Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add updateBoostStrength and rezeroWeights to examples #9

Merged
merged 1 commit into from
Apr 14, 2019
Merged

Add updateBoostStrength and rezeroWeights to examples #9

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
96 changes: 67 additions & 29 deletions examples/sparse_cnn.ipynb
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,12 +30,12 @@
},
{
"cell_type": "code",
"execution_count": 2,
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# Uncomment the following line to install nupic.torch\n",
"#!pip install -e git+https://github.com/numenta/nupic.torch.git#egg=nupic.torch"
"!pip install -e git+https://github.com/numenta/nupic.torch.git#egg=nupic.torch"
]
},
{
Expand Down Expand Up @@ -196,7 +196,8 @@
"metadata": {},
"outputs": [],
"source": [
"from nupic.torch.modules import KWinners2d, KWinners, SparseWeights, Flatten\n",
"from nupic.torch.modules import (\n",
" KWinners2d, KWinners, SparseWeights, Flatten, rezeroWeights, updateBoostStrength)\n",
"\n",
"sparseCNN = nn.Sequential(\n",
" # Sparse CNN layer\n",
Expand Down Expand Up @@ -255,20 +256,55 @@
"cell_type": "code",
"execution_count": 8,
"metadata": {},
"outputs": [],
"source": [
"sgd = optim.SGD(sparseCNN.parameters(), lr=LEARNING_RATE, momentum=MOMENTUM)\n",
"train(model=sparseCNN, loader=first_loader, optimizer=sgd, criterion=F.nll_loss)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"After each epoch we rezero the weights to keep the initial sparsity constant during training. We also apply the boost strength factor after each epoch"
]
},
{
"cell_type": "code",
"execution_count": 9,
"metadata": {},
"outputs": [],
"source": [
"%%capture\n",
"sparseCNN.apply(rezeroWeights)\n",
"sparseCNN.apply(updateBoostStrength)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Test and print results"
]
},
{
"cell_type": "code",
"execution_count": 10,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"{'accuracy': 0.978, 'loss': 0.06757185096740723, 'total_correct': 9780}\n"
]
"data": {
"text/plain": [
"{'accuracy': 0.9782, 'loss': 0.06787856979370117, 'total_correct': 9782}"
]
},
"execution_count": 10,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"sgd = optim.SGD(sparseCNN.parameters(), lr=LEARNING_RATE, momentum=MOMENTUM)\n",
"train(model=sparseCNN, loader=first_loader, optimizer=sgd, criterion=F.nll_loss)\n",
"results = test(model=sparseCNN, loader=test_loader, criterion=F.nll_loss)\n",
"print(results)"
"test(model=sparseCNN, loader=test_loader, criterion=F.nll_loss)"
]
},
{
Expand All @@ -280,28 +316,30 @@
},
{
"cell_type": "code",
"execution_count": 9,
"execution_count": 11,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"{'accuracy': 0.9856, 'loss': 0.04167602500915527, 'total_correct': 9856}\n",
"{'accuracy': 0.9868, 'loss': 0.040336785697937014, 'total_correct': 9868}\n",
"{'accuracy': 0.9872, 'loss': 0.03909029731750488, 'total_correct': 9872}\n",
"{'accuracy': 0.9874, 'loss': 0.037309212684631346, 'total_correct': 9874}\n",
"{'accuracy': 0.9876, 'loss': 0.037822017288208006, 'total_correct': 9876}\n",
"{'accuracy': 0.9877, 'loss': 0.03768303909301758, 'total_correct': 9877}\n",
"{'accuracy': 0.988, 'loss': 0.03783873291015625, 'total_correct': 9880}\n",
"{'accuracy': 0.9881, 'loss': 0.038252718925476076, 'total_correct': 9881}\n",
"{'accuracy': 0.9873, 'loss': 0.03846522216796875, 'total_correct': 9873}\n"
"{'accuracy': 0.9862, 'loss': 0.0412615779876709, 'total_correct': 9862}\n",
"{'accuracy': 0.9868, 'loss': 0.04029187545776367, 'total_correct': 9868}\n",
"{'accuracy': 0.9867, 'loss': 0.03934368209838867, 'total_correct': 9867}\n",
"{'accuracy': 0.9876, 'loss': 0.03759277114868164, 'total_correct': 9876}\n",
"{'accuracy': 0.9877, 'loss': 0.03777754402160644, 'total_correct': 9877}\n",
"{'accuracy': 0.9872, 'loss': 0.03784116630554199, 'total_correct': 9872}\n",
"{'accuracy': 0.9872, 'loss': 0.03829168930053711, 'total_correct': 9872}\n",
"{'accuracy': 0.9876, 'loss': 0.03837165260314941, 'total_correct': 9876}\n",
"{'accuracy': 0.9871, 'loss': 0.03919747161865234, 'total_correct': 9871}\n"
]
}
],
"source": [
"for epoch in range(1, EPOCHS):\n",
" train(model=sparseCNN, loader=train_loader, optimizer=sgd, criterion=F.nll_loss)\n",
" sparseCNN.apply(rezeroWeights)\n",
" sparseCNN.apply(updateBoostStrength)\n",
" results = test(model=sparseCNN, loader=test_loader, criterion=F.nll_loss)\n",
" print(results)"
]
Expand All @@ -316,18 +354,18 @@
},
{
"cell_type": "code",
"execution_count": 10,
"execution_count": 12,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"0.05 : {'accuracy': 0.9842, 'loss': 0.04920544128417969, 'total_correct': 9842}\n",
"0.1 : {'accuracy': 0.9782, 'loss': 0.067373148727417, 'total_correct': 9782}\n",
"0.15 : {'accuracy': 0.9697, 'loss': 0.0976463264465332, 'total_correct': 9697}\n",
"0.2 : {'accuracy': 0.9524, 'loss': 0.15406905670166016, 'total_correct': 9524}\n",
"0.25 : {'accuracy': 0.9238, 'loss': 0.23606297912597657, 'total_correct': 9238}\n"
"0.05 : {'accuracy': 0.9841, 'loss': 0.05029125137329102, 'total_correct': 9841}\n",
"0.1 : {'accuracy': 0.9777, 'loss': 0.06853677139282227, 'total_correct': 9777}\n",
"0.15 : {'accuracy': 0.9697, 'loss': 0.09935387115478515, 'total_correct': 9697}\n",
"0.2 : {'accuracy': 0.9509, 'loss': 0.1598887924194336, 'total_correct': 9509}\n",
"0.25 : {'accuracy': 0.9225, 'loss': 0.2412475631713867, 'total_correct': 9225}\n"
]
}
],
Expand Down Expand Up @@ -359,7 +397,7 @@
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.6.7"
"version": "3.7.3"
}
},
"nbformat": 4,
Expand Down
104 changes: 73 additions & 31 deletions examples/sparse_linear.ipynb
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,8 +30,10 @@
},
{
"cell_type": "code",
"execution_count": 2,
"metadata": {},
"execution_count": null,
"metadata": {
"scrolled": true
},
"outputs": [],
"source": [
"# Uncomment the following line to install nupic.torch\n",
Expand Down Expand Up @@ -188,7 +190,8 @@
"metadata": {},
"outputs": [],
"source": [
"from nupic.torch.modules import KWinners, SparseWeights, Flatten\n",
"from nupic.torch.modules import (\n",
" KWinners, SparseWeights, Flatten, rezeroWeights, updateBoostStrength)\n",
"\n",
"sparseNN = nn.Sequential(\n",
" Flatten(),\n",
Expand Down Expand Up @@ -244,53 +247,90 @@
"cell_type": "code",
"execution_count": 8,
"metadata": {},
"outputs": [],
"source": [
"sgd = optim.SGD(sparseNN.parameters(), lr=LEARNING_RATE, momentum=MOMENTUM)\n",
"train(model=sparseNN, loader=first_loader, optimizer=sgd, criterion=F.nll_loss)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"After each epoch we rezero the weights to keep the initial sparsity constant during training. We also apply the boost strength factor after each epoch"
]
},
{
"cell_type": "code",
"execution_count": 9,
"metadata": {},
"outputs": [],
"source": [
"%%capture\n",
"sparseNN.apply(rezeroWeights)\n",
"sparseNN.apply(updateBoostStrength)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Test and print results"
]
},
{
"cell_type": "code",
"execution_count": 10,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"{'accuracy': 0.9506, 'loss': 0.16315889587402344, 'total_correct': 9506}\n"
]
"data": {
"text/plain": [
"{'accuracy': 0.9486, 'loss': 0.16556934661865233, 'total_correct': 9486}"
]
},
"execution_count": 10,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"sgd = optim.SGD(sparseNN.parameters(), lr=LEARNING_RATE, momentum=MOMENTUM)\n",
"train(model=sparseNN, loader=first_loader, optimizer=sgd, criterion=F.nll_loss)\n",
"results = test(model=sparseNN, loader=test_loader, criterion=F.nll_loss)\n",
"print(results)"
"test(model=sparseNN, loader=test_loader, criterion=F.nll_loss)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"At this point the duty cycles should be stable and we can train on larger batch sizes"
"At this point the duty cycles should be stable and we can train the rest of the epochs on larger batch sizes"
]
},
{
"cell_type": "code",
"execution_count": 9,
"execution_count": 11,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"{'accuracy': 0.9627, 'loss': 0.12201591415405273, 'total_correct': 9627}\n",
"{'accuracy': 0.9634, 'loss': 0.12190820617675781, 'total_correct': 9634}\n",
"{'accuracy': 0.9623, 'loss': 0.12345575714111329, 'total_correct': 9623}\n",
"{'accuracy': 0.9639, 'loss': 0.1185587173461914, 'total_correct': 9639}\n",
"{'accuracy': 0.9611, 'loss': 0.11994301071166992, 'total_correct': 9611}\n",
"{'accuracy': 0.9633, 'loss': 0.11600606689453125, 'total_correct': 9633}\n",
"{'accuracy': 0.9634, 'loss': 0.11699238586425781, 'total_correct': 9634}\n",
"{'accuracy': 0.9639, 'loss': 0.11530724716186523, 'total_correct': 9639}\n",
"{'accuracy': 0.9633, 'loss': 0.11797227020263672, 'total_correct': 9633}\n"
"{'accuracy': 0.9621, 'loss': 0.12309523391723633, 'total_correct': 9621}\n",
"{'accuracy': 0.9625, 'loss': 0.12451380462646484, 'total_correct': 9625}\n",
"{'accuracy': 0.9621, 'loss': 0.12468773880004883, 'total_correct': 9621}\n",
"{'accuracy': 0.9638, 'loss': 0.11706881561279296, 'total_correct': 9638}\n",
"{'accuracy': 0.9623, 'loss': 0.120688623046875, 'total_correct': 9623}\n",
"{'accuracy': 0.9645, 'loss': 0.11490174255371094, 'total_correct': 9645}\n",
"{'accuracy': 0.9648, 'loss': 0.1163398452758789, 'total_correct': 9648}\n",
"{'accuracy': 0.9643, 'loss': 0.1144802864074707, 'total_correct': 9643}\n",
"{'accuracy': 0.9657, 'loss': 0.11591439743041992, 'total_correct': 9657}\n"
]
}
],
"source": [
"for epoch in range(1, EPOCHS):\n",
" train(model=sparseNN, loader=train_loader, optimizer=sgd, criterion=F.nll_loss)\n",
" sparseNN.apply(updateBoostStrength)\n",
" sparseNN.apply(rezeroWeights)\n",
" results = test(model=sparseNN, loader=test_loader, criterion=F.nll_loss)\n",
" print(results)"
]
Expand All @@ -305,18 +345,20 @@
},
{
"cell_type": "code",
"execution_count": 10,
"metadata": {},
"execution_count": 12,
"metadata": {
"scrolled": true
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"0.05 : {'accuracy': 0.9541, 'loss': 0.14802040100097658, 'total_correct': 9541}\n",
"0.1 : {'accuracy': 0.939, 'loss': 0.1916733642578125, 'total_correct': 9390}\n",
"0.15 : {'accuracy': 0.9094, 'loss': 0.2852293640136719, 'total_correct': 9094}\n",
"0.2 : {'accuracy': 0.8639, 'loss': 0.4125948455810547, 'total_correct': 8639}\n",
"0.25 : {'accuracy': 0.8043, 'loss': 0.5801907653808593, 'total_correct': 8043}\n"
"0.05 : {'accuracy': 0.9563, 'loss': 0.14633502197265624, 'total_correct': 9563}\n",
"0.1 : {'accuracy': 0.9406, 'loss': 0.19336406707763673, 'total_correct': 9406}\n",
"0.15 : {'accuracy': 0.9114, 'loss': 0.28374275512695313, 'total_correct': 9114}\n",
"0.2 : {'accuracy': 0.8656, 'loss': 0.4086779327392578, 'total_correct': 8656}\n",
"0.25 : {'accuracy': 0.8082, 'loss': 0.5770498291015626, 'total_correct': 8082}\n"
]
}
],
Expand Down Expand Up @@ -348,7 +390,7 @@
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.6.7"
"version": "3.7.3"
}
},
"nbformat": 4,
Expand Down