Fix some stuff

run.sh

@@ -47,13 +47,13 @@ fi
 
 ngrok http "$VIS_PORT" > /dev/null &
 echo "> Tunneling URL with Ngrok"
+sleep 3
 
 WEB_HOOK_URL=$(curl --silent --connect-timeout 10 https://localhost:4040/api/tunnels | \
  pipenv run python -c "import json,sys;obj=json.load(sys.stdin);print(obj['tunnels'][0]['public_url'])")
 
 echo "Visdom tunneled in $WEB_HOOK_URL"
 
-
 # RUN MODEL
 
 pipenv run python src/main.py --viz

src/constants.py

@@ -36,19 +36,19 @@
 SAMPLE_TIMES = 100
 
 # Generator
-LR_G = 0.4
+LR_G = 0.1
 LR_PAT_G = 10
 L2_G = 0.25
 HIDDEN_DIM_G = 150
 BIDIRECTIONAL_G = False
 TYPE_G = 'LSTM'
-LAYERS_G = 3
+LAYERS_G = 1
 
 # Discriminator
-LR_D = 0.4
+LR_D = 0.1
 LR_PAT_D = 10
 L2_D = 0.25
 HIDDEN_DIM_D = 150
 BIDIRECTIONAL_D = True
 TYPE_D = 'LSTM'
-LAYERS_D = 3
+LAYERS_D = 1

src/model/train.py

@@ -11,7 +11,7 @@
 from model.GANGenerator import GANGenerator
 from model.GANModel import GANModel
 from constants import EPOCHS, NUM_NOTES, CKPT_STEPS, CHECKPOINTS_PATH, SAMPLE_STEPS, FLAGS, PLOT_COL
-from utils.tensors import device
+from utils.tensors import device, zeros_target, ones_target
 
 
 class VisdomLinePlotter:
@@ -144,7 +144,7 @@ def _train_generator(self, data):
  prediction = self.model.discriminator(fake_data)
 
  # Calculate gradients w.r.t parameters and backpropagate
- loss = self.model.g_criterion(prediction, torch.ones(batch_size).to(device))
+ loss = self.model.g_criterion(prediction, ones_target((batch_size,)))
  loss.backward()
 
  # Update parameters
@@ -163,14 +163,14 @@ def _train_discriminator(self, real_data) -> float:
 
  # Train on real data
  real_predictions = self.model.discriminator(real_data)
- real_loss = self.model.d_criterion(real_predictions, torch.ones(batch_size).to(device))
+ real_loss = self.model.d_criterion(real_predictions, ones_target((batch_size,)))
  real_loss.backward()
 
  # Train on fake data
  noise_data = GANGenerator.noise((batch_size, time_steps, NUM_NOTES))
  fake_data = self.model.generator(noise_data).detach()
  fake_predictions = self.model.discriminator(fake_data)
- fake_loss = self.model.d_criterion(fake_predictions, torch.zeros(batch_size).to(device))
+ fake_loss = self.model.d_criterion(fake_predictions, zeros_target((batch_size,)))
  fake_loss.backward()
 
  # Update parameters

src/utils/tensors.py

@@ -1,4 +1,15 @@
+from typing import Tuple
+
+import torch
 from torch import cuda, device as dev
 
 use_cuda = cuda.is_available()
 device = dev('cuda' if use_cuda else 'cpu')
+
+
+def zeros_target(dims: Tuple):
+ return torch.zeros(size=dims).to(device)
+
+
+def ones_target(dims: Tuple):
+ return torch.ones(size=dims).to(device)