def __init__(self,
width_mult=1.0,
classifier=True,
classifier_out_features=1000):
super(MobileNetV3, self).__init__()
conv_0_0 = nn.Conv2d(in_channels=3,
out_channels=int(16 * width_mult),
kernel_size=(3, 3),
stride=2,
padding=3 // 2,
bias=False)
conv_0_1 = nn.BatchNorm2d(num_features=int(16 * width_mult))
conv_0_2 = nn.Hardswish(inplace=True)
conv_1_0 = Bottleneck(in_channels=int(16 * width_mult),
out_channels=int(16 * width_mult),
dw_kernel_size=(3, 3),
expand_size=16,
squeeze_excite=True,
nonlinearity='relu',
stride=2)
conv_2_0 = Bottleneck(in_channels=int(16 * width_mult),
out_channels=int(24 * width_mult),
dw_kernel_size=(3, 3),
expand_size=72,
squeeze_excite=False,
nonlinearity='relu',
stride=2)
conv_3_0 = Bottleneck(in_channels=int(24 * width_mult),
out_channels=int(24 * width_mult),
dw_kernel_size=(3, 3),
expand_size=88,
squeeze_excite=False,
nonlinearity='relu',
stride=1)
conv_4_0 = Bottleneck(in_channels=int(24 * width_mult),
out_channels=int(40 * width_mult),
dw_kernel_size=(5, 5),
expand_size=96,
squeeze_excite=True,
nonlinearity='hardswish',
stride=2)
conv_5_0 = Bottleneck(in_channels=int(40 * width_mult),
out_channels=int(40 * width_mult),
dw_kernel_size=(5, 5),
expand_size=240,
squeeze_excite=True,
nonlinearity='hardswish',
stride=1)
conv_6_0 = Bottleneck(in_channels=int(40 * width_mult),
out_channels=int(40 * width_mult),
dw_kernel_size=(5, 5),
expand_size=240,
squeeze_excite=True,
nonlinearity='hardswish',
stride=1)
conv_7_0 = Bottleneck(in_channels=int(40 * width_mult),
out_channels=int(48 * width_mult),
dw_kernel_size=(5, 5),
expand_size=120,
squeeze_excite=True,
nonlinearity='hardswish',
stride=1)
conv_8_0 = Bottleneck(in_channels=int(48 * width_mult),
out_channels=int(48 * width_mult),
dw_kernel_size=(5, 5),
expand_size=144,
squeeze_excite=True,
nonlinearity='hardswish',
stride=1)
conv_9_0 = Bottleneck(in_channels=int(48 * width_mult),
out_channels=int(96 * width_mult),
dw_kernel_size=(5, 5),
expand_size=288,
squeeze_excite=True,
nonlinearity='hardswish',
stride=2)
conv_10_0 = Bottleneck(in_channels=int(96 * width_mult),
out_channels=int(96 * width_mult),
dw_kernel_size=(5, 5),
expand_size=576,
squeeze_excite=True,
nonlinearity='hardswish',
stride=1)
conv_11_0 = Bottleneck(in_channels=int(96 * width_mult),
out_channels=int(96 * width_mult),
dw_kernel_size=(5, 5),
expand_size=576,
squeeze_excite=True,
nonlinearity='hardswish',
stride=1)
conv_12_0 = nn.Conv2d(in_channels=int(96 * width_mult),
out_channels=int(576 * width_mult),
kernel_size=(1, 1),
bias=False)
conv_12_1 = nn.Hardswish(inplace=True)
conv_12_2 = nn.BatchNorm2d(num_features=int(576 * width_mult))
self.features = nn.Sequential(conv_0_0, conv_0_1, conv_0_2, conv_1_0,
conv_2_0, conv_3_0, conv_4_0, conv_5_0,
conv_6_0, conv_7_0, conv_8_0, conv_9_0,
conv_10_0, conv_11_0, conv_12_0,
conv_12_1, conv_12_2)
if classifier:
self.classifiers = nn.Sequential(
nn.AdaptiveAvgPool2d(output_size=1), nn.Flatten(start_dim=1),
nn.Linear(int(576 * width_mult), int(1024 * width_mult)),
nn.Dropout(p=0.2),
nn.Linear(int(1024 * width_mult), classifier_out_features))
else:
self.classifiers = nn.Identity()
ACTIONS = 4 # number of valid actions
GAMMA = 0.99 # decay rate of past observations
OBSERVE = 3000. # timesteps to observe before training
EXPLORE = 2000000. # frames over which to anneal epsilon
FINAL_EPSILON = 0.0001 # final value of epsilon
INITIAL_EPSILON = 0.1 # starting value of epsilon
REPLAY_MEMORY = 50000 # number of previous transitions to remember
BATCH = 128 # size of minibatch
LEARNING_RATE = 1e-6 # learning rate
weightfile = 'weight2048.pt'
# if gpu is to be used
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
net = nn.Sequential(nn.Conv2d(1, 1024, 2), nn.ReLU(), nn.Conv2d(1024, 1024, 2),
nn.ReLU(), nn.Flatten(), nn.Linear(4096, 1024), nn.ReLU(),
nn.Linear(1024, 256), nn.ReLU(), nn.Linear(256, 4),
nn.ReLU()).to(device)
optimizer = torch.optim.Adam(net.parameters(), lr=LEARNING_RATE)
try:
net.load_state_dict(torch.load(weightfile))
net.eval()
print("Successfully loaded:")
except:
print("Could not find old network weights")
def trainNetwork(model):
game_state = game.GameState()
# store the previous observations in replay memory
def __init__(self, n_base_channels=16):
super(PoseNet, self).__init__()
self.vgg_part = nn.ModuleList([
VggBlock(in_channels=3,
out_channels=n_base_channels,
kernel_size=7,
padding=3,
maxpool=False), # 16
VggBlock(in_channels=n_base_channels,
out_channels=n_base_channels,
kernel_size=7,
padding=3,
maxpool=True), # 16
VggBlock(in_channels=n_base_channels,
out_channels=n_base_channels * 2,
kernel_size=5,
padding=2,
maxpool=False), # 32
VggBlock(in_channels=n_base_channels * 2,
out_channels=n_base_channels * 2,
kernel_size=5,
padding=2,
maxpool=True), # 32
VggBlock(in_channels=n_base_channels * 2,
out_channels=n_base_channels * 4,
kernel_size=3,
padding=1,
maxpool=False), # 64
VggBlock(in_channels=n_base_channels * 4,
out_channels=n_base_channels * 4,
kernel_size=3,
padding=1,
maxpool=True), # 64
VggBlock(in_channels=n_base_channels * 4,
out_channels=n_base_channels * 8,
kernel_size=3,
padding=1,
maxpool=False), # 128
VggBlock(in_channels=n_base_channels * 8,
out_channels=n_base_channels * 8,
kernel_size=3,
padding=1,
maxpool=True), # 128
VggBlock(in_channels=n_base_channels * 8,
out_channels=n_base_channels * 16,
kernel_size=3,
padding=1,
maxpool=False), # 256
VggBlock(in_channels=n_base_channels * 16,
out_channels=n_base_channels * 16,
kernel_size=3,
padding=1,
maxpool=True), # 256
VggBlock(in_channels=n_base_channels * 16,
out_channels=n_base_channels * 16,
kernel_size=3,
padding=1,
maxpool=False), # 256
VggBlock(in_channels=n_base_channels * 16,
out_channels=n_base_channels * 16,
kernel_size=3,
padding=1,
maxpool=True), # 256
VggBlock(in_channels=n_base_channels * 16,
out_channels=n_base_channels * 32,
kernel_size=3,
padding=1,
maxpool=False), # 512
VggBlock(in_channels=n_base_channels * 32,
out_channels=n_base_channels * 32,
kernel_size=3,
padding=1,
maxpool=True), # 512
])
self.avgpool = nn.AdaptiveAvgPool2d((7, 7))
self.flatten = nn.Flatten()
self.rot1 = nn.Linear(n_base_channels * 32 * 7 * 7, 512)
self.rot2 = nn.Linear(512, 512)
self.rot3 = nn.Linear(512, 3)
self.transl1 = nn.Linear(n_base_channels * 32 * 7 * 7, 512)
self.transl2 = nn.Linear(512, 512)
self.transl3 = nn.Linear(512, 3)
def graft_net(args):
global logger_net
logger_net = Logger('log/graft_net_{}_{}_{}perclass.txt'.\
format(args.dataset, time.strftime("%Y-%m-%d_%H-%M-%S", time.localtime()),
args.num_per_class))
# ---------------------- Datasets ----------------------
if args.dataset == 'CIFAR10':
train_loader = DataLoader(CIFAR10Few(args.data_path,
args.num_per_class,
transform=get_transformer(
args.dataset,
cropsize=32,
crop_padding=4,
hflip=True)),
batch_size=args.batch_size,
num_workers=4,
shuffle=True)
elif args.dataset == 'CIFAR100':
train_loader = DataLoader(CIFAR100Few(args.data_path,
args.num_per_class,
transform=get_transformer(
args.dataset,
cropsize=32,
crop_padding=4,
hflip=True)),
batch_size=args.batch_size,
num_workers=4,
shuffle=True)
test_loader = DataLoader(get_dataset(args, train_flag=False),
batch_size=args.batch_size,
num_workers=4,
shuffle=False)
cfg_t = cfgs['vgg16']
cfg_s = cfgs['vgg16-graft']
cfg_blocks_t = split_block(cfg_t)
cfg_blocks_s = split_block(cfg_s)
num_block = len(block_graft_ids)
# ---------------------- Adaption ----------------------
adaptions_t2s = [
nn.Conv2d(cfg_blocks_t[block_graft_ids[i]][-2],
cfg_blocks_s[block_graft_ids[i]][-2],
kernel_size=1).cuda() for i in range(0, num_block - 1)
]
adaptions_s2t = [
nn.Conv2d(cfg_blocks_s[block_graft_ids[i]][-2],
cfg_blocks_t[block_graft_ids[i]][-2],
kernel_size=1).cuda() for i in range(0, num_block - 1)
]
# ---------------------- Teacher ----------------------
teacher = vgg_stock(cfg_t, args.dataset, args.num_class)
params_t = torch.load(args.ckpt)
teacher.cuda().eval()
teacher.load_state_dict(params_t)
# ---------------------- Blocks ----------------------
params_s = {}
for key in params_t.keys():
key_split = key.split('.')
if key_split[0] == 'features' and \
key_split[1] in ['0', '1', '2']:
params_s[key] = params_t[key]
student = vgg_bw(cfg_s, True, args.dataset, args.num_class)
student.cuda().train()
student.load_state_dict(params_s, strict=False)
blocks_s = [student.features[i] for i in block_graft_ids[:-1]]
blocks_s += [nn.Sequential(nn.Flatten().cuda(), student.classifier)]
blocks = []
for block_id in range(num_block):
blocks.append(
warp_block(blocks_s, block_id, adaptions_t2s,
adaptions_s2t).cuda())
params = torch.load('ckpt/student/vgg16-student-graft-block-{}-{}perclass.pth'.\
format(args.dataset, args.num_per_class))
for block_id in range(num_block):
blocks[block_id].load_state_dict(params['block-{}'.format(block_id)])
for i in range(num_block - 1):
block = nn.Sequential(*blocks[:(i + 2)])
optimizer = optim.Adam(block.parameters(), lr=0.0001)
scion_len = sum(blocks_s_len[:(i + 2)])
accuracy_best_block = 0.0
params_best_save = None
for epoch in range(args.num_epoch[i]):
if logger_net: logger_net.write('Epoch', epoch)
loss_value = train_epoch(args, teacher, block, scion_len,
train_loader, optimizer)
accuracy = test(teacher, test_loader)
if accuracy > accuracy_best_block:
accuracy_best_block = accuracy
params_tmp = block.cpu().state_dict()
params_best_save = params_tmp.copy()
block.cuda()
if epoch == (args.num_epoch[i] - 1) and \
i == (num_block - 2):
block.load_state_dict(params_best_save)
if logger_net:
logger_net.write('Accuracy-length-{}'.format(scion_len),
accuracy)
if logger_net:
logger_net.write('Student Best Accuracy', accuracy_best_block)
with open('ckpt/student/vgg16-student-graft-net-{}-{}perclass.pth'\
.format(args.dataset, args.num_per_class), 'wb') as f:
torch.save(block.state_dict(), f)
if logger_net:
logger_net.close()
return accuracy_best_block
def __init__(self, c1, c2, k=1, s=1, p=None, g=1):
super(Classify, self).__init__()
self.aap = nn.AdaptiveAvgPool2d(1) # to x(b,c1,1,1)
self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p),
groups=g) # to x(b,c2,1,1)
self.flat = nn.Flatten()
def get_classifier(self):
return nn.Sequential(self.avgpool, nn.Flatten(start_dim=1), self.fc)
def forward(self,x):
x = nn.Flatten(x)
x = F.relu(self.fc1(x))
x = self.fc2(x)
return x
#!/usr/bin/env python
# coding: utf-8
# In[1]:
#4-3 Concise Implementation of Multilayer Perceptrons
import torch
from torch import nn
from pytorch_d2l.d2l import torch as d2l
#Model: hidden layer 256 units, output layer and ReLU activation function.
net = nn.Sequential(nn.Flatten(), nn.Linear(784, 256), nn.ReLU(), nn.Linear(256, 10))
def init_weights(m):
if type(m) == nn.Linear:
nn.init.normal_(m.weight, std=0.01)
net.apply(init_weights)
#Parameters, hyperparameters
batch_size, lr, num_epochs = 256, 0.1, 10
loss = nn.CrossEntropyLoss()
trainer = torch.optim.SGD(net.parameters(), lr=lr)
train_iter, test_iter = d2l.load_data_fashion_mnist(batch_size)
d2l.train_ch3(net, train_iter, test_iter, loss, num_epochs, trainer)
def __init__(self, lr, inputDim, outputDim):
super(forwardNet, self).__init__()
modelStr = [
nn.Linear(inputDim, 300),
nn.BatchNorm1d(300),
nn.LeakyReLU(),
nn.Linear(300, 300),
nn.BatchNorm1d(300),
nn.LeakyReLU(),
nn.Dropout(p=0.5),
nn.Linear(300, 300),
nn.BatchNorm1d(300),
nn.LeakyReLU(),
nn.Dropout(p=0.5),
nn.Linear(300, 300),
nn.BatchNorm1d(300),
nn.LeakyReLU(),
nn.Dropout(p=0.5),
nn.Linear(300, 300),
nn.BatchNorm1d(300),
nn.LeakyReLU(),
nn.Dropout(p=0.5),
nn.Linear(300, outputDimS),
nn.Sigmoid()
]
modelHole = [
nn.Conv2d(3, 32, kernel_size=3, padding=1),
nn.ReLU(),
nn.Conv2d(32, 128, kernel_size=3, stride=1, padding=1),
nn.ReLU(),
nn.MaxPool2d(1, 2), # output: 128x5x10
nn.Conv2d(128, 200, kernel_size=3, stride=1, padding=1),
nn.ReLU(),
nn.Conv2d(200, 200, kernel_size=3, stride=1, padding=1),
nn.ReLU(),
nn.MaxPool2d(1, 2), # output: 200x5x5
nn.Conv2d(200, 250, kernel_size=3, stride=1, padding=1),
nn.ReLU(),
nn.Conv2d(250, 250, kernel_size=3, stride=1, padding=1),
nn.ReLU(),
#nn.MaxPool2d(2, 2), # output: 250 x 4 x 4
nn.Flatten(),
nn.Linear(2500, 1000),
nn.ReLU(),
nn.Linear(1000, 500),
nn.ReLU(),
nn.Dropout(0.2),
nn.Linear(500, outputDimH)
]
outputDecoder = [
nn.Linear(outputDimH + outputDimS, 100),
nn.BatchNorm1d(100),
nn.ReLU(),
nn.Dropout(p=0.5),
nn.Linear(100, 200),
nn.BatchNorm1d(200),
nn.ReLU(),
nn.Dropout(p=0.5),
nn.Linear(200, 300),
nn.BatchNorm1d(300),
nn.LeakyReLU(),
nn.Dropout(p=0.5),
nn.Linear(300, outputDim),
nn.ReLU()
]
self.modelStr = nn.Sequential(*modelStr)
self.modelHole = nn.Sequential(*modelHole)
self.outputDecoder = nn.Sequential(*outputDecoder)
self.lr = lr
self.decayRate = decayRate
self.momentum = momentum
self.optim = torch.optim.RMSprop(
list(self.modelStr.parameters()) +
list(self.modelHole.parameters()) +
list(self.outputDecoder.parameters()), self.lr, self.decayRate)
self.criterion = nn.MSELoss()
self.accuracy = nn.L1Loss()
def __init__(self,
input_size,
num_classes,
l0_strength=7e-6,
l2_strength=0,
droprate_init=0.5,
learn_weight=True,
random_weight=True,
decay_mean=False,
deterministic=False,
use_batch_norm=True,
cnn_out_channels=(64, 64),
kernel_size=5,
linear_units=1000,
maxpool_stride=2,
bn_track_running_stats=True):
feature_map_sidelength = (((
(input_size[1] - kernel_size + 1) / maxpool_stride) - kernel_size +
1) / maxpool_stride)
assert (feature_map_sidelength == int(feature_map_sidelength))
feature_map_sidelength = int(feature_map_sidelength)
modules = [
# -------------
# Conv Block
# -------------
("cnn1",
VDropConv2d(input_size[0], cnn_out_channels[0], kernel_size)),
("cnn1_maxpool", nn.MaxPool2d(maxpool_stride)),
]
if use_batch_norm:
modules.append(
("cnn1_bn",
nn.BatchNorm2d(cnn_out_channels[0],
affine=False,
track_running_stats=bn_track_running_stats)))
modules += [
("cnn1_relu", nn.ReLU(inplace=True)),
# -------------
# Conv Block
# -------------
("cnn2",
VDropConv2d(cnn_out_channels[0], cnn_out_channels[1],
kernel_size)),
("cnn2_maxpool", nn.MaxPool2d(maxpool_stride)),
]
if use_batch_norm:
modules.append(
("cnn2_bn",
nn.BatchNorm2d(cnn_out_channels[1],
affine=False,
track_running_stats=bn_track_running_stats)))
modules += [
("cnn2_relu", nn.ReLU(inplace=True)),
("flatten", nn.Flatten()),
# -------------
# Linear Block
# -------------
("fc1",
VDropLinear((feature_map_sidelength**2) * cnn_out_channels[1],
linear_units)),
]
if use_batch_norm:
modules.append(
("fc1_bn",
nn.BatchNorm1d(linear_units,
affine=False,
track_running_stats=bn_track_running_stats)))
modules += [
("fc1_relu", nn.ReLU(inplace=True)),
# -------------
# Output Layer
# -------------
("fc2", VDropLinear(linear_units, num_classes)),
]
super().__init__(OrderedDict(modules))
def __init__(self,
embedding_dim,
dropout_rate,
num_class,
vocab_size=0,
seq_length=0,
num_blocks=3,
num_filters=250,
kernel_sizes=3,
embedding_matrix=None,
requires_grads=False):
'''
initialization
⚠️In default,the way to initialize embedding is loading pretrained embedding look-up table!
:param embedding_dim: embedding dim
:param num_class: the number of label
:param dropout_rate: dropout rate
:param vocab_size: vocabulary size
:param seq_length: max length of sequence after padding
:param num_blocks: the number of block in DPCNN model
:param num_filters: the number of filters of conv kernel
:param kernel_sizes: conv kernel size
:param embedding_matrix: pretrained embedding look up table
:param requires_grads: whether to update gradient of embedding in training stage
'''
super(DPCNN, self).__init__()
self.vocab_size = vocab_size
self.seq_length = seq_length
self.embedding_dim = embedding_dim
self.num_filters = num_filters
self.dropout_rate=dropout_rate
self.num_blocks = num_blocks
self.num_class = num_class
self.kernel_sizes = kernel_sizes
self.embedding_matrix = embedding_matrix
self.requires_grads = requires_grads
# embedding
if self.embedding_matrix is None:
self.embedding = nn.Embedding(num_embeddings=self.vocab_size,
embedding_dim=self.embedding_dim,
padding_idx=0)
else:
self.embedding = nn.Embedding.from_pretrained(self.embedding_matrix, freeze=self.requires_grads)
self.vocab_size = self.embedding_matrix.shape[0]
# text region embedding
self.region_embedding = nn.Conv2d(in_channels=1, out_channels=self.num_filters,
stride=1, kernel_size=(self.kernel_sizes, self.embedding_dim))
# two conv
self.conv2d1 = nn.Conv2d(in_channels=self.num_filters, out_channels=self.num_filters,
stride=2, kernel_size=(self.kernel_sizes, 1), padding=0)
self.conv2d2 = nn.Conv2d(in_channels=self.num_filters, out_channels=self.num_filters,
stride=2, kernel_size=(self.kernel_sizes, 1), padding=0)
self.padding1 = nn.ZeroPad2d((0, 0, (self.kernel_sizes-1)//2, (self.kernel_sizes-1)-((self.kernel_sizes-1)//2))) # top bottom
self.padding2 = nn.ZeroPad2d((0, 0, 0, self.kernel_sizes-2)) # bottom
# one block
self.block_max_pool = nn.MaxPool2d(kernel_size=(self.kernel_sizes, 1), stride=2)
self.conv2d3 = nn.Conv2d(in_channels=self.num_filters, out_channels=self.num_filters,
stride=1, kernel_size=(self.kernel_sizes, 1), padding=0)
self.conv2d4 = nn.Conv2d(in_channels=self.num_filters, out_channels=self.num_filters,
stride=1, kernel_size=(self.kernel_sizes, 1), padding=0)
# final pool and softmax
self.flatten = nn.Flatten()
self.dropout=nn.Dropout(p=self.dropout_rate)
self.classifier = nn.Linear(in_features=self.num_filters, out_features=self.num_class)
def __init__(
self,
input_shape=(1, 32, 32),
cnn_out_channels=(64, 64),
num_classes=12,
use_batch_norm=True,
z_logvar_init=-10,
vdrop_data_class=VDropCentralData,
kernel_size=5,
linear_units=1000,
maxpool_stride=2,
bn_track_running_stats=True,
conv_target_density=(1.0, 1.0),
linear_target_density=(1.0, 1.0),
):
feature_map_sidelength = ((
(input_shape[1] - kernel_size + 1) / maxpool_stride) -
kernel_size + 1) / maxpool_stride
vdrop_data = vdrop_data_class(z_logvar_init=z_logvar_init)
assert feature_map_sidelength == int(feature_map_sidelength)
feature_map_sidelength = int(feature_map_sidelength)
modules = [
# -------------
# Conv Block
# -------------
(
"vdrop_cnn1",
prunable_vdrop_conv2d(
input_shape[0],
cnn_out_channels[0],
kernel_size,
vdrop_data,
target_density=conv_target_density[0],
),
),
("cnn1_maxpool", nn.MaxPool2d(maxpool_stride)),
]
if use_batch_norm:
modules.append((
"cnn1_bn",
nn.BatchNorm2d(
cnn_out_channels[0],
affine=False,
track_running_stats=bn_track_running_stats,
),
))
modules += [
("cnn1_relu", nn.ReLU(inplace=True)),
# -------------
# Conv Block
# -------------
(
"vdrop_cnn2",
prunable_vdrop_conv2d(
cnn_out_channels[0],
cnn_out_channels[1],
kernel_size,
vdrop_data,
target_density=conv_target_density[1],
),
),
("cnn2_maxpool", nn.MaxPool2d(maxpool_stride)),
]
if use_batch_norm:
modules.append((
"cnn2_bn",
nn.BatchNorm2d(
cnn_out_channels[1],
affine=False,
track_running_stats=bn_track_running_stats,
),
))
modules += [
("cnn2_relu", nn.ReLU(inplace=True)),
("flatten", nn.Flatten()),
# -------------
# Linear Block
# -------------
(
"vdrop_fc1",
prunable_vdrop_linear(
(feature_map_sidelength**2) * cnn_out_channels[1],
linear_units,
vdrop_data,
target_density=linear_target_density[0],
),
),
]
if use_batch_norm:
modules.append((
"fc1_bn",
nn.BatchNorm1d(
linear_units,
affine=False,
track_running_stats=bn_track_running_stats,
),
))
modules += [
("fc1_relu", nn.ReLU(inplace=True)),
# -------------
# Output Layer
# -------------
(
"vdrop_fc2",
prunable_vdrop_linear(
linear_units,
num_classes,
vdrop_data,
target_density=linear_target_density[1],
),
),
]
super().__init__(OrderedDict(modules))
vdrop_data.finalize()
self.vdrop_data = vdrop_data
def __init__(self):
super(Convolution3D, self).__init__()
self.Convolution1 = nn.Conv3d(in_channels=3,
out_channels=64,
kernel_size=(3, 3, 3),
stride=1,
padding=(1, 0, 0),
dilation=1,
groups=1,
bias=True,
padding_mode='zeros')
self.BatchN1 = nn.BatchNorm3d(num_features=64,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.MaxPooling1 = nn.MaxPool3d(kernel_size=(1, 2, 2),
stride=(1, 2, 2),
padding=(0, 0, 0),
dilation=1,
return_indices=False,
ceil_mode=False)
self.MaxPooling2 = nn.MaxPool3d(kernel_size=(1, 2, 2),
stride=(1, 2, 2),
padding=(0, 0, 0),
dilation=1,
return_indices=False,
ceil_mode=False)
self.Convolution2 = nn.Conv3d(in_channels=64,
out_channels=64,
kernel_size=3,
stride=1,
padding=(1, 0, 0))
self.BatchN2 = nn.BatchNorm3d(num_features=64,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.ResConvolution1 = nn.Conv3d(in_channels=64,
out_channels=64,
kernel_size=3,
stride=1,
padding=(1, 1, 1))
self.averagePool1 = nn.AvgPool3d(kernel_size=1,
stride=1,
padding=(0, 0, 0))
self.ResBatchN1 = nn.BatchNorm3d(num_features=64,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.Convolution3 = nn.Conv3d(in_channels=64,
out_channels=64,
kernel_size=3,
stride=1,
padding=(1, 0, 0))
self.BatchN3 = nn.BatchNorm3d(num_features=64,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.ResConvolution2 = nn.Conv3d(in_channels=64,
out_channels=64,
kernel_size=3,
stride=1,
padding=(1, 1, 1))
self.averagePool2 = nn.AvgPool3d(kernel_size=1,
stride=1,
padding=(0, 0, 0))
self.ResBatchN2 = nn.BatchNorm3d(num_features=64,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.Convolution4 = nn.Conv3d(in_channels=64,
out_channels=8,
kernel_size=(3, 3, 3),
stride=1,
padding=(1, 0, 0),
dilation=1,
groups=1,
bias=True,
padding_mode='zeros')
self.BatchN4 = nn.BatchNorm3d(num_features=8,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.Convolution5 = nn.Conv3d(in_channels=8,
out_channels=8,
kernel_size=(3, 3, 3),
stride=1,
padding=(1, 0, 0),
dilation=1,
groups=1,
bias=True,
padding_mode='zeros')
self.BatchN5 = nn.BatchNorm3d(num_features=8,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.Convolution6 = nn.Conv3d(in_channels=8,
out_channels=8,
kernel_size=(3, 3, 3),
stride=1,
padding=(1, 0, 0),
dilation=1,
groups=1,
bias=True,
padding_mode='zeros')
self.BatchN6 = nn.BatchNorm3d(num_features=8,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.Flatten1 = nn.Flatten(start_dim=2)
self.LSTM1 = nn.LSTM(input_size=10488,
hidden_size=64,
num_layers=1,
batch_first=True)
self.LSTM2 = nn.LSTM(input_size=64,
hidden_size=16,
num_layers=1,
batch_first=True)
self.fc1 = nn.Linear(in_features=16, out_features=512, bias=True)
self.fc2 = nn.Linear(in_features=512, out_features=128, bias=True)
self.fc3 = nn.Linear(in_features=128, out_features=64, bias=True)
self.fc4 = nn.Linear(in_features=64, out_features=16, bias=True)
self.fc5 = nn.Linear(in_features=16, out_features=1, bias=True)
nn.MaxPool2d(kernel_size=2, stride=2),
nn.Conv2d(6, 16, kernel_size=5), nn.ReLU(),
nn.MaxPool2d(kernel_size=2, stride=2),
nn.Flatten(),
nn.Linear(16 * 5 * 5, 120), nn.ReLU(),
nn.Linear(120, 84), nn.Sigmoid(),
nn.Linear(84, 10))
loss 0.203, train acc 0.923, test acc 0.897
50960.2 examples/sec on cuda:0
"""
net = torch.nn.Sequential(Reshape(), nn.Conv2d(1, 6, kernel_size=5, padding=2),
nn.ReLU(), nn.MaxPool2d(kernel_size=2, stride=2),
nn.Conv2d(6, 16, kernel_size=5), nn.ReLU(),
nn.MaxPool2d(kernel_size=2, stride=2), nn.Flatten(),
nn.Linear(16 * 5 * 5, 120), nn.ReLU(),
nn.Linear(120, 84), nn.Sigmoid(), nn.Linear(84, 10))
""" Original LeNet
net = torch.nn.Sequential(
Reshape(),
nn.Conv2d(1, 6, kernel_size=5, padding=2), nn.Sigmoid(),
nn.AvgPool2d(kernel_size=2, stride=2),
nn.Conv2d(6, 16, kernel_size=5), nn.Sigmoid(),
nn.AvgPool2d(kernel_size=2, stride=2),
nn.Flatten(),
nn.Linear(16 * 5 * 5, 120), nn.Sigmoid(),
nn.Linear(120, 84), nn.Sigmoid(),
nn.Linear(84, 10))
"""
X = torch.rand(size=(1, 1, 28, 28), dtype=torch.float32)
def __init__(self, device, in_nc=3, kernel_size=4, nf=48, im_size=256):
"""
:param in_nc: number of in channels
:param kernel_size: kernel size
:param nf: number of convolution filters after 1 layer
"""
super(InpaintingDiscriminator, self).__init__()
self.patch_dis = nn.ModuleList([
SNBlock(in_channels=in_nc,
out_channels=nf,
kernel_size=kernel_size,
stride=2,
padding=get_pad(im_size, kernel_size, 2),
norm='in',
activation='relu',
pad_type='zero'),
SNBlock(in_channels=nf,
out_channels=nf * 2,
kernel_size=kernel_size,
stride=2,
padding=get_pad(im_size // 2, kernel_size, 2),
norm='in',
activation='relu',
pad_type='zero'),
SNBlock(in_channels=nf * 2,
out_channels=nf * 2,
kernel_size=kernel_size,
stride=2,
padding=get_pad(im_size // 4, kernel_size, 2),
norm='in',
activation='relu',
pad_type='zero'),
SNBlock(in_channels=nf * 2,
out_channels=nf * 4,
kernel_size=kernel_size,
stride=2,
padding=get_pad(im_size // 8, kernel_size, 2),
norm='in',
activation='relu',
pad_type='zero'),
SNBlock(in_channels=nf * 4,
out_channels=nf * 4,
kernel_size=kernel_size,
stride=2,
padding=get_pad(im_size // 16, kernel_size, 2),
norm='in',
activation='relu',
pad_type='zero'),
SNBlock(in_channels=nf * 4,
out_channels=nf * 4,
kernel_size=kernel_size,
stride=2,
padding=get_pad(im_size // 32, kernel_size, 2),
norm='in',
activation='relu',
pad_type='zero'),
nn.Flatten(),
nn.Linear(nf * 4 * im_size // 64 * im_size // 64, 512)
])
self.flat = nn.Flatten()
self.edge_dis = nn.Sequential(
SobelFilter(device,
in_nc=3,
filter_c=1,
padding=get_pad(256, 3, 1),
stride=1),
SNBlock(in_channels=2,
out_channels=nf // 2,
kernel_size=kernel_size,
stride=4,
padding=get_pad(im_size, kernel_size, 2),
norm='in',
activation='relu',
pad_type='zero'),
SNBlock(in_channels=nf // 2,
out_channels=nf,
kernel_size=kernel_size,
stride=2,
padding=get_pad(im_size // 4, kernel_size, 2),
norm='in',
activation='relu',
pad_type='zero'),
SNBlock(in_channels=nf,
out_channels=nf * 2,
kernel_size=kernel_size,
stride=2,
padding=get_pad(im_size // 8, kernel_size, 2),
norm='in',
activation='relu',
pad_type='zero'),
SNBlock(in_channels=nf * 2,
out_channels=nf * 4,
kernel_size=kernel_size,
stride=2,
padding=get_pad(im_size // 16, kernel_size, 2),
norm='in',
activation='relu',
pad_type='zero'),
SNBlock(in_channels=nf * 4,
out_channels=nf * 4,
kernel_size=kernel_size,
stride=2,
padding=get_pad(im_size // 32, kernel_size, 2),
norm='in',
activation='relu',
pad_type='zero'),
SNBlock(in_channels=nf * 4,
out_channels=nf * 4,
kernel_size=kernel_size,
stride=2,
padding=get_pad(im_size // 64, kernel_size, 2),
norm='in',
activation='relu',
pad_type='zero'), nn.Flatten(),
nn.Linear(nf * 4 * im_size // 128 * im_size // 128, 512))
self.out = nn.Sequential(_activation('relu'), nn.Linear(1024, 1))
data, labels = next(train_it)
#imshow(torchvision.utils.make_grid(data))
# Resnet
# resnet = models.resnet152(pretrained=True)
# num_ftrs_resnet = resnet.fc.in_features # Number of features before FC
# modules = list(resnet.children())[:-1]
# resnet = nn.Sequential(*modules)
# for p in resnet.parameters():
# p.requires_grad = False
resnet = models.resnet152(pretrained=True)
num_ftrs_resnet = resnet.fc.in_features
for param in resnet.parameters():
param.requires_grad = False
resnet.fc = nn.Flatten()
# Vgg16
vgg16 = models.vgg16(pretrained=True)
vgg16 = vgg16.features
for p in vgg16.parameters():
p.requires_grad = False
num_ftrs_vgg16 = 512 * 7 * 7
# Choose extractor
feature_extractor = resnet
num_ftrs = num_ftrs_resnet
feature_extractor = feature_extractor.to(device)
# summary(feature_extractor, input_size=(TRAIN_BATCH_SIZE, 3, IMAGE_SIZE, IMAGE_SIZE))
import torch
import torch.nn as nn
from torch.utils.data import random_split
from torchvision.datasets import MNIST
from torchvision.transforms import ToTensor
from poutyne import Model
# Instanciate the MNIST dataset
train_valid_dataset = MNIST('./datasets', train=True, download=True, transform=ToTensor())
test_dataset = MNIST('./datasets', train=False, download=True, transform=ToTensor())
train_dataset, valid_dataset = random_split(train_valid_dataset, [50_000, 10_000],
generator=torch.Generator().manual_seed(42))
# Select CUDA device if available
cuda_device = 0
device = torch.device('cuda:%d' % cuda_device if torch.cuda.is_available() else 'cpu')
# Define the network
network = nn.Sequential(
nn.Flatten(),
nn.Linear(28 * 28, 100),
nn.ReLU(),
nn.Linear(100, 10)
)
epochs = 5
# Define the Model and train
model = Model(network, 'sgd', 'cross_entropy', batch_metrics=['accuracy'], epoch_metrics=['f1'], device=device)
model.fit_dataset(train_dataset, valid_dataset, epochs=epochs)
def __init__(self, channels_prev: int, num_classes: int):
super().__init__()
self.pool = nn.AvgPool2d(7)
self.flat = nn.Flatten()
self.fc = nn.Linear(channels_prev, num_classes)
def build_resnet(layers: List[int],
num_classes: int = 1000,
inplace: bool = False
) -> nn.Sequential:
"""Builds a ResNet as a simple sequential model.
Note:
The implementation is copied from :mod:`torchvision.models.resnet`.
"""
inplanes = 64
def make_layer(planes: int,
blocks: int,
stride: int = 1,
inplace: bool = False,
) -> nn.Sequential:
nonlocal inplanes
downsample = None
if stride != 1 or inplanes != planes * 4:
downsample = nn.Sequential(
nn.Conv2d(inplanes, planes * 4,
kernel_size=1, stride=stride, bias=False),
nn.BatchNorm2d(planes * 4),
)
layers = []
layers.append(bottleneck(inplanes, planes, stride, downsample, inplace))
inplanes = planes * 4
for _ in range(1, blocks):
layers.append(bottleneck(inplanes, planes, inplace=inplace))
return nn.Sequential(*layers)
# Build ResNet as a sequential model.
model = nn.Sequential(OrderedDict([
('conv1', nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False)),
('bn1', nn.BatchNorm2d(64)),
('relu', nn.ReLU()),
('maxpool', nn.MaxPool2d(kernel_size=3, stride=2, padding=1)),
('layer1', make_layer(64, layers[0], inplace=inplace)),
('layer2', make_layer(128, layers[1], stride=2, inplace=inplace)),
('layer3', make_layer(256, layers[2], stride=2, inplace=inplace)),
('layer4', make_layer(512, layers[3], stride=2, inplace=inplace)),
('avgpool', nn.AdaptiveAvgPool2d((1, 1))),
('flat', nn.Flatten()),
('fc', nn.Linear(512 * 4, num_classes)),
]))
# Flatten nested sequentials.
model = flatten_sequential(model)
# Initialize weights for Conv2d and BatchNorm2d layers.
# Stolen from torchvision-0.4.0.
def init_weight(m: nn.Module) -> None:
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
return
if isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
return
model.apply(init_weight)
return model
model = NeuralNetwork()
print(model)
X = torch.rand(1, 28, 28)
logits = model(X)
pred_probab = nn.Softmax(dim=1)(logits)
y_pred = pred_probab.argmax(1)
print(f"Predicted class: {y_pred}")
input_image = torch.rand(3, 28, 28)
print(input_image.size())
flatten = nn.Flatten()
flat_image = flatten(input_image)
print(flat_image.size())
layer1 = nn.Linear(in_features=28 * 28, out_features=20)
hidden1 = layer1(flat_image)
print(hidden1.size())
print(f"Before ReLU: {hidden1}\n\n")
hidden1 = nn.ReLU()(hidden1)
print(f"After ReLU: {hidden1}")
seq_modules = nn.Sequential(flatten, layer1, nn.ReLU(), nn.Linear(20, 10))
input_image = torch.rand(3, 28, 28)
logits = seq_modules(input_image)
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--env", type=str, default="BreakoutNoFrameskip-v4")
parser.add_argument(
"--outdir",
type=str,
default="results",
help=("Directory path to save output files."
" If it does not exist, it will be created."),
)
parser.add_argument("--seed",
type=int,
default=0,
help="Random seed [0, 2 ** 31)")
parser.add_argument("--gpu", type=int, default=0)
parser.add_argument("--demo", action="store_true", default=False)
parser.add_argument("--load-pretrained",
action="store_true",
default=False)
parser.add_argument("--pretrained-type",
type=str,
default="best",
choices=["best", "final"])
parser.add_argument("--load", type=str, default=None)
parser.add_argument("--final-exploration-frames", type=int, default=10**6)
parser.add_argument("--final-epsilon", type=float, default=0.01)
parser.add_argument("--eval-epsilon", type=float, default=0.001)
parser.add_argument("--steps", type=int, default=5 * 10**7)
parser.add_argument(
"--max-frames",
type=int,
default=30 * 60 * 60, # 30 minutes with 60 fps
help="Maximum number of frames for each episode.",
)
parser.add_argument("--replay-start-size", type=int, default=5 * 10**4)
parser.add_argument("--target-update-interval", type=int, default=10**4)
parser.add_argument("--eval-interval", type=int, default=250000)
parser.add_argument("--eval-n-steps", type=int, default=125000)
parser.add_argument("--update-interval", type=int, default=4)
parser.add_argument("--batch-size", type=int, default=32)
parser.add_argument(
"--log-level",
type=int,
default=20,
help="Logging level. 10:DEBUG, 20:INFO etc.",
)
parser.add_argument(
"--render",
action="store_true",
default=False,
help="Render env states in a GUI window.",
)
parser.add_argument(
"--monitor",
action="store_true",
default=False,
help=
("Monitor env. Videos and additional information are saved as output files."
),
)
parser.add_argument("--batch-accumulator",
type=str,
default="mean",
choices=["mean", "sum"])
parser.add_argument("--quantile-thresholds-N", type=int, default=64)
parser.add_argument("--quantile-thresholds-N-prime", type=int, default=64)
parser.add_argument("--quantile-thresholds-K", type=int, default=32)
parser.add_argument("--n-best-episodes", type=int, default=200)
args = parser.parse_args()
import logging
logging.basicConfig(level=args.log_level)
# Set a random seed used in PFRL.
utils.set_random_seed(args.seed)
# Set different random seeds for train and test envs.
train_seed = args.seed
test_seed = 2**31 - 1 - args.seed
args.outdir = experiments.prepare_output_dir(args, args.outdir)
print("Output files are saved in {}".format(args.outdir))
def make_env(test):
# Use different random seeds for train and test envs
env_seed = test_seed if test else train_seed
env = atari_wrappers.wrap_deepmind(
atari_wrappers.make_atari(args.env, max_frames=args.max_frames),
episode_life=not test,
clip_rewards=not test,
)
env.seed(int(env_seed))
if test:
# Randomize actions like epsilon-greedy in evaluation as well
env = pfrl.wrappers.RandomizeAction(env, args.eval_epsilon)
if args.monitor:
env = pfrl.wrappers.Monitor(
env, args.outdir, mode="evaluation" if test else "training")
if args.render:
env = pfrl.wrappers.Render(env)
return env
env = make_env(test=False)
eval_env = make_env(test=True)
n_actions = env.action_space.n
q_func = pfrl.agents.iqn.ImplicitQuantileQFunction(
psi=nn.Sequential(
nn.Conv2d(4, 32, 8, stride=4),
nn.ReLU(),
nn.Conv2d(32, 64, 4, stride=2),
nn.ReLU(),
nn.Conv2d(64, 64, 3, stride=1),
nn.ReLU(),
nn.Flatten(),
),
phi=nn.Sequential(
pfrl.agents.iqn.CosineBasisLinear(64, 3136),
nn.ReLU(),
),
f=nn.Sequential(
nn.Linear(3136, 512),
nn.ReLU(),
nn.Linear(512, n_actions),
),
)
# Use the same hyper parameters as https://arxiv.org/abs/1710.10044
opt = torch.optim.Adam(q_func.parameters(),
lr=5e-5,
eps=1e-2 / args.batch_size)
rbuf = replay_buffers.ReplayBuffer(10**6)
explorer = explorers.LinearDecayEpsilonGreedy(
1.0,
args.final_epsilon,
args.final_exploration_frames,
lambda: np.random.randint(n_actions),
)
def phi(x):
# Feature extractor
return np.asarray(x, dtype=np.float32) / 255
agent = pfrl.agents.IQN(
q_func,
opt,
rbuf,
gpu=args.gpu,
gamma=0.99,
explorer=explorer,
replay_start_size=args.replay_start_size,
target_update_interval=args.target_update_interval,
update_interval=args.update_interval,
batch_accumulator=args.batch_accumulator,
phi=phi,
quantile_thresholds_N=args.quantile_thresholds_N,
quantile_thresholds_N_prime=args.quantile_thresholds_N_prime,
quantile_thresholds_K=args.quantile_thresholds_K,
)
if args.load or args.load_pretrained:
# either load or load_pretrained must be false
assert not args.load or not args.load_pretrained
if args.load:
agent.load(args.load)
else:
agent.load(
utils.download_model("IQN",
args.env,
model_type=args.pretrained_type)[0])
if args.demo:
eval_stats = experiments.eval_performance(
env=eval_env,
agent=agent,
n_steps=args.eval_n_steps,
n_episodes=None,
)
print("n_steps: {} mean: {} median: {} stdev {}".format(
args.eval_n_steps,
eval_stats["mean"],
eval_stats["median"],
eval_stats["stdev"],
))
else:
experiments.train_agent_with_evaluation(
agent=agent,
env=env,
steps=args.steps,
eval_n_steps=args.eval_n_steps,
eval_n_episodes=None,
eval_interval=args.eval_interval,
outdir=args.outdir,
save_best_so_far_agent=True,
eval_env=eval_env,
)
dir_of_best_network = os.path.join(args.outdir, "best")
agent.load(dir_of_best_network)
# run 200 evaluation episodes, each capped at 30 mins of play
stats = experiments.evaluator.eval_performance(
env=eval_env,
agent=agent,
n_steps=None,
n_episodes=args.n_best_episodes,
max_episode_len=args.max_frames / 4,
logger=None,
)
with open(os.path.join(args.outdir, "bestscores.json"), "w") as f:
json.dump(stats, f)
print("The results of the best scoring network:")
for stat in stats:
print(str(stat) + ":" + str(stats[stat]))
def train_experiment(device):
with TemporaryDirectory() as logdir:
model = nn.Sequential(nn.Flatten(), nn.Linear(28 * 28, 10))
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=0.02)
loaders = {
"train":
DataLoader(MNIST(os.getcwd(),
train=False,
download=True,
transform=ToTensor()),
batch_size=32),
"valid":
DataLoader(MNIST(os.getcwd(),
train=False,
download=True,
transform=ToTensor()),
batch_size=32),
}
runner = dl.SupervisedRunner(input_key="features",
output_key="logits",
target_key="targets",
loss_key="loss")
# model training
runner.train(
engine=dl.DeviceEngine(device),
model=model,
criterion=criterion,
optimizer=optimizer,
loaders=loaders,
num_epochs=1,
callbacks=[
dl.AccuracyCallback(input_key="logits",
target_key="targets",
topk_args=(1, 3, 5)),
dl.PrecisionRecallF1SupportCallback(input_key="logits",
target_key="targets",
num_classes=10),
dl.AUCCallback(input_key="logits", target_key="targets"),
dl.ConfusionMatrixCallback(input_key="logits",
target_key="targets",
num_classes=10),
],
logdir=logdir,
valid_loader="valid",
valid_metric="loss",
minimize_valid_metric=True,
verbose=False,
load_best_on_end=True,
timeit=False,
check=False,
overfit=False,
fp16=False,
ddp=False,
)
# model inference
for prediction in runner.predict_loader(loader=loaders["valid"]):
assert prediction["logits"].detach().cpu().numpy().shape[-1] == 10
# model post-processing
features_batch = next(iter(loaders["valid"]))[0]
# model stochastic weight averaging
model.load_state_dict(
utils.get_averaged_weights_by_path_mask(logdir=logdir,
path_mask="*.pth"))
# model tracing
utils.trace_model(model=runner.model, batch=features_batch)
# model to onnx
# utils.onnx_export(
# model=runner.model, batch=features_batch, file=f"./{logdir}/mnist.onnx", verbose=False
# )
if SETTINGS.quantization_required:
# model quantization
utils.quantize_model(model=runner.model)
if SETTINGS.pruning_required:
# model pruning
utils.prune_model(model=runner.model,
pruning_fn="l1_unstructured",
amount=0.8)
def _atari_arch(self, input_shape, action_dim, config):
channels = input_shape[0]
layers_encoder = [
nn.Conv2d(channels, 32, kernel_size=7, stride=3, padding=2),
nn.LeakyReLU(),
nn.Conv2d(32, 32, kernel_size=5, stride=3, padding=0),
nn.LeakyReLU(),
nn.Conv2d(32, 64, kernel_size=5, stride=1, padding=0),
nn.LeakyReLU(),
nn.Conv2d(64, 64, kernel_size=3, stride=1, padding=0),
nn.LeakyReLU(),
nn.Conv2d(64, 128, kernel_size=3, stride=1, padding=0),
nn.LeakyReLU(),
nn.Conv2d(128, 128, kernel_size=2, stride=1, padding=0),
nn.Flatten()
]
nn.init.xavier_uniform_(layers_encoder[0].weight)
nn.init.xavier_uniform_(layers_encoder[2].weight)
nn.init.xavier_uniform_(layers_encoder[4].weight)
nn.init.xavier_uniform_(layers_encoder[6].weight)
nn.init.xavier_uniform_(layers_encoder[8].weight)
nn.init.xavier_uniform_(layers_encoder[10].weight)
layers_model = [
Linear(in_features=256,
out_features=config.forward_model_h1,
bias=True),
LeakyReLU(),
Linear(in_features=config.forward_model_h1,
out_features=config.forward_model_h1,
bias=True),
LeakyReLU(),
Linear(in_features=config.forward_model_h1,
out_features=config.forward_model_h2,
bias=True),
LeakyReLU(),
Linear(in_features=config.forward_model_h2,
out_features=config.forward_model_h2,
bias=True),
LeakyReLU(),
Linear(in_features=config.forward_model_h2,
out_features=128,
bias=True)
]
nn.init.xavier_uniform_(layers_model[0].weight)
nn.init.xavier_uniform_(layers_model[2].weight)
nn.init.xavier_uniform_(layers_model[4].weight)
nn.init.xavier_uniform_(layers_model[6].weight)
nn.init.xavier_uniform_(layers_model[8].weight)
return layers_encoder, layers_model
#AE arch
# def _atari_arch(self, input_shape, action_dim, config):
# channels = input_shape[0]
#
# layers_encoder = [
# nn.Conv2d(channels, 64, kernel_size=7, stride=3, padding=2),
# nn.LeakyReLU(),
#
# nn.Conv2d(64, 64, kernel_size=5, stride=3, padding=0),
# nn.LeakyReLU(),
#
# nn.Conv2d(64, 128, kernel_size=5, stride=1, padding=0),
# nn.LeakyReLU(),
#
# nn.Conv2d(128, 128, kernel_size=3, stride=1, padding=0),
# nn.LeakyReLU(),
#
# nn.Conv2d(128, 256, kernel_size=3, stride=1, padding=0),
# nn.LeakyReLU(),
#
# nn.Conv2d(256, 256, kernel_size=2, stride=1, padding=0),
# nn.LeakyReLU(),
#
# nn.Flatten()
# ]
#
# nn.init.xavier_uniform_(layers_encoder[0].weight)
# nn.init.xavier_uniform_(layers_encoder[2].weight)
# nn.init.xavier_uniform_(layers_encoder[4].weight)
# nn.init.xavier_uniform_(layers_encoder[6].weight)
# nn.init.xavier_uniform_(layers_encoder[8].weight)
# nn.init.xavier_uniform_(layers_encoder[10].weight)
#
# layers_model = [
# nn.Linear(512, 256),
# nn.LeakyReLU(),
# nn.Linear(256, 256),
# nn.LeakyReLU(),
# ]
#
# nn.init.xavier_uniform_(layers_model[0].weight)
# nn.init.xavier_uniform_(layers_model[2].weight)
#
# layers_decoder = [
# nn.ConvTranspose2d(256, 256, kernel_size=2, stride=1, padding=0),
# nn.LeakyReLU(),
#
# nn.ConvTranspose2d(256, 128, kernel_size=3, stride=1, padding=0),
# nn.LeakyReLU(),
#
# nn.ConvTranspose2d(128, 128, kernel_size=3, stride=1, padding=0),
# nn.LeakyReLU(),
#
# nn.ConvTranspose2d(128, 64, kernel_size=5, stride=1, padding=0),
# nn.LeakyReLU(),
#
# nn.ConvTranspose2d(64, 64, kernel_size=5, stride=3, padding=0),
# nn.LeakyReLU(),
#
# nn.ConvTranspose2d(64, channels, kernel_size=7, stride=3, padding=2),
# ]
#
# nn.init.xavier_uniform_(layers_decoder[0].weight)
# nn.init.xavier_uniform_(layers_decoder[2].weight)
# nn.init.xavier_uniform_(layers_decoder[4].weight)
# nn.init.xavier_uniform_(layers_decoder[6].weight)
# nn.init.xavier_uniform_(layers_decoder[8].weight)
# nn.init.xavier_uniform_(layers_decoder[10].weight)
#
# return layers_encoder, layers_model, layers_decoder
def __init__(self,
in_channels,
device,
len_sequence,
n_conv_layers=3,
feed_layers=[256, 128],
output_size=None,
out_activation=None):
"""
:param in_channels: number of input channels
:param n_conv_layers: how many convolutional layers (output layer excluded)
:param feed_layers: list of feedforward layers size
:param output_size: None if output does not have to be computed. An integer otherwise. Default None.
:param out_activation: if None last layer is linear, else out_activation is used as output function.
out_activation must be passed in the form torch.nn.Function(args).
If output_size is None this option has no effect. Default None.
"""
super(CNN1DSpeechWords, self).__init__()
self.output_type = OUTPUT_TYPE.ALL_OUTS
self.is_recurrent = False
self.in_channels = in_channels
self.n_conv_layers = n_conv_layers
self.feed_layers = feed_layers
self.len_sequence = len_sequence
self.output_size = output_size
self.device = device
out_chs = [self.in_channels] + \
[ 40*(i+1) for i in range(n_conv_layers) ]
ks = [3**(i + 1) for i in range(n_conv_layers)]
self.out_activation = out_activation
self.layers = nn.ModuleDict()
output_size_conv = self.len_sequence
for i in range(self.n_conv_layers):
output_size_conv = compute_conv_out_shape_1d(
output_size_conv, 0, 1, ks[i], 1) # convolution
output_size_conv = compute_conv_out_shape_1d(
output_size_conv, 0, 1, 2, 1) # pooling
self.layers.update([[
f'conv{i}',
nn.Conv1d(out_chs[i],
out_chs[i + 1],
kernel_size=ks[i],
stride=1)
], [f'relu{i}', nn.ReLU()
], [f'pool{i}',
nn.MaxPool1d(kernel_size=2, stride=1)]])
self.layers.update([['flatten', nn.Flatten()]])
for i, el in enumerate(feed_layers):
if i == 0:
input_size = output_size_conv * out_chs[-1]
else:
input_size = feed_layers[i - 1]
self.layers.update([
[f'l{i}', nn.Linear(input_size, el, bias=True)],
[f'relu_l{i}', nn.ReLU()],
])
if self.output_size is not None:
self.layers.update({
'out':
nn.Linear(feed_layers[-1], self.output_size, bias=True)
})
self.layers = self.layers.to(self.device)
def _init_model(self):
'''
input:
flatten_att_nbhd_inputs: shape=[att_lstm_num * att_lstm_seq_len, (batch_size, nbhd_size, nbhd_size, nbhd_type)]
-> att_nbhd_inputs: shape=[att_lstm_num, att_lstm_seq_len, (batch_size, nbhd_size, nbhd_size, nbhd_type)]
flatten_att_flow_inputs: shape=[att_lstm_num * att_lstm_seq_len, (batch_size, nbhd_size, nbhd_size, flow_type)]
-> att_flow_inputs: shape=[att_lstm_num, att_lstm_seq_len, (batch_size, nbhd_size, nbhd_size, flow_type)]
att_lstm_inputs: shape=[att_lstm_num, (batch_size, att_lstm_seq_len, feature_vec_len)]
nbhd_inputs: shaoe=[lstm_seq_len, (batch_size, nbhd_size, nbhd_size, nbhd_type)]
flow_inputs: shape=[lstm_seq_len, (batch_size, nbhd_size, nbhd_size, flow_type)]
lstm_inputs: shape=(batch_size, lstm_seq_len, feature_vec_len)
remark:
tensor part should have shape of (batch_size, input_channel, H, W), use permute
'''
# 1st level gate
self.nbhd_cnns_1st = nn.ModuleList([
nn.Sequential(
nn.Conv2d(in_channels=self.nbhd_type,
out_channels=64,
kernel_size=(3, 3),
padding=1), nn.ReLU()).to(self.device)
for i in range(self.lstm_seq_len)
])
self.flow_cnns_1st = nn.ModuleList([
nn.Sequential(
nn.Conv2d(in_channels=self.flow_type,
out_channels=64,
kernel_size=(3, 3),
padding=1), nn.ReLU(), nn.Sigmoid()).to(self.device)
for i in range(self.lstm_seq_len)
])
# [nbhd * flow] shape=[lstm_seq_len, (batch_size, 64, nbhd_size, nbhd_size)]
# 2nd level gate
self.nbhd_cnns_2nd = nn.ModuleList([
nn.Sequential(
nn.Conv2d(in_channels=64,
out_channels=64,
kernel_size=(3, 3),
padding=1), nn.ReLU()).to(self.device)
for i in range(self.lstm_seq_len)
])
self.flow_cnns_2nd = nn.ModuleList([
nn.Sequential(
nn.Conv2d(in_channels=self.flow_type,
out_channels=64,
kernel_size=(3, 3),
padding=1), nn.ReLU(), nn.Sigmoid()).to(self.device)
for i in range(self.lstm_seq_len)
])
# [nbhd * flow] shape=[lstm_seq_len, (batch_size, 64, nbhd_size, nbhd_size)]
# 3rd level gate
self.nbhd_cnns_3rd = nn.ModuleList([
nn.Sequential(
nn.Conv2d(in_channels=64,
out_channels=64,
kernel_size=(3, 3),
padding=1), nn.ReLU()).to(self.device)
for i in range(self.lstm_seq_len)
])
self.flow_cnns_3rd = nn.ModuleList([
nn.Sequential(
nn.Conv2d(in_channels=self.flow_type,
out_channels=64,
kernel_size=(3, 3),
padding=1), nn.ReLU(), nn.Sigmoid()).to(self.device)
for i in range(self.lstm_seq_len)
])
# [nbhd * flow] shape=[lstm_seq_len, (batch_size, 64, nbhd_size, nbhd_size)]
# dense part
self.nbhd_vecs = nn.ModuleList([
nn.Sequential(
nn.Flatten(),
nn.Linear(64 * self.nbhd_size * self.nbhd_size,
self.cnn_flat_size), nn.ReLU()).to(self.device)
for i in range(self.lstm_seq_len)
]) # shape=[lstm_seq_len, (batch_size, cnn_flat_size)]
# feature concatenate
# torch.cat(list, dim=-1), shape=(batch_size, cnn_flat_size * lstm_seq_len)
# torch.reshape(tensor, (tensor.shape[0], lstm_seq_len, cnn_flat_size))
# torch.cat(list, dim=-1), shape=(batch_size, lstm_seq_len, feature_vec_len + cnn_flat_size)
# lstm
self.lstm = nn.LSTM(input_size=self.feature_vec_len +
self.cnn_flat_size,
hidden_size=self.lstm_out_size,
batch_first=True,
dropout=0.1).to(self.device)
# result shape=(batch_size, lstm_seq_len, lstm_out_size)
# result, (hn, cn) = lstm -> hn[-1] shape=(batch, lstm_out_size)
# attention part
self.att_nbhd_cnns_1st = nn.ModuleList([
nn.ModuleList([
nn.Sequential(
nn.Conv2d(in_channels=self.nbhd_type,
out_channels=64,
kernel_size=(3, 3),
padding=1), nn.ReLU()).to(self.device)
for j in range(self.att_lstm_seq_len)
]) for i in range(self.att_lstm_num)
])
self.att_flow_cnns_1st = nn.ModuleList([
nn.ModuleList([
nn.Sequential(
nn.Conv2d(in_channels=self.flow_type,
out_channels=64,
kernel_size=(3, 3),
padding=1), nn.ReLU()).to(self.device)
for j in range(self.att_lstm_seq_len)
]) for i in range(self.att_lstm_num)
])
self.att_flow_gate_1st = nn.ModuleList([
nn.ModuleList([
nn.Sigmoid().to(self.device)
for j in range(self.att_lstm_seq_len)
]) for i in range(self.att_lstm_num)
])
# [[nbhd * flow]] shape=[att_lstm_num, att_lstm_seq_len, (batch_size, 64, nbhd_size, nbhd_size)]
self.att_nbhd_cnns_2nd = nn.ModuleList([
nn.ModuleList([
nn.Sequential(
nn.Conv2d(in_channels=64,
out_channels=64,
kernel_size=(3, 3),
padding=1), nn.ReLU()).to(self.device)
for j in range(self.att_lstm_seq_len)
]) for i in range(self.att_lstm_num)
])
self.att_flow_cnns_2nd = nn.ModuleList([
nn.ModuleList([
nn.Sequential(
nn.Conv2d(in_channels=64,
out_channels=64,
kernel_size=(3, 3),
padding=1), nn.ReLU()).to(self.device)
for j in range(self.att_lstm_seq_len)
]) for i in range(self.att_lstm_num)
])
self.att_flow_gate_2nd = nn.ModuleList([
nn.ModuleList([
nn.Sigmoid().to(self.device)
for j in range(self.att_lstm_seq_len)
]) for i in range(self.att_lstm_num)
])
# [[nbhd * flow]] shape=[att_lstm_num, att_lstm_seq_len, (batch_size, 64, nbhd_size, nbhd_size)]
self.att_nbhd_cnns_3rd = nn.ModuleList([
nn.ModuleList([
nn.Sequential(
nn.Conv2d(in_channels=64,
out_channels=64,
kernel_size=(3, 3),
padding=1), nn.ReLU()).to(self.device)
for j in range(self.att_lstm_seq_len)
]) for i in range(self.att_lstm_num)
])
self.att_flow_cnns_3rd = nn.ModuleList([
nn.ModuleList([
nn.Sequential(
nn.Conv2d(in_channels=64,
out_channels=64,
kernel_size=(3, 3),
padding=1), nn.ReLU()).to(self.device)
for j in range(self.att_lstm_seq_len)
]) for i in range(self.att_lstm_num)
])
self.att_flow_gate_3rd = nn.ModuleList([
nn.ModuleList([
nn.Sigmoid().to(self.device)
for j in range(self.att_lstm_seq_len)
]) for i in range(self.att_lstm_num)
])
# [[nbhd * flow]] shape=[att_lstm_num, att_lstm_seq_len, (batch_size, 64, nbhd_size, nbhd_size)]
self.att_nbhd_vecs = nn.ModuleList([
nn.ModuleList([
nn.Sequential(
nn.Flatten(),
nn.Linear(64 * self.nbhd_size * self.nbhd_size,
self.cnn_flat_size), nn.ReLU()).to(self.device)
for j in range(self.att_lstm_seq_len)
]) for i in range(self.att_lstm_num)
])
# shape=[att_lstm_num, att_lstm_seq_len, (batch_size, cnn_flat_size)]
# [torch.cat(list, dim=-1)], shape=[att_lstm_num, (batch_size, cnn_flat_size * att_lstm_seq_len)]
# [torch.reshape(tensor, (tensor.shape[0], att_lstm_seq_len, cnn_flat_size))]
# [torch.cat(list, dim=-1)], shape=[att_lstm_num, (batch_size, att_lstm_seq_len, feature_vec_len + cnn_flat_size)]
self.att_lstms = nn.ModuleList([
nn.LSTM(input_size=self.feature_vec_len + self.cnn_flat_size,
hidden_size=self.lstm_out_size,
batch_first=True,
dropout=0.1).to(self.device)
for i in range(self.att_lstm_num)
])
# [result] shape=[att_lstm_num, (batch_size, lstm_seq_len, lstm_out_size)]
# [result, (hn, cn) = lstm -> result] [att_lstm_num, (batch_size, lstm_seq_len, lstm_out_size)]
# compare
self.att_low_level = nn.ModuleList([
CBAAttention(self.device, self.lstm_out_size, self.lstm_out_size)
for i in range(self.att_lstm_num)
])
# shape=[att_lstm_num, (batch_size, lstm_out_size)]
# torch.cat(list, dim=-1), shape=(batch_size, lstm_out_size * att_lstm_num)
# torch.reshape(tensor, (tensor.shape[0], att_lstm_num, lstm_out_size))
# shape=(batch_size, att_lstm_num, lstm_out_size)
self.att_high_level = nn.LSTM(input_size=self.lstm_out_size,
hidden_size=self.lstm_out_size,
batch_first=True,
dropout=0.1).to(self.device)
# result shape=(batch_size, att_lstm_num, lstm_out_size)
# result, (hn, cn) = lstm -> hn[-1] shape=(batch_size, lstm_out_size)
self.lstm_all = nn.Linear(self.lstm_out_size + self.lstm_out_size,
self.output_dim).to(self.device)
self.pred_volume = nn.Tanh().to(self.device)
def __init__(self):
super().__init__()
self.classifier = nn.Sequential(nn.Flatten(), nn.Linear(12544, 10))
def __init__(self, hidden_size=128, feature_size=40):
super(GeneratorModel, self).__init__()
self.hidden_size = hidden_size
self.feature_size = feature_size
self.total_size = self.hidden_size + self.feature_size
self.ones = torch.ones([1, 1, 64, 64]).to('cuda')
self.encoder = nn.Sequential(
nn.Conv2d(in_channels=3 + self.feature_size,
out_channels=32,
kernel_size=4,
stride=2,
padding=1), # 32 * 32 * 32
nn.BatchNorm2d(32),
nn.GELU(),
nn.Conv2d(in_channels=32,
out_channels=64,
kernel_size=4,
stride=2,
padding=1), # 64 * 16 * 16
nn.BatchNorm2d(64),
nn.GELU(),
nn.Conv2d(in_channels=64,
out_channels=128,
kernel_size=4,
stride=2,
padding=1), # 128 * 8 * 8
nn.BatchNorm2d(128),
nn.GELU(),
nn.Conv2d(in_channels=128,
out_channels=256,
kernel_size=4,
stride=2,
padding=1), # 256 * 4 * 4
nn.BatchNorm2d(256),
nn.GELU(),
nn.Conv2d(in_channels=256,
out_channels=512,
kernel_size=4,
stride=2,
padding=1), # 512 * 2 * 2
nn.BatchNorm2d(512),
nn.GELU(),
nn.Conv2d(in_channels=512,
out_channels=1024,
kernel_size=4,
stride=2,
padding=1), # 1024 * 1 * 1
nn.BatchNorm2d(1024),
nn.GELU(),
nn.Flatten(),
nn.Linear(in_features=1024, out_features=self.hidden_size * 2))
self.decoder = nn.Sequential(
nn.Linear(in_features=self.total_size, out_features=1024),
nn.Unflatten(dim=1, unflattened_size=(1024, 1,
1)), # batch x 1024 x 1 x 1
nn.GELU(),
nn.ConvTranspose2d(in_channels=1024,
out_channels=512,
stride=2,
kernel_size=4,
padding=1), # batch x 512 x 2 x 2
nn.BatchNorm2d(512),
nn.GELU(),
nn.ConvTranspose2d(in_channels=512,
out_channels=256,
stride=2,
kernel_size=4,
padding=1), # batch x 256 x 4 x 4
nn.BatchNorm2d(256),
nn.GELU(),
nn.ConvTranspose2d(in_channels=256,
out_channels=128,
stride=2,
kernel_size=4,
padding=1), # batch x 128 x 8 x 8
nn.BatchNorm2d(128),
nn.GELU(),
nn.ConvTranspose2d(in_channels=128,
out_channels=64,
stride=2,
kernel_size=4,
padding=1), # batch x 64 x 16 x 16
nn.BatchNorm2d(64),
nn.GELU(),
nn.ConvTranspose2d(in_channels=64,
out_channels=32,
stride=2,
kernel_size=4,
padding=1), # batch x 32 x 32 x 32
nn.BatchNorm2d(32),
nn.GELU(),
nn.ConvTranspose2d(in_channels=32,
out_channels=3,
stride=2,
kernel_size=4,
padding=1), # batch x 3 x 64 x 64
nn.Sigmoid())
def lin_1(input_dim=3072, num_classes=10):
model = nn.Sequential(nn.Flatten(), nn.Linear(input_dim, num_classes))
return model
def __init__(self,
image_channels,
num_classes,
batch_normalization: bool = False,
drop_out: bool = False,
conv_stride: bool = False,
init_weights: bool = False):
"""
Is called when model is initialized.
Args:
image_channels. Number of color channels in image (3)
num_classes: Number of classes we want to predict (10)
"""
super().__init__()
num_filters = 32 # Set number of filters in first conv layer
self.num_classes = num_classes
# Define the convolutional layers
self.feature_extractor = nn.Sequential(
# Layer 1
nn.Conv2d(in_channels=image_channels,
out_channels=num_filters,
kernel_size=5,
stride=1,
padding=2),
nn.ReLU(),
nn.BatchNorm2d(num_features=num_filters)
if batch_normalization else nn.Identity(),
nn.Conv2d(
num_filters, num_filters, kernel_size=2, stride=2, padding=0)
if conv_stride else nn.MaxPool2d(kernel_size=2, stride=2),
nn.Dropout() if drop_out else nn.Identity(),
# Layer 2
nn.Conv2d(num_filters, num_filters * 2, 5, 1, 2),
nn.ReLU(),
nn.BatchNorm2d(num_filters *
2) if batch_normalization else nn.Identity(),
nn.Conv2d(num_filters * 2,
num_filters * 2,
kernel_size=2,
stride=2,
padding=0) if conv_stride else nn.MaxPool2d(2, 2),
nn.Dropout() if drop_out else nn.Identity(),
# Layer 3
nn.Conv2d(num_filters * 2, num_filters * 4, 5, 1, 2),
nn.ReLU(),
nn.BatchNorm2d(num_filters *
4) if batch_normalization else nn.Identity(),
nn.Conv2d(num_filters * 4,
num_filters * 4,
kernel_size=2,
stride=2,
padding=0) if conv_stride else nn.MaxPool2d(2, 2),
nn.Dropout() if drop_out else nn.Identity(),
# Flatten
nn.Flatten())
self.num_output_features = num_filters * 4 * 4 * 4
self.classifier = nn.Sequential(
nn.Linear(self.num_output_features, 64),
nn.ReLU(),
nn.BatchNorm1d(64) if batch_normalization else nn.Identity(),
nn.Linear(64, num_classes),
)
if init_weights:
self.feature_extractor.apply(self.initialize_weights)
self.classifier.apply(self.initialize_weights)
import torch.nn as nn
from . import LATENT_SIZE
discriminator = nn.Sequential(
nn.Conv2d(3, 64, kernel_size=4, stride=2, padding=1, bias=False),
nn.BatchNorm2d(64), nn.LeakyReLU(0.2, inplace=True),
nn.Conv2d(64, 128, kernel_size=4, stride=2, padding=1, bias=False),
nn.BatchNorm2d(128), nn.LeakyReLU(0.2, inplace=True),
nn.Conv2d(128, 256, kernel_size=4, stride=2, padding=1, bias=False),
nn.BatchNorm2d(256), nn.LeakyReLU(0.2, inplace=True),
nn.Conv2d(256, 512, kernel_size=4, stride=2, padding=1, bias=False),
nn.BatchNorm2d(512), nn.LeakyReLU(0.2, inplace=True),
nn.Conv2d(512, 1, kernel_size=4, stride=1, padding=0, bias=False),
nn.Flatten(), nn.Sigmoid())
generator = nn.Sequential(
nn.ConvTranspose2d(LATENT_SIZE,
512,
kernel_size=4,
stride=1,
padding=0,
bias=False), nn.BatchNorm2d(512), nn.ReLU(True),
nn.ConvTranspose2d(512,
256,
kernel_size=4,
stride=2,
padding=1,
bias=False), nn.BatchNorm2d(256), nn.ReLU(True),
nn.ConvTranspose2d(256,
128,