Solves issue #3

.gitignore

@@ -1 +1,2 @@
 pixel-nerf/checkpoints
+*.pyc

pixel-nerf/conf/exp/srn.conf

@@ -1,5 +1,108 @@
 # SRN experiments config
-include required("../default_mv.conf")
+# Single-view only base model
+# (Not used in experiments; resnet_fine_mv.conf inherits)
+model {
+    # MLP architecture
+    # Adapted for multiview
+    # Possibly too big
+
+    # Condition on local encoder
+    use_encoder = True
+
+    # Condition also on a global encoder?
+    use_global_encoder = False
+
+    # Use xyz input instead of just z
+    # (didn't ablate)
+    use_xyz = True
+
+    # Canonical space xyz (default view space)
+    canon_xyz = False
+
+    # Positional encoding
+    use_code = True
+    code {
+        num_freqs = 6
+        freq_factor = 1.5
+        include_input = True
+    }
+
+    # View directions
+    use_viewdirs = True
+    # Apply pos. enc. to viewdirs?
+    use_code_viewdirs = False
+
+    # MLP architecture
+    mlp_coarse {
+        type = resnet  # Can change to mlp
+        n_blocks = 3
+        d_hidden = 512
+        # Combine after 3rd layer by average
+        combine_layer = 3
+        combine_type = average
+    }
+    mlp_fine {
+        type = resnet
+        n_blocks = 3
+        d_hidden = 512
+        combine_layer = 3
+        combine_type = average
+    }
+
+    # Encoder architecture
+    encoder {
+        backbone = resnet34
+        pretrained = True
+        num_layers = 4
+    }
+}
+renderer {
+    n_coarse = 64
+    n_fine = 32
+    # Try using expected depth sample
+    n_fine_depth = 16
+    # Noise to add to depth sample
+    depth_std = 0.01
+    # Decay schedule, not used
+    sched = []
+    # White background color (false : black)
+    white_bkgd = True
+}
+loss {
+    # RGB losses coarse/fine
+    rgb {
+        use_l1 = False
+    }
+    rgb_fine {
+        use_l1 = False
+    }
+    # Alpha regularization (disabled in final version)
+    alpha {
+        # lambda_alpha = 0.0001
+        lambda_alpha = 0.0
+        clamp_alpha = 100
+        init_epoch = 5
+    }
+    # Coarse/fine weighting (nerf = equal)
+    lambda_coarse = 1.0  # loss = lambda_coarse * loss_coarse + loss_fine
+    lambda_fine = 1.0  # loss = lambda_coarse * loss_coarse + loss_fine
+}
+train {
+    # Training 
+    print_interval = 2
+    save_interval = 50
+    vis_interval = 100
+    eval_interval = 50
+
+    # Accumulating gradients. Not really recommended.
+    # 1 = disable
+    accu_grad = 1
+
+    # Number of times to repeat dataset per 'epoch'
+    # Useful if dataset is extremely small, like DTU
+    num_epoch_repeats = 1
+}
+
 data {
     format = srn
 }

pixel-nerf/environment.yml

@@ -7,7 +7,8 @@ dependencies:
   - python>=3.8
   - pip
   - pip:
-      - pyhocon
+      - pyhocon==0.3.6
+      - pyparsing==2.3.1
       - opencv-python
       - dotmap
       - tensorboard
@@ -19,7 +20,7 @@ dependencies:
   - scipy
   - numpy
   - matplotlib
-  - pytorch==1.6.0
-  - torchvision==0.7.0
+  - pytorch==1.8.0
+  - torchvision==0.9.0
   - scikit-image==0.17.2
   - tqdm

pixel-nerf/src/model/models.py

@@ -265,7 +265,7 @@ def forward(self, xyz, coarse=True, viewdirs=None, far=False):
             output = output.reshape(SB, B, -1)
         return output
 
-    def load_weights(self, args, opt_init=False, strict=True, device=None):
+    def load_weights(self, args, opt_init=False, strict=False, device=None):
         """
         Helper for loading weights according to argparse arguments.
         Your can put a checkpoint at checkpoints/<exp>/pixel_nerf_init to use as initialization.