def main():
# Initialize experiment options and parameters
suite = MNISTSparseExperiment()
suite.parse_opt()
suite.parse_cfg()
experiments = suite.options.experiments or suite.cfgparser.sections()
pool = multiprocessing.Pool()
for expName in experiments:
path = suite.get_exp(expName)[0]
results = suite.get_exps(path=path)
# Build argument list for multiprocessing pool
args = []
for exp in results:
params = suite.get_params(exp)
for i, minWeight in enumerate(np.linspace(0.0, 0.1, 21)):
args.append((exp, params, minWeight, -1, i))
args = np.array(args)
# Analyze weight pruning alone. No dutycycle pruning
args[:, 3] = -1 # set minDutycycle to -1 for all experiments
run(pool, expName, "Weight Pruning", args)
# Analyze dutycycle pruning units with dutycycle below 5% from target density
args[:, 3] = 0.05 # set minDutycycle to 5% for all experiments
run(pool, expName, "Dutycycle Pruning (5%)", args)
# Analyze dutycycle pruning units with dutycycle below 10% from target density
args[:, 3] = 0.10 # set minDutycycle to 10% for all experiments
run(pool, expName, "Dutycycle Pruning (10%)", args)
def main():
# Set fixed seed. Only works on cpu
random.seed(42)
np.random.seed(42)
torch.manual_seed(42)
dataset = create_union_mnist_dataset()
# Load experiment
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
suite = MNISTSparseExperiment()
suite.parse_opt()
suite.parse_cfg()
experiments = suite.options.experiments or suite.cfgparser.sections()
table = {}
progress = tqdm(experiments)
for expName in progress:
progress.set_description(expName)
path = suite.get_exp(expName)[0]
results = suite.get_exps(path=path)
for exp in results:
model_file = os.path.join(exp, "model.pt")
if os.path.exists(model_file):
model = torch.load(model_file, map_location=device)
params = suite.get_params(exp)
test_loader = torch.utils.data.DataLoader(
dataset,
shuffle=True,
batch_size=params["test_batch_size"])
table[params['name']] = evaluate(model=model,
loader=test_loader,
device=device)
# Random model
test_loader = torch.utils.data.DataLoader(dataset,
shuffle=True,
batch_size=4)
table["random"] = evaluate(model=random_model,
loader=test_loader,
device=device)
# Save results
df = pd.DataFrame.from_dict(table)
df.to_csv("union_experiment_results.csv")
print(tabulate(df, tablefmt='fancy_grid', headers='keys',
numalign="right"))
# Plot first 100 images in the dataset
fig = plt.figure(figsize=(10, 10))
for i in range(100):
ax = fig.add_subplot(10, 10, i + 1)
ax.set_axis_off()
img, label = dataset[i]
ax.imshow(img.numpy().reshape((28, 28)), cmap='gray')
ax.set_title(str(label.numpy()))
plt.tight_layout()
plt.savefig("union_experiment_sample_images.png")
plt.close()
def main():
# Initialize experiment options and parameters
suite = MNISTSparseExperiment()
suite.parse_opt()
suite.parse_cfg()
experiments = suite.options.experiments or suite.cfgparser.sections()
pool = multiprocessing.Pool()
for expName in experiments:
path = suite.get_exp(expName)[0]
results = suite.get_exps(path=path)
# Build argument list for multiprocessing pool
args = []
for exp in results:
params = suite.get_params(exp)
for i, minWeight in enumerate(np.linspace(0.0, 0.1, 21)):
args.append((exp, params, minWeight, -1, i))
args = np.array(args)
# Analyze weight pruning alone. No dutycycle pruning
args[:, 3] = -1 # set minDutycycle to -1 for all experiments
run(pool, expName, "weight_prunning", args)
# Analyze dutycycle pruning units with dutycycle below 5% from target density
args[:, 3] = 0.05 # set minDutycycle to 5% for all experiments
run(pool, expName, "dutycycle_pruning", args)
testScores = np.zeros(numExps)
noiseScores = np.zeros(numExps)
for i, scoresForRepetition in enumerate(
zip(results["testerror"], results["totalCorrect"])):
maxTestScore, bestEpoch, maxNoiseScore = bestScore(scoresForRepetition)
testScores[i] = maxTestScore
noiseScores[i] = maxNoiseScore
return {
"test_score": (testScores.mean(), testScores.std()),
"noise_score": (noiseScores.mean(), noiseScores.std())
}
if __name__ == '__main__':
suite = MNISTSparseExperiment()
suite.parse_opt()
suite.parse_cfg()
experiments = suite.options.experiments or suite.cfgparser.sections()
testScoresTable = [["Network", "Test Score", "Noise Score"]]
for expName in experiments:
results = suite.get_exp(expName)
for exp in results:
if not os.path.exists(exp):
continue
errorBars = getErrorBars(exp, suite)
test_score = u"{0:.2f} ± {1:.2f}".format(*errorBars["test_score"])
noise_score = u"{0:,.0f} ± {1:.2f}".format(
*errorBars["noise_score"])
#
# This program is free software: you can redistribute it and/or modify
# it under the terms of the GNU Affero Public License version 3 as
# published by the Free Software Foundation.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
# See the GNU Affero Public License for more details.
#
# You should have received a copy of the GNU Affero Public License
# along with this program. If not, see http://www.gnu.org/licenses.
#
# http://numenta.org/licenses/
# ----------------------------------------------------------------------
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
from htmresearch.frameworks.pytorch.mnist_sparse_experiment import \
MNISTSparseExperiment
# Run from htmresearch:
# domino run projects/sdr_paper/pytorch_experiments/run_experiment.py -c projects/sdr_paper/pytorch_experiments/experiments.cfg
# OR
# python projects/sdr_paper/pytorch_experiments/run_experiment.py -c projects/sdr_paper/pytorch_experiments/experiments.cfg -d
if __name__ == '__main__':
suite = MNISTSparseExperiment()
suite.start()
noiseScores[i] = maxTotalAccuracy
print("")
print("Experiment:", expPath, "Number of sub-experiments", len(exps))
print("test score mean and standard deviation:", testScores.mean(),
testScores.std())
print("noise score mean and standard deviation:", noiseScores.mean(),
noiseScores.std())
# To run it from htmresearch top level:
# python projects/sdr_paper/pytorch_experiments/analyze_experiment.py -c projects/sdr_paper/pytorch_experiments/experiments.cfg
if __name__ == '__main__':
suite = MNISTSparseExperiment()
# summarizeResults("./results", suite)
#
for expName in [
# "./results/standardOneLayer",
# Best sparse CNN net so far
# "./results/bestSparseNetLuiz",
# "./results/bestSparseNet",
# "./results/bestSparseCNN",
# This is the best sparse net (non CNN) so far, as of Jan 7.
# "./results/bestSparseNet/k50.0n500.0",
# "./results/DenseCNN1Seeds/seed47.0",
]:
# See sparse_net.py
KERNEL_SIZE = 5
def computeMaxPool(input_width):
"""
Compute CNN max pool width. see 'cnn_sdr.py'
"""
wout = math.floor((input_width + 2 * PADDING - KERNEL_SIZE) / STRIDE + 1)
return int(math.floor(wout / 2.0))
if __name__ == '__main__':
suite = MNISTSparseExperiment()
suite.parse_opt()
suite.parse_cfg()
experiments = suite.options.experiments or suite.cfgparser.sections()
paramsTable = [["Network",
"L1 F", "L1 Sparsity",
"L2 F", "L2 Sparsity",
"L3 N", "L3 Sparsity",
"Wt Sparsity"]]
for name in experiments:
exps = suite.get_exps(suite.get_exp(name)[0])
for exp in exps:
if not os.path.exists(exp):
continue
params = suite.get_params(exp=exp)
# This program is free software: you can redistribute it and/or modify
# it under the terms of the GNU Affero Public License version 3 as
# published by the Free Software Foundation.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
# See the GNU Affero Public License for more details.
#
# You should have received a copy of the GNU Affero Public License
# along with this program. If not, see http://www.gnu.org/licenses.
#
# http://numenta.org/licenses/
# ----------------------------------------------------------------------
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
from htmresearch.frameworks.pytorch.mnist_sparse_experiment import \
MNISTSparseExperiment
# Run from htmresearch:
# domino run projects/sdr_paper/pytorch_experiments/run_experiment.py -c projects/sdr_paper/pytorch_experiments/experiments.cfg
# OR
# python projects/sdr_paper/pytorch_experiments/run_experiment.py -c projects/sdr_paper/pytorch_experiments/experiments.cfg -d
if __name__ == '__main__':
suite = MNISTSparseExperiment()
suite.start()
transform.transforms.pop()
ax = fig.add_subplot(4, num_noise, y * num_noise + x + 1)
ax.set_axis_off()
ax.imshow(img.numpy().reshape((28, 28)), cmap='gray')
if y == 0:
ax.set_title("{0}%".format(noise_values[x] * 100))
plt.tight_layout()
plt.savefig(plotPath)
plt.close()
if __name__ == '__main__':
# Initialize experiment options and parameters
suite = MNISTSparseExperiment()
suite.parse_opt()
suite.parse_cfg()
path = suite.cfgparser.defaults()['path']
# Plot Noise Curve (MNIST)
results, format = configureNoisePlot(
suite,
experiments=[
# "BestDenseOneLayer2",
"DenseCNN1Seeds/seed49.0",
"twoLayerDenseCNNs/k1000.0n1000.0",
"bestSparseCNNOneLayerSeeds/seed48.0",
"bestSparseCNNTwoLayerSeeds/seed50.0",
],
labels=["dense-CNN1", "dense-CNN2", "sparse-CNN1", "sparse-CNN2"],