def dataset(request):
dt = cnn.Cifar10_data()
def fin():
print("Teardown dataset")
request.addfinalizer(fin)
return dt
"optim_args": {
"lr": LR,
"weight_decay": 0.00001
},
},
}
results = pd.DataFrame(
index=range(REPEATS * len(group_config)),
columns=["Group", "Train Accuracy", "Test Accuracy", "Repeat"],
)
counter = 0
# DATASET
data_loader = cnn.Cifar10_data(mode="test", batch_size=BATCH_SIZE)
for i in range(REPEATS):
print(f"Running trial {i+1}/{REPEATS}")
net = cnn.NgnCnn(args.neurons, seed=i)
for group in group_config:
net_copy = copy.deepcopy(net)
if "Dropout" in group:
net_copy.dropout = 0.2
if "Neural Noise" in group:
net_copy.neural_noise = (-0.2, 0.5)
net_copy.to(device)
parameters = group_config[group]
log, _ = cnn.train_model(
model=net_copy,
dataset=data_loader,
"lr": LR,
},
},
"Control": {
"epochs": EPOCHS,
"neurogenesis": 0,
"early_stop": False,
"optim_args":{
"lr": LR,
}
},
}
# DATASET
data_loader = cnn.Cifar10_data(mode="validation", data_folder=DATA_DIR, batch_size=BATCH_SIZE, num_workers=16)
REPEATS = 1
for i in range(REPEATS):
print(f"Running trial {i+1}/{REPEATS}")
net = cnn.NgnCnn(args.neurons, channels=3, seed=i, excite=True)
for group in group_config:
net_copy = copy.deepcopy(net)
net_copy.to(device)
if group == 'Dropout':
net_copy.dropout = 0.2
parameters = group_config[group]
log, optimizer = cnn.train_model(
model=net_copy,
dataset=data_loader,
dtype=dtype,
device=device,
counter = 0
def sim_matrix(a, b, eps=1e-8):
"""
added eps for numerical stability
"""
a_n, b_n = a.norm(dim=1)[:, None], b.norm(dim=1)[:, None]
a_norm = a / torch.max(a_n, eps * torch.ones_like(a_n))
b_norm = b / torch.max(b_n, eps * torch.ones_like(b_n))
sim_mt = torch.mm(a_norm, b_norm.transpose(0, 1))
return sim_mt
# DATASET
data_loader = cnn.Cifar10_data(mode="test", batch_size=4)
for i in range(REPEATS):
print(f"Running trial {i+1}/{REPEATS}")
net = cnn.NgnCnn(args.neurons, seed=i)
for group_num, group in enumerate(group_config):
net_copy = copy.deepcopy(net)
if "Dropout" in group:
net_copy.dropout = 0.2
net_copy.to(device)
parameters = group_config[group]
log, _ = cnn.train_model(
model=net_copy,
dataset=data_loader,
dtype=dtype,
device=device,
import numpy as np
import neurodl.cnn as cnn
import argparse
import torch.optim as optim
from datetime import datetime
from pathlib import Path
from ax.service.managed_loop import optimize
from ax import save
import joblib
torch.manual_seed(12345)
datatype = torch.float
DEVICE = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
BATCH_SIZE = 1000
data_loader = cnn.Cifar10_data(mode="validation", batch_size=BATCH_SIZE)
parametrization = [
{
"name": "lr",
"type": "range",
"bounds": [0.00001, 0.001],
"value_type":
"float", # Optional, defaults to inference from type of "bounds".
"log_scale": True,
},
{
"name": "momentum",
"type": "range",
"bounds": [0, 1],
"value_type":