Minor changes

drafts/mlctnet_simple_inference.py

@@ -15,6 +15,7 @@
 import sys
 import torch
 from torch import nn
+from torchinfo import summary
 from torchgeo.datasets.utils import BoundingBox
 import numpy as np
 import rasterio.crs
@@ -169,7 +170,18 @@
 
 # Show results
 #-----------
-
+inner_shape = [min(s1,s2) for (s1,s2) in zip(y.shape[1:], y_true["mask"].shape[1:])]
+ccrop = transforms.t.CenterCrop(size=inner_shape)
+acc = (ccrop(y) == ccrop(y_true["mask"])).sum()/ccrop(y).numel()
+print(f"Overall accuracy over this patch: {acc}")
 out_lc.plot({"mask":y})
 plt.show(block=False)
 
+# Export
+#-----------
+model = nn.Sequential(
+    esawc_autoencoder.encoder,
+    esgp_autoencoder.decoder
+)
+summary(model, x.shape)
+# torch.onnx.export(model, "esawc2esgp_test.onnx", verbose=True)

mmt/graphs/models/universal_embedding.py

@@ -14,7 +14,6 @@
 class AtrouMMU(nn.Module):
     def __init__(self, inf, scale_factor=10, bias=False):
         super(AtrouMMU, self).__init__()
-        print(inf, scale_factor)
         self.conv1 = nn.Sequential(
             nn.Conv2d(
                 inf, inf, kernel_size=3, padding=1, stride=scale_factor, bias=bias
@@ -365,6 +364,19 @@ def MemoryMonged_forward(self, x):
 
     def forward(self, x):
         return self.forward_method(x)
+
+    def check_shapes(self, x = None):
+        """Display shapes of some tensors"""
+        if x is None:
+            x = torch.rand(10, 3, 60, 120)
+            print(f"Random input: x = {x.shape}")
+        else:
+            print(f"Given input: x = {x.shape}")
+
+        states = self.encoder_part(x)
+        for i, xx in enumerate(states):
+            print(f"x{i} = {xx.shape}")
+
 
 
 class UnivEmb(nn.Module):

questions/what_are_the_architectures_of_the_tested_models.py

@@ -41,7 +41,8 @@
         "universal_embedding.json",
     )
 )
-n_labels = 12
+n_labels = 43
+resize = None
 
 baseline_model = universal_embedding.UnivEmb(
     in_channels = n_labels + 1,
@@ -54,7 +55,7 @@
     num_groups = mlulcconfig.group_norm,
     decoder_depth = mlulcconfig.decoder_depth,
     mode = mlulcconfig.mode,
-    resize = 6,
+    resize = resize,
     cat=False,
     pooling_factors = mlulcconfig.pooling_factors,
     decoder_atrou = mlulcconfig.decoder_atrou,
@@ -74,7 +75,7 @@
     num_groups = mlulcconfig.group_norm,
     decoder_depth = mlulcconfig.decoder_depth,
     mode = mlulcconfig.mode,
-    resize = 6,
+    resize = resize,
     cat=False,
     pooling_factors = mlulcconfig.pooling_factors,
     decoder_atrou = mlulcconfig.decoder_atrou,
@@ -84,15 +85,14 @@
     candidate_model.decoder,
 )
 
-x = torch.rand(mlulcconfig.train_batch_size, n_labels + 1, 100, 100)
+x = torch.rand(15, n_labels + 1, 600, 600)
 
 print("     Architecture of the BASELINE model")
 baseline_summary = summary(baseline_model, x.shape)
-print(baseline_summary)
 
 print("\n     Architecture of the CANDIDATE model")
 candidate_summary = summary(candidate_model, x.shape)
-print(candidate_summary)
+
 print(" ")
 print(f"The BASELINE model has {baseline_summary.trainable_params} trainable parameters and needs {baseline_summary.total_output_bytes/10**9} GB to run")
 print(f"The CANDIDATE model has {candidate_summary.trainable_params} trainable parameters and needs {candidate_summary.total_output_bytes/10**9} GB to run")

tests/import_test.py

@@ -44,6 +44,7 @@
 import mmt.graphs
 import mmt.graphs.models
 import mmt.graphs.models.custom_layers
+from mmt import _repopath_ as mmt_repopath
 
 from mmt.datasets import landcover_to_landcover
 landcover_to_landcover.LandcoverToLandcoverDataLoader
@@ -61,6 +62,8 @@
 
 from mmt.utils import plt_utils
 plt_utils.plt_loss2
+from mmt.utils import config as utilconf
+utilconf.get_config_from_json
 
 print("All imports passed successfully")
 print(f"Package {mmt.__name__}-{mmt.__version__} from {mmt._repopath_}")