Skip to content

Commit

Permalink
Merge branch 'dev' of https://github.com/SchotmanG/deepLearning into dev
Browse files Browse the repository at this point in the history
  • Loading branch information
@TaekedeHaan
TaekedeHaan committed Jun 17, 2018
2 parents 22d0202 + ab3435a commit 0d74669
Show file tree
Hide file tree
Showing 2 changed files with 142 additions and 29 deletions.
100 changes: 91 additions & 9 deletions experiments/final.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -11,20 +11,59 @@
filenames_valid = ['2016']
features_train = ['air_temperature']
features_predict = ['air_temperature']
regularization = None
"""
Experiment I: Nodes per layer
"""
n_nodes = [0.2, 0.5, 1.0, 1.5, 2.0, 2.5, 3.0]
act = 'relu'
n_memory = 1
n_dense_pre = 3
n_dense_pos = 3
memory_unit = 'lstm'
for n_node in n_nodes:
model = create_model(batch_size=batch_size,
t_train=t_train_h,
t_pred=int(t_pred_d * 24 / t_pred_resolution_h),
n_features_train=len(features_train),
n_stations=n_stations,
memory_unit=memory_unit,
width=n_dense_pre,
n_layers_memory=n_memory,
n_layers_preprocessing=n_dense_pre,
n_layers_postprocessing=n_dense_pos,
n_features_pred=len(features_predict),
activation=act,
mask_value=MASK_VALUE,
regularization=regularization)

# '{layer_pre}x{n_nodes}*{act}->[{memory}]{n_lstm}->{layer_pos}{n_nodes}*{act}'
log_dir = 'out/{}-{}-{}-{}{}-{}-{}-{}'.format(n_dense_pre, int(n_node * 10), act, memory_unit, n_memory,
n_dense_pos, int(n_stations / n_dense_pre), act)

train(radius=radius,
batch_size=batch_size,
log_dir=log_dir,
t_train_h=t_train_h,
t_pred_d=t_pred_d,
t_pred_resolution_h=t_pred_resolution_h,
model_name=model,
filenames_train=filenames_train,
filenames_valid=filenames_valid,
features_train=features_train,
features_predict=features_predict,
)

"""
Experiment I: Final
Experiment II: Memory Depth
"""
n_dense_pres = [1, 2, 3]
n_lstms = [1, 2, 4, 8]
n_node = 1.0
act = 'relu'
n_memory = 2
n_dense_poss = [1, 2, 3]
n_dense_pre = 3
n_dense_pos = 3
memory_unit = 'lstm'
reg = None
for i in range(3):
n_dense_pos = n_dense_poss[i]
n_dense_pre = n_dense_pres[i]
for n_memory in n_lstms:
model = create_model(batch_size=batch_size,
t_train=t_train_h,
t_pred=int(t_pred_d * 24 / t_pred_resolution_h),
Expand All @@ -38,7 +77,7 @@
n_features_pred=len(features_predict),
activation=act,
mask_value=MASK_VALUE,
regularization=None)
regularization=regularization)

# '{layer_pre}x{n_nodes}*{act}->[{memory}]{n_lstm}->{layer_pos}{n_nodes}*{act}'
log_dir = 'out/{}-{}-{}-{}{}-{}-{}-{}'.format(n_dense_pre, int(n_node * 10), act, memory_unit, n_memory,
Expand All @@ -56,3 +95,46 @@
features_train=features_train,
features_predict=features_predict,
)

"""
Experiment III: Regularization
"""
n_nodes = 1.0
act = 'relu'
regs = ['Dropout50', 'Dropout25', 'l1', 'l2']
n_memory = 2
n_dense_pre = 3
n_dense_pos = 3
memory_unit = 'lstm'
for reg in regs:
model = create_model(batch_size=batch_size,
t_train=t_train_h,
t_pred=int(t_pred_d * 24 / t_pred_resolution_h),
n_features_train=len(features_train),
n_stations=n_stations,
memory_unit=memory_unit,
width=n_dense_pre,
n_layers_memory=n_memory,
n_layers_preprocessing=n_dense_pre,
n_layers_postprocessing=n_dense_pos,
n_features_pred=len(features_predict),
activation=act,
mask_value=MASK_VALUE,
regularization=reg)

# '{layer_pre}x{n_nodes}*{act}->[{memory}]{n_lstm}->{layer_pos}{n_nodes}*{act}'
log_dir = 'out/{}-{}-{}-{}{}-{}-{}-{}'.format(n_dense_pre, int(n_node * 10), act, memory_unit, n_memory,
n_dense_pos, int(n_stations / n_dense_pre), act)

train(radius=radius,
batch_size=batch_size,
log_dir=log_dir,
t_train_h=t_train_h,
t_pred_d=t_pred_d,
t_pred_resolution_h=t_pred_resolution_h,
model_name=model,
filenames_train=filenames_train,
filenames_valid=filenames_valid,
features_train=features_train,
features_predict=features_predict,
)
71 changes: 51 additions & 20 deletions plots/plot_architecture.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -3,30 +3,61 @@

from downloadData.functions.file_utils import load_file

model_names = ['1-10-relu-lstm1-3-41-relu', '2-10-relu-lstm1-3-20-relu', '3-2-relu-lstm1-3-13-relu',
csfont = {'fontname': 'Century Gothic'}
std = 10.67839587
post = ['3-10-relu-lstm1-1-55-relu', '3-10-relu-lstm1-2-55-relu' ,'3-10-relu-lstm1-4-55-relu',
'3-10-relu-lstm1-8-55-relu']
pre = ['1-10-relu-lstm1-1-165-relu', '2-10-relu-lstm1-2-82-relu', '4-10-relu-lstm1-3-41-relu',
'8-10-relu-lstm1-8-20-relu']
memory = ['3-10-relu-lstm1-3-13-relu', '3-10-relu-lstm2-3-13-relu', '3-10-relu-lstm4-3-13-relu',
'3-10-relu-lstm8-3-13-relu']
layers = np.array([1, 2, 4, 8])
plt.subplot(2, 1, 1)
performances = []
for i, r in enumerate(post):
mat = np.array(load_file('../out/' + r + '/log.csv'))
valid_loss = mat[1:, 2].astype(np.float32)
y = np.min(valid_loss)*std**2
performances.append(y)
plt.bar(layers, performances, width=0.2, color='b', align='center')
performances = []
for i, r in enumerate(pre):
mat = np.array(load_file('../out/' + r + '/log.csv'))
valid_loss = mat[1:, 2].astype(np.float32)
y = np.min(valid_loss)*std**2
performances.append(y)
plt.bar(layers - 0.2, performances, width=0.2, color='g', align='center')
layers = np.array([1, 2, 4, 8])
performances = []
for i, r in enumerate(memory):
mat = np.array(load_file('../out/' + r + '/log.csv'))
valid_loss = mat[1:, 2].astype(np.float32)
y = np.min(valid_loss)*std**2
performances.append(y)
plt.bar(layers + 0.2, performances, width=0.2, color='r', align='center')
plt.legend(['Postprocessing', 'Preprocessing', 'Memory'], fontsize=16, prop={'family': 'Century Gothic'})
plt.xlabel('Layers',fontsize=16, **csfont)
plt.ylabel('Mean Squared Error [$^\circ$C$^2$]', fontsize=16, **csfont)
plt.grid('on')
plt.title('Evaluation of Model Architecture in Terms of Validation Loss ', fontsize=20, **csfont)
plt.subplot(2, 1, 2)

width = [0.2, 0.5, 1.0, 1.5, 2.0, 2.5, 3.0]
model_names = ['3-2-relu-lstm1-3-13-relu',
'3-5-relu-lstm1-3-13-relu',
'3-10-relu-lstm1-1-13-relu', '3-10-relu-lstm1-2-13-relu', '3-10-relu-lstm1-4-13-relu',
'3-10-relu-lstm1-8-13-relu', '3-10-relu-lstm2-3-13-relu', '3-10-relu-lstm4-3-13-relu',
'3-10-relu-lstm8-3-13-relu', '3-15-relu-lstm1-3-13-relu', '3-20-relu-lstm1-3-13-relu',
'3-25-relu-lstm1-3-13-relu', '3-30-relu-lstm1-3-13-relu', '4-10-relu-lstm1-3-10-relu',
'8-10-relu-lstm1-3-5-relu']
legend = ['pre1-mem1-post3', 'pre2-mem1-post3', 'pre2-mem1-post3','pre3-mem1-post3', 'pre3-mem1-post1', 'pre3-mem1-post2',
'pre3-mem1-post4',
'pre3-mem1-post8', 'pre3-mem2-post3', 'pre3-mem4-post3', 'pre3-mem8-post3', 'pre3-mem1-post3',
'pre3-mem1-post3',
'pre3-mem1-post3', 'pre3-mem1-post3', 'pre4-mem1-post3', 'pre8-mem1-post3']
legend = model_names
layers = [5, 6, 6,7,5, 6, 8, 12, 8, 10, 14, 7, 7, 7, 7, 8, 12]
'3-10-relu-lstm1-3-13-relu',
'3-15-relu-lstm1-3-13-relu',
'3-20-relu-lstm1-3-13-relu',
'3-25-relu-lstm1-3-13-relu', '3-30-relu-lstm1-3-13-relu']
performances = []
for i, r in enumerate(model_names):
mat = np.array(load_file('../out/' + r + '/log.csv'))
valid_loss = mat[1:, 2].astype(np.float32)
x = layers[i]
y = np.min(valid_loss)
plt.plot(x, y, '*')
plt.annotate(legend[i], xy=(x, y))
y = np.min(valid_loss)*std**2
performances.append(y)
plt.plot(width, performances, '--*')

plt.xlabel('Layers')
plt.ylabel('Normalized Mean Squared Error [°C^2]')
plt.xlabel('*N_Nodes', fontsize=16, **csfont)
plt.ylabel('Mean Squared Error [$^\circ$C$^2$]', fontsize=16, **csfont)
plt.grid('on')
plt.title('Evaluation of Model Architecture in Terms of Validation Loss')
plt.show()

0 comments on commit 0d74669

Please sign in to comment.