Readme Update und Residuals GN Training

This commit is contained in:
Steffen Illium 2022-02-03 12:21:17 +01:00
parent 3fe4f49bca
commit 6b1efd0c49
3 changed files with 59 additions and 22 deletions

View File

@ -31,10 +31,15 @@
### Tasks für Steffen:
- [ ] Training mit kleineren GNs -| Running
- [x] Training mit kleineren GNs
- Accuracy leidet enorm (_0.56_)
![image info](./figures/training_lineplot.png)
- Es entstehen mehr SRNN
- Der Dropout Effekt wird stärker (diff_ohne_SRNN = _0.0_)
![image info](./figures/dropout_stacked_barplot.png)
- [ ] Weiter Trainieren -> 500 Epochs?
- [ ] Loss Gewichtung anpassen
- [ ] Training ohne Residual Skip Connection
- [ ] Training ohne Residual Skip Connection | - Running
- [ ] Test mit Baseline Dense Network
- [ ] mit vergleichbaren Neuron Count
- [ ] mit gesamt Weight Count

View File

@ -133,7 +133,7 @@ def checkpoint_and_validate(model, out_path, epoch_n, final_model=False):
def plot_training_result(path_to_dataframe):
# load from Drive
df = pd.read_csv(path_to_dataframe, index_col=0)
df = pd.read_csv(path_to_dataframe, index_col=False)
# Set up figure
fig, ax1 = plt.subplots() # initializes figure and plots
@ -163,6 +163,9 @@ def plot_training_result(path_to_dataframe):
else:
plt.savefig(Path(path_to_dataframe.parent / 'training_lineplot.png'), dpi=300)
def flat_for_store(parameters):
return (x.item() for y in parameters for x in y.detach().flatten())
if __name__ == '__main__':
@ -175,7 +178,7 @@ if __name__ == '__main__':
data_path = Path('data')
data_path.mkdir(exist_ok=True, parents=True)
run_path = Path('output') / 'mn_st_smaller'
run_path = Path('output') / 'mn_st_NoRes'
model_path = run_path / '0000_trained_model.zip'
df_store_path = run_path / 'train_store.csv'
weight_store_path = run_path / 'weight_store.csv'
@ -189,14 +192,14 @@ if __name__ == '__main__':
d = DataLoader(dataset, batch_size=BATCHSIZE, shuffle=True, drop_last=True, num_workers=WORKER)
interface = np.prod(dataset[0][0].shape)
metanet = MetaNet(interface, depth=5, width=6, out=10).to(DEVICE)
metanet = MetaNet(interface, depth=5, width=6, out=10, residual_skip=False).to(DEVICE)
meta_weight_count = sum(p.numel() for p in next(metanet.particles).parameters())
loss_fn = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(metanet.parameters(), lr=0.008, momentum=0.9)
train_store = new_storage_df('train', None)
weight_store = new_storage_df('train', meta_weight_count)
weight_store = new_storage_df('weights', meta_weight_count)
for epoch in tqdm(range(EPOCH), desc='MetaNet Train - Epochs'):
is_validation_epoch = epoch % VALIDATION_FRQ == 0 if not debug else True
is_self_train_epoch = epoch % SELF_TRAIN_FRQ == 0 if not debug else True
@ -254,29 +257,30 @@ if __name__ == '__main__':
step_log = dict(Epoch=int(epoch), Batch=BATCHSIZE, Metric=key, Score=value)
train_store.loc[train_store.shape[0]] = step_log
for particle in metanet.particles:
weight_log = (epoch, particle.name, *(x for y in particle.parameters() for x in y))
train_store.to_csv(df_store_path, mode='a', header=not df_store_path.exists())
weight_store.to_csv(weight_store_path, mode='a', header=not weight_store_path.exists())
weight_log = (epoch, particle.name, *flat_for_store(particle.parameters()))
weight_store.loc[weight_store.shape[0]] = weight_log
train_store.to_csv(df_store_path, mode='a', header=not df_store_path.exists(), index=False)
weight_store.to_csv(weight_store_path, mode='a', header=not weight_store_path.exists(), index=False)
train_store = new_storage_df('train', None)
weight_store = new_storage_df('train', meta_weight_count)
weight_store = new_storage_df('weights', meta_weight_count)
metanet.eval()
accuracy = checkpoint_and_validate(metanet, run_path, EPOCH, final_model=True)
validation_log = dict(Epoch=EPOCH, Batch=BATCHSIZE,
Metric='Test Accuracy', Score=accuracy.item())
for particle in metanet.particles:
weight_log = (EPOCH, particle.name, *(x for y in particle.parameters() for x in y))
weight_log = (EPOCH, particle.name, *(flat_for_store(particle.parameters())))
weight_store.loc[weight_store.shape[0]] = weight_log
train_store.loc[train_store.shape[0]] = validation_log
train_store.to_csv(df_store_path, mode='a', header=not df_store_path.exists())
weight_store.to_csv(weight_store_path, mode='a', header=not weight_store_path.exists())
train_store.to_csv(df_store_path, mode='a', header=not df_store_path.exists(), index=False)
weight_store.to_csv(weight_store_path, mode='a', header=not weight_store_path.exists(), index=False)
if plotting:
plot_training_result(df_store_path)
if particle_analysis:
model_path = next(run_path.glob(f'*e{EPOCH}.tp'))
model_path = next(run_path.glob(f'*e100.tp'))
latest_model = torch.load(model_path, map_location=DEVICE).eval()
counter_dict = defaultdict(lambda: 0)
_ = test_for_fixpoints(counter_dict, list(latest_model.particles))
@ -284,10 +288,38 @@ if __name__ == '__main__':
zero_ident = torch.load(model_path, map_location=DEVICE).eval().replace_with_zero('identity_func')
zero_other = torch.load(model_path, map_location=DEVICE).eval().replace_with_zero('other_func')
if as_sparse_network_test:
acc_pre = validate(model_path, ratio=1)
acc_pre = validate(model_path, ratio=0.01).item()
ident_ckpt = set_checkpoint(zero_ident, model_path.parent, -1, final_model=True)
ident_acc_post = validate(ident_ckpt, ratio=1)
ident_acc_post = validate(ident_ckpt, ratio=0.01).item()
tqdm.write(f'Zero_ident diff = {abs(ident_acc_post-acc_pre)}')
other_ckpt = set_checkpoint(zero_other, model_path.parent, -2, final_model=True)
other_acc_post = validate(other_ckpt, ratio=1)
other_acc_post = validate(other_ckpt, ratio=0.01).item()
tqdm.write(f'Zero_other diff = {abs(other_acc_post - acc_pre)}')
if plotting:
plt.clf()
fig, ax = plt.subplots(ncols=2)
data = [acc_pre, ident_acc_post, other_acc_post]
labels = ['Full Network', 'Sparse, No Identity', 'Sparse, No Other']
for idx, (score, name) in enumerate(zip(data, labels)):
l = sns.barplot(y=[score], x=['Networks'], color=sns.color_palette()[idx], label=name, ax=ax[0])
# noinspection PyUnboundLocalVariable
for idx, patch in enumerate(l.patches):
if idx != 0:
# we recenter the bar
patch.set_x(patch.get_x() + idx * 0.035)
ax[0].set_title('Accuracy after particle dropout')
ax[0].set_xlabel('Accuracy')
# ax[0].legend()
counter_dict['full_network'] = sum(counter_dict.values())
ax[1].pie(counter_dict.values(), labels=counter_dict.keys(), colors=sns.color_palette()[:3], )
ax[1].set_title('Particle Count for ')
# ax[1].set_xlabel('')
plt.tight_layout()
if debug:
plt.show()
else:
plt.savefig(Path(run_path / 'dropout_stacked_barplot.png'), dpi=300)

View File

@ -296,7 +296,7 @@ class MetaCell(nn.Module):
self.name = name
self.interface = interface
self.weight_interface = 5
self.net_hidden_size = 3
self.net_hidden_size = 4
self.net_ouput_size = 1
self.meta_weight_list = nn.ModuleList()
self.meta_weight_list.extend(
@ -371,7 +371,7 @@ class MetaLayer(nn.Module):
class MetaNet(nn.Module):
def __init__(self, interface=4, depth=3, width=4, out=1, activation=None):
def __init__(self, interface=4, depth=3, width=4, out=1, activation=None, residual_skip=True):
super().__init__()
self.activation = activation
self.out = out
@ -382,14 +382,14 @@ class MetaNet(nn.Module):
self._meta_layer_list = nn.ModuleList()
self._meta_layer_list.append(MetaLayer(name=f'L{0}',
interface=self.interface,
width=self.width)
width=self.width, residual_skip=residual_skip)
)
self._meta_layer_list.extend([MetaLayer(name=f'L{layer_idx + 1}',
interface=self.width, width=self.width
interface=self.width, width=self.width, residual_skip=residual_skip
) for layer_idx in range(self.depth - 2)]
)
self._meta_layer_list.append(MetaLayer(name=f'L{len(self._meta_layer_list)}',
interface=self.width, width=self.out)
interface=self.width, width=self.out, residual_skip=residual_skip)
)
def replace_with_zero(self, ident_key):