.info("FMReconstructionChannelPruner - param: %s pruned=%.3f goal=%.3f (%d/%d)",
param_name,
distiller.sparsity_ch(zeros_mask_dict[param_name].mask),
diff --git a/examples/auto_compression/amc/jupyter/amc_plain20.ipynb b/examples/auto_compression/amc/jupyter/amc_plain20.ipynb
index e58ec54..d2e2543 100644
--- a/examples/auto_compression/amc/jupyter/amc_plain20.ipynb
+++ b/examples/auto_compression/amc/jupyter/amc_plain20.ipynb
@@ -79,7 +79,7 @@
"\n",
"The command-line is provided below:\n",
" \n",
- " time python3 ../../classifier_compression/multi-run.py ${AMC_EXP_PATH}/plain20-ddpg-private amc.py --arch=plain20_cifar ${CIFAR10_PATH} --resume=checkpoint.plain20_cifar.pth.tar --lr=0.05 --amc --amc-protocol=mac-constrained --amc-action-range 0.05 1.0 --amc-target-density=0.5 -p=50 --etes=0.075 --amc-ft-epochs=0 --amc-prune-pattern=channels --amc-prune-method=fm-reconstruction --amc-agent-algo=DDPG --amc-cfg=auto_compression_channels.yaml --evs=0.5 --etrs=0.5 --amc-rllib=hanlab -j=1\n",
+ " time python3 ../../classifier_compression/multi-run.py ${AMC_EXP_PATH}/plain20-ddpg-private amc.py --arch=plain20_cifar ${CIFAR10_PATH} --resume=${CHECKPOINTS_PATH}/checkpoint.plain20_cifar.pth.tar --lr=0.05 --amc-protocol=mac-constrained --amc-action-range 0.05 1.0 --amc-target-density=0.5 -p=50 --etes=0.075 --amc-ft-epochs=0 --amc-prune-pattern=channels --amc-prune-method=fm-reconstruction --amc-agent-algo=DDPG --amc-cfg=auto_compression_channels.yaml --evs=0.5 --etrs=0.5 --amc-rllib=hanlab -j=1\n",
" \n",
"Each colored line represents one experiment execution instance. We plot the search-Top1 score of discovered networks as the RL-based AMC system learns to find better compressed networks. You might be impressed by:\n",
"* The variability in behavior, which is typical for RL algorithms.\n",
@@ -133,12 +133,12 @@
"source": [
"Fine-tune for 2 epochs (using 16 processes): notice the low LR which we use because we only FT for 2 episodes.\n",
"\n",
- " time python parallel-finetune.py --scan-dir=${AMC_EXP_PATH}/plain20-ddpg-private/2019.07.24-174636 --arch=plain20_cifar --lr=0.001 --vs=0 -p=50 --compress=plain20_fine_tune.yaml ${CIFAR10_PATH} -j=1 --epochs=2 --output-csv=ft_2epoch_results.csv --processes=16\n",
+ " time python parallel-finetune.py --scan-dir=${AMC_EXP_PATH}/plain20-ddpg-private/2019.07.24-174636 --arch=plain20_cifar --lr=0.001 --vs=0 -p=50 --compress=../plain20_fine_tune.yaml ${CIFAR10_PATH} -j=1 --epochs=2 --output-csv=ft_2epoch_results.csv --processes=16\n",
"\n",
"\n",
"Fine-tune for 60 epochs (using 16 processes): notice the x100 higher LR.\n",
"\n",
- " time python parallel-finetune.py --scan-dir=${AMC_EXP_PATH}/plain20-ddpg-private/2019.07.24-174636 --arch=plain20_cifar --lr=0.1 --vs=0 -p=50 --compress=plain20_fine_tune.yaml ${CIFAR10_PATH} -j=1 --epochs=60 --output-csv=ft_60epoch_results.csv --processes=16"
+ " time python parallel-finetune.py --scan-dir=${AMC_EXP_PATH}/plain20-ddpg-private/2019.07.24-174636 --arch=plain20_cifar --lr=0.1 --vs=0 -p=50 --compress=../plain20_fine_tune.yaml ${CIFAR10_PATH} -j=1 --epochs=60 --output-csv=ft_60epoch_results.csv --processes=16"
]
},
{
@@ -286,7 +286,7 @@
"Results are interesting and encouraging as there is learning. However, this is less sample-efficient compared to DDPG, and therefore takes longer.\n",
"There might be some room to tune the PPO hyper-params to generate better results.\n",
"\n",
- " time python3 ../../classifier_compression/multi-run.py ${AMC_EXP_PATH}/plain20-ppo-coach amc.py --arch=plain20_cifar ${CIFAR10_PATH} --resume=checkpoint.plain20_cifar.pth.tar --lr=0.05 --amc --amc-protocol=mac-constrained --amc-action-range 0.05 1.0 --amc-target-density=0.5 -p=50 --etes=0.075 --amc-ft-epochs=0 --amc-prune-pattern=channels --amc-prune-method=fm-reconstruction --amc-agent-algo=ClippedPPO-continuous --amc-cfg=auto_compression_channels.yaml --evs=0.5 --etrs=0.5 --amc-rllib=coach -j=1"
+ " time python3 ../../classifier_compression/multi-run.py ${AMC_EXP_PATH}/plain20-ppo-coach amc.py --arch=plain20_cifar ${CIFAR10_PATH} --resume=${CHECKPOINTS_PATH}/checkpoint.plain20_cifar.pth.tar --lr=0.05 --amc-protocol=mac-constrained --amc-action-range 0.05 1.0 --amc-target-density=0.5 -p=50 --etes=0.075 --amc-ft-epochs=0 --amc-prune-pattern=channels --amc-prune-method=fm-reconstruction --amc-agent-algo=ClippedPPO-continuous --amc-cfg=auto_compression_channels.yaml --evs=0.5 --etrs=0.5 --amc-rllib=coach -j=1"
]
},
{
@@ -321,7 +321,7 @@
"cell_type": "markdown",
"metadata": {},
"source": [
- " time python parallel-finetune.py --scan-dir=${AMC_EXP_PATH}/plain20-ppo-coach/2019.07.28-012356 --arch=plain20_cifar --lr=0.1 --vs=0 -p=50 --compress=plain20_fine_tune.yaml ${CIFAR10_PATH} -j=1 --epochs=60 --output-csv=ft_60epoch_results.csv --processes=16"
+ " time python3 parallel-finetune.py --scan-dir=${AMC_EXP_PATH}/plain20-ppo-coach/2019.07.28-012356 --arch=plain20_cifar --lr=0.1 --vs=0 -p=50 --compress=../plain20_fine_tune.yaml ${CIFAR10_PATH} -j=1 --epochs=60 --output-csv=ft_60epoch_results.csv --processes=16"
]
},
{
@@ -358,7 +358,7 @@
"source": [
"## Using a different reward function\n",
"\n",
- " time python3 ../../classifier_compression/multi-run.py ${AMC_EXP_PATH}/plain20-ddpg-private-punish compress_classifier.py --arch=plain20_cifar ${CIFAR10_PATH} --resume=checkpoint.plain20_cifar.pth.tar --lr=0.05 --amc-protocol=punish-agent --amc-action-range 0.05 1.0 --amc-target-density=0.5 -p=50 --etes=0.075 --amc-ft-epochs=0 --amc-prune-pattern=channels --amc-prune-method=fm-reconstruction --amc-agent-algo=DDPG --amc-cfg=../automated_deep_compression/auto_compression_channels.yaml --evs=0.5 --etrs=0.5 --amc-rllib=hanlab -j=1"
+ " time python3 ../../classifier_compression/multi-run.py ${AMC_EXP_PATH}/plain20-ddpg-private-punish amc.py --arch=plain20_cifar ${CIFAR10_PATH} --resume=${CHECKPOINTS_PATH}/checkpoint.plain20_cifar.pth.tar --lr=0.05 --amc-protocol=punish-agent --amc-action-range 0.05 1.0 --amc-target-density=0.5 -p=50 --etes=0.075 --amc-ft-epochs=0 --amc-prune-pattern=channels --amc-prune-method=fm-reconstruction --amc-agent-algo=DDPG --amc-cfg=auto_compression_channels.yaml --evs=0.5 --etrs=0.5 --amc-rllib=hanlab -j=1"
]
},
{
@@ -395,7 +395,7 @@
"cell_type": "markdown",
"metadata": {},
"source": [
- " time python parallel-finetune.py --scan-dir=${AMC_EXP_PATH}/plain20-ddpg-private-punish/2019.07.29-171102 --arch=plain20_cifar --lr=0.1 --vs=0 -p=50 --compress=plain20_fine_tune.yaml ${CIFAR10_PATH} -j=1 --epochs=60 --output-csv=ft_60epoch_results.csv --processes=16"
+ " time python parallel-finetune.py --scan-dir=${AMC_EXP_PATH}/plain20-ddpg-private-punish/2019.07.29-171102 --arch=plain20_cifar --lr=0.1 --vs=0 -p=50 --compress=../plain20_fine_tune.yaml ${CIFAR10_PATH} -j=1 --epochs=60 --output-csv=ft_60epoch_results.csv --processes=16"
]
},
{
@@ -450,7 +450,7 @@
"\n",
" time python3 ../../classifier_compression/multi-run.py ${AMC_EXP_PATH}/plain20-ddpg-private amc.py --arch=plain20_cifar ${CIFAR10_PATH} --resume=${CHECKPOINTS_PATH}/checkpoint.plain20_cifar.pth.tar --lr=0.05 --amc-protocol=mac-constrained --amc-action-range 0.05 1.0 --amc-target-density=0.5 -p=50 --etes=0.075 --amc-ft-epochs=0 --amc-prune-pattern=channels --amc-prune-method=fm-reconstruction --amc-agent-algo=DDPG --amc-cfg=auto_compression_channels.yaml --evs=1.0 --etrs=0.01 --amc-rllib=hanlab -j=1\n",
" \n",
- " time python parallel-finetune.py --scan-dir=${AMC_EXP_PATH}/plain20-ddpg-private/2019.08.01-181040 --arch=plain20_cifar --lr=0.1 --vs=0 -p=50 --compress=plain20_fine_tune.yaml ${CIFAR10_PATH} -j=1 --epochs=60 --output-csv=ft_60epoch_results.csv --processes=16"
+ " time python parallel-finetune.py --scan-dir=${AMC_EXP_PATH}/plain20-ddpg-private/2019.08.01-181040 --arch=plain20_cifar --lr=0.1 --vs=0 -p=50 --compress=../plain20_fine_tune.yaml ${CIFAR10_PATH} -j=1 --epochs=60 --output-csv=ft_60epoch_results.csv --processes=16"
]
},
{
@@ -493,7 +493,7 @@
"\n",
" time python3 ../../classifier_compression/multi-run.py ${AMC_EXP_PATH}/plain20-ddpg-private amc.py --arch=plain20_cifar ${CIFAR10_PATH} --resume=${CHECKPOINTS_PATH}/checkpoint.plain20_cifar.pth.tar --lr=0.05 --amc-protocol=mac-constrained --amc-action-range 0.05 1.0 --amc-target-density=0.5 -p=50 --etes=0.075 --amc-ft-epochs=0 --amc-prune-pattern=channels --amc-prune-method=fm-reconstruction --amc-agent-algo=DDPG --amc-cfg=auto_compression_channels.yaml --evs=0.5 --etrs=0.01 --amc-rllib=hanlab -j=1 --amc-save-chkpts\n",
" \n",
- " time python ../../classifier_compression/parallel-finetune.py --scan-dir=${AMC_EXP_PATH}/plain20-ddpg-private/2019.08.03-000628 --arch=plain20_cifar --lr=0.1 --vs=0 -p=50 --compress=../plain20_fine_tune.yaml ${CIFAR10_PATH} -j=1 --epochs=60 --output-csv=ft_60epoch_results.csv --processes=16 --top-performing-chkpts"
+ " time python3 ../../classifier_compression/parallel-finetune.py --scan-dir=${AMC_EXP_PATH}/plain20-ddpg-private/2019.08.03-000628 --arch=plain20_cifar --lr=0.1 --vs=0 -p=50 --compress=../plain20_fine_tune.yaml ${CIFAR10_PATH} -j=1 --epochs=60 --output-csv=ft_60epoch_results.csv --processes=16 --top-performing-chkpts"
]
},
{
diff --git a/examples/auto_compression/amc/jupyter/amc_random.ipynb b/examples/auto_compression/amc/jupyter/amc_random.ipynb
index 0de236e..259d057 100644
--- a/examples/auto_compression/amc/jupyter/amc_random.ipynb
+++ b/examples/auto_compression/amc/jupyter/amc_random.ipynb
@@ -60,19 +60,19 @@
"source": [
"**Test**\n",
"\n",
- " time python3 ../../classifier_compression/multi-run.py ${AMC_EXP_PATH}/plain20-random-reconstruction_nondeterministic amc.py --arch=plain20_cifar ${CIFAR10_PATH} --resume=checkpoint.plain20_cifar.pth.tar --lr=0.05 --amc-protocol=mac-constrained --amc-action-range 0.05 1.0 --amc-target-density=0.5 -p=50 --etes=0.075 --amc-ft-epochs=0 --amc-prune-pattern=channels --amc-prune-method=fm-reconstruction --amc-agent-algo=Random-policy --amc-cfg=../automated_deep_compression/auto_compression_channels.yaml --evs=0.5 --etrs=0.5 --amc-rllib=random -j=1\n",
+ " time python3 ../../classifier_compression/multi-run.py ${AMC_EXP_PATH}/plain20-random-reconstruction_nondeterministic amc.py --arch=plain20_cifar ${CIFAR10_PATH} --resume=${CHECKPOINTS_PATH}/checkpoint.plain20_cifar.pth.tar --lr=0.05 --amc-protocol=mac-constrained --amc-action-range 0.05 1.0 --amc-target-density=0.5 -p=50 --etes=0.075 --amc-ft-epochs=0 --amc-prune-pattern=channels --amc-prune-method=fm-reconstruction --amc-agent-algo=Random-policy --amc-cfg=auto_compression_channels.yaml --evs=0.5 --etrs=0.5 --amc-rllib=random -j=1\n",
"\n",
"==> experiments/plain20-random-reconstruction/2019.07.22-120953/\n",
"\n",
- " time python parallel-finetune.py --scan-dir=${AMC_EXP_PATH}/plain20-random-reconstruction_nondeterministic/2019.07.23-124600 --arch=plain20_cifar --lr=0.005 --vs=0 -p=50 --epochs=60 --compress=../automated_deep_compression/fine_tune.yaml ${CIFAR10_PATH} -j=1 --deterministic --ft-epochs=1 --output-csv=ft_1epoch_results.csv\n",
+ " time python parallel-finetune.py --scan-dir=${AMC_EXP_PATH}/plain20-random-reconstruction_nondeterministic/2019.07.23-124600 --arch=plain20_cifar --lr=0.005 --vs=0 -p=50 --epochs=60 --compress=../plain20_fine_tune.yaml ${CIFAR10_PATH} -j=1 --deterministic --epoch=1 --output-csv=ft_1epoch_results.csv\n",
" \n",
"**Baseline**\n",
"\n",
- " time python3 ../../classifier_compression/multi-run.py ${AMC_EXP_PATH}/plain20-random-l1_rank amc.py --arch=plain20_cifar ${CIFAR10_PATH} --resume=checkpoint.plain20_cifar.pth.tar --lr=0.05 --amc-protocol=mac-constrained --amc-action-range 0.05 1.0 --amc-target-density=0.5 -p=50 --etes=0.075 --amc-ft-epochs=0 --amc-prune-pattern=channels --amc-prune-method=l1-rank --amc-agent-algo=Random-policy --amc-cfg=../automated_deep_compression/auto_compression_channels.yaml --evs=0.5 --etrs=0.5 --amc-rllib=random -j=1\n",
+ " time python3 ../../classifier_compression/multi-run.py ${AMC_EXP_PATH}/plain20-random-l1_rank amc.py --arch=plain20_cifar ${CIFAR10_PATH} --resume=${CHECKPOINTS_PATH}/checkpoint.plain20_cifar.pth.tar --lr=0.05 --amc-protocol=mac-constrained --amc-action-range 0.05 1.0 --amc-target-density=0.5 -p=50 --etes=0.075 --amc-ft-epochs=0 --amc-prune-pattern=channels --amc-prune-method=l1-rank --amc-agent-algo=Random-policy --amc-cfg=auto_compression_channels.yaml --evs=0.5 --etrs=0.5 --amc-rllib=random -j=1\n",
"\n",
"==> classifier_compression/experiments/plain20-random-l1_rank/2019.07.21-004045\n",
"\n",
- " time python parallel-finetune.py --scan-dir=${AMC_EXP_PATH}/plain20-random-l1_rank/2019.07.21-004045/ --arch=plain20_cifar --lr=0.005 --vs=0 -p=50 --epochs=60 --compress=../automated_deep_compression/fine_tune.yaml ${CIFAR10_PATH} -j=1 --deterministic --ft-epochs=1 --output-csv=ft_1epoch_results.csv\n",
+ " time python parallel-finetune.py --scan-dir=${AMC_EXP_PATH}/plain20-random-l1_rank/2019.07.21-004045/ --arch=plain20_cifar --lr=0.005 --vs=0 -p=50 --epochs=60 --compress=../plain20_fine_tune.yaml ${CIFAR10_PATH} -j=1 --deterministic --epoch=1 --output-csv=ft_1epoch_results.csv\n",
"\n",
"==> classifier_compression/experiments/plain20-random-l1_rank/2019.07.21-004045/ft_1epoch_results.csv"
]
@@ -165,7 +165,7 @@
" NOTE: we fine-tune _AFTER_ the RL search is done, so this FT process does not help the agent. And, in any case, we are using a random agent.\n",
" \n",
"<code>\n",
- "time python parallel-finetune.py --scan-dir=${AMC_EXP_PATH}/plain20-random-reconstruction_nondeterministic/2019.07.23-124600 --arch=plain20_cifar --lr=0.005 --vs=0 -p=50 --epochs=60 --compress=plain20_fine_tune.yaml ${CIFAR10_PATH} -j=1 --deterministic --ft-epochs=20 --output-csv=ft_20epoch_results.csv\n",
+ "time python parallel-finetune.py --scan-dir=${AMC_EXP_PATH}/plain20-random-reconstruction_nondeterministic/2019.07.23-124600 --arch=plain20_cifar --lr=0.005 --vs=0 -p=50 --epochs=60 --compress=../plain20_fine_tune.yaml ${CIFAR10_PATH} -j=1 --deterministic --epoch=20 --output-csv=ft_20epoch_results.csv\n",
"</code>\n",
"\n",
"#### Answer:\n",
@@ -280,11 +280,11 @@
"* In other words, how many fine-tuning epochs do we need to perform during the RL search to get a cleaner, less noisy (i.e. more stable) search-Top1 (used in the multi-objective reward signal)?\n",
"\n",
"\n",
- " time python parallel-finetune.py --scan-dir=${AMC_EXP_PATH}/plain20-random-l1_rank/2019.07.21-004045/ --arch=plain20_cifar --lr=0.005 --vs=0 -p=50 --epochs=60 --compress=../automated_deep_compression/fine_tune.yaml ${CIFAR10_PATH} -j=1 --deterministic --ft-epochs=3 --output-csv=ft_3epoch_results.csv\n",
+ " time python parallel-finetune.py --scan-dir=${AMC_EXP_PATH}/plain20-random-l1_rank/2019.07.21-004045/ --arch=plain20_cifar --lr=0.005 --vs=0 -p=50 --epochs=60 --compress=../plain20_fine_tune.yaml ${CIFAR10_PATH} -j=1 --deterministic --epoch=3 --output-csv=ft_3epoch_results.csv\n",
"\n",
- " time python para,,e,-finetune.py --scan-dir=${AMC_EXP_PATH}/plain20-random-l1_rank/2019.07.21-004045/ --arch=plain20_cifar --lr=0.005 --vs=0 -p=50 --epochs=60 --compress=../automated_deep_compression/fine_tune.yaml ${CIFAR10_PATH} -j=1 --deterministic --ft-epochs=6 --output-csv=ft_6epoch_results.csv\n",
+ " time python parallel-finetune.py --scan-dir=${AMC_EXP_PATH}/plain20-random-l1_rank/2019.07.21-004045/ --arch=plain20_cifar --lr=0.005 --vs=0 -p=50 --epochs=60 --compress=../plain20_fine_tune.yaml ${CIFAR10_PATH} -j=1 --deterministic --epoch=6 --output-csv=ft_6epoch_results.csv\n",
"\n",
- " time python parallel-finetune.py --scan-dir=${AMC_EXP_PATH}/plain20-random-l1_rank/2019.07.21-004045/ --arch=plain20_cifar --lr=0.005 --vs=0 -p=50 --epochs=60 --compress=../automated_deep_compression/fine_tune.yaml ${CIFAR10_PATH} -j=1 --deterministic --ft-epochs=20 --output-csv=ft_20epoch_results.csv"
+ " time python parallel-finetune.py --scan-dir=${AMC_EXP_PATH}/plain20-random-l1_rank/2019.07.21-004045/ --arch=plain20_cifar --lr=0.005 --vs=0 -p=50 --epochs=60 --compress=../plain20_fine_tune.yaml ${CIFAR10_PATH} -j=1 --deterministic --epoch=20 --output-csv=ft_20epoch_results.csv"
]
},
{
@@ -441,13 +441,13 @@
"pycharm": {
"stem_cell": {
"cell_type": "raw",
+ "source": [],
"metadata": {
"collapsed": false
- },
- "source": []
+ }
}
}
},
"nbformat": 4,
"nbformat_minor": 2
-}
+}
\ No newline at end of file
diff --git a/examples/auto_compression/amc/jupyter/amc_resnet20.ipynb b/examples/auto_compression/amc/jupyter/amc_resnet20.ipynb
index 4232c7b..29477ab 100644
--- a/examples/auto_compression/amc/jupyter/amc_resnet20.ipynb
+++ b/examples/auto_compression/amc/jupyter/amc_resnet20.ipynb
@@ -79,7 +79,7 @@
"\n",
"The command-line is provided below:\n",
" \n",
- " time python3 ../../classifier_compression/multi-run.py ${AMC_EXP_PATH}/resnet20-ddpg-private amc.py --arch=resnet20_cifar ${CIFAR10_PATH} --resume=../ssl/checkpoints/checkpoint_trained_dense.pth.tar --amc-protocol=mac-constrained --amc-action-range 0.05 1.0 --amc-target-density=0.5 -p=50 --etes=0.075 --amc-ft-epochs=0 --amc-prune-pattern=channels --amc-prune-method=fm-reconstruction --amc-agent-algo=DDPG --amc-cfg=auto_compression_channels.yaml --amc-rllib=hanlab -j=1\n",
+ " time python3 ../../classifier_compression/multi-run.py ${AMC_EXP_PATH}/resnet20-ddpg-private amc.py --arch=resnet20_cifar ${CIFAR10_PATH} --resume=../../ssl/checkpoints/checkpoint_trained_dense.pth.tar --amc-protocol=mac-constrained --amc-action-range 0.05 1.0 --amc-target-density=0.5 -p=50 --etes=0.075 --amc-ft-epochs=0 --amc-prune-pattern=channels --amc-prune-method=fm-reconstruction --amc-agent-algo=DDPG --amc-cfg=auto_compression_channels.yaml --amc-rllib=hanlab -j=1\n",
" \n",
"Each colored line represents one experiment execution instance. We plot the search-Top1 score of discovered networks as the RL-based AMC system learns to find better compressed networks. You might be impressed by:\n",
"* The variability in behavior, which is typical for RL algorithms.\n",
@@ -120,7 +120,7 @@
"\n",
"Fine-tune for 60 epochs (using 16 processes): **notice the large LR**.\n",
"\n",
- " time python parallel-finetune.py --scan-dir=${AMC_EXP_PATH}/resnet20-ddpg-private/2019.07.30-015225 --arch=resnet20_cifar --lr=0.1 --vs=0 -p=50 --compress=../automated_deep_compression/fine_tune.yaml ${CIFAR10_PATH} -j=1 --epochs=60 --output-csv=ft_60epoch_results.csv --processes=16"
+ " time python parallel-finetune.py --scan-dir=${AMC_EXP_PATH}/resnet20-ddpg-private/2019.07.30-015225 --arch=resnet20_cifar --lr=0.1 --vs=0 -p=50 --compress=../plain20_fine_tune.yaml ${CIFAR10_PATH} -j=1 --epochs=60 --output-csv=ft_60epoch_results.csv --processes=16"
]
},
{
--
GitLab