sync remote changes

cgan_earth/code_files/model.py

@@ -6,7 +6,7 @@
 
 def train(pth, gen, disc, imgs, labels, img_length, n_classes, num_epochs, z_dim, batch_size, lr, device, wandb_name):
 
-    rt = 1
+    rt = 0
     lz = 6
     beta1 = 0.5
     beta2 = 0.999

cgan_earth/code_files/networks.py

@@ -8,14 +8,14 @@ def cgan_earth_nets(path, Training, g_dim, d_dim):
 
     hidden_dim = 64
     if Training == True:
-        layers_g = [g_dim, hidden_dim*8, hidden_dim*8, hidden_dim*8, hidden_dim*8, hidden_dim*8, 3]
+        layers_g = [g_dim, hidden_dim*32, hidden_dim*16, hidden_dim*8, hidden_dim*8, hidden_dim*8, 3]
         kernel_g = [4, 4, 4, 4, 4, 3]
         stride_g = [2, 2, 2, 2, 2, 1]
         pad_g = [2, 2, 2, 2, 2, 0]
-        layers_d = [d_dim, hidden_dim*8, hidden_dim*8, hidden_dim*8, hidden_dim*8, hidden_dim*8, 1]
+        layers_d = [d_dim, hidden_dim*8, hidden_dim*8, hidden_dim*8, hidden_dim*16, hidden_dim*32, 1]
         kernel_d = [4, 4, 4, 4, 4, 4]
         stride_d = [2, 2, 2, 2, 2, 1]
-        pad_d = [2, 2, 2, 2, 1, 0]
+        pad_d = [1, 1, 1, 1, 1, 0]
         params = [layers_g, kernel_g, stride_g, pad_g, layers_d, kernel_d, stride_d, pad_d]
         with open(path + '_params.data', 'wb') as filehandle:
             # store the data as binary data stream
@@ -35,9 +35,9 @@ def __init__(self, g_dim, img_length, im_chan=3, hidden_dim=64):
             self.final_conv = nn.Conv2d(hidden_dim * 8, im_chan, 3, 1, 0)
             # Build the neural network
             self.gen = nn.Sequential(
-                self.make_gen_block(g_dim, hidden_dim * 8),
-                self.make_gen_block(hidden_dim * 8, hidden_dim * 8),
-                self.make_gen_block(hidden_dim * 8, hidden_dim * 8),
+                self.make_gen_block(g_dim, hidden_dim * 32),
+                self.make_gen_block(hidden_dim * 32, hidden_dim * 16),
+                self.make_gen_block(hidden_dim * 16, hidden_dim * 8),
                 self.make_gen_block(hidden_dim * 8, hidden_dim * 8),
                 self.make_gen_block(hidden_dim * 8, hidden_dim * 8)
             )
@@ -85,12 +85,12 @@ def __init__(self, d_dim, hidden_dim=64):
                 self.make_crit_block(d_dim, hidden_dim * 8),
                 self.make_crit_block(hidden_dim * 8, hidden_dim * 8),
                 self.make_crit_block(hidden_dim * 8, hidden_dim * 8),
-                self.make_crit_block(hidden_dim * 8, hidden_dim * 8),
-                self.make_crit_block(hidden_dim * 8, hidden_dim * 8, padding=1),
-                self.make_crit_block(hidden_dim * 8, 1, stride=1, final_layer=True),
+                self.make_crit_block(hidden_dim * 8, hidden_dim * 16),
+                self.make_crit_block(hidden_dim * 16, hidden_dim * 32),
+                self.make_crit_block(hidden_dim * 32, 1, stride=1, final_layer=True),
             )
 
-        def make_crit_block(self, input_channels, output_channels, kernel_size=4, stride=2, padding=2, final_layer=False):
+        def make_crit_block(self, input_channels, output_channels, kernel_size=4, stride=2, padding=1, final_layer=False):
             '''
             Function to return a sequence of operations corresponding to a critic block of DCGAN;
             a convolution, a batchnorm (except in the final layer), and an activation (except in the final layer).

cgan_earth/run_cgan_earth.py

@@ -3,15 +3,15 @@
 import os
 
 PATH = os.path.dirname(os.path.realpath(__file__))
-Project_name = 'earth_cylinder_t'
+Project_name = 'earth_cylinder_r'
 Project_dir = PATH + '/trained_generators/'
 wandb_name = Project_name
 
 # import training images and define labels
 data_path = []
 labels = []
 
-for img_path, label in zip(['forest1'], [0]):
+for img_path, label in zip(['sea1'], [0]):
     file = PATH + '/earth_screenshots/{}.jpg'.format(img_path)
     data_path.append(file) # path to training data
     labels.append(label)
@@ -21,7 +21,7 @@
 # define hyperparameters and architecture
 ngpu = 1
 z_dim = 64
-lr = 0.0001
+lr = 0.0002
 Training = 1
 n_classes = 1
 batch_size = 2