From 2a6e2e8975961d93ab14a1a469cd73f421ea7f04 Mon Sep 17 00:00:00 2001
From: Thomas Rieutord <thomas.rieutord@gmail.com>
Date: Tue, 5 Sep 2023 12:29:36 +0000
Subject: [PATCH] Add folder to give answers to specific questions. Starting
 with architectures

---
 ..._the_architectures_of_the_tested_models.py | 98 +++++++++++++++++++
 1 file changed, 98 insertions(+)
 create mode 100644 questions/what_are_the_architectures_of_the_tested_models.py

diff --git a/questions/what_are_the_architectures_of_the_tested_models.py b/questions/what_are_the_architectures_of_the_tested_models.py
new file mode 100644
index 0000000..6bd13b4
--- /dev/null
+++ b/questions/what_are_the_architectures_of_the_tested_models.py
@@ -0,0 +1,98 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""Multiple land-cover/land-use Maps Translation (MMT)
+
+https://github.com/ThomasRieutord/MT-MLULC
+
+
+ QUESTION
+==========
+What are the architectures of the tested models?
+In particular, how many parameters do they have and what is their memory footprint?
+
+ ANSWER
+=========
+The BASELINE model has 350031 trainable parameters and needs 20.20513536 GB to run
+The CANDIDATE model has 355182 trainable parameters and needs 20.20956992 GB to run
+
+The BASELINE model is the original MLCnet proposed by Baudoux et al. (2022)
+The CANDIDATE model is the new version of model that includes attention layers
+
+Reproducibility
+---------------
+python what_are_the_architectures_of_the_tested_models.py
+"""
+import os
+import torch
+from torchinfo import summary
+
+from mmt.graphs.models import universal_embedding
+from mmt.graphs.models import transformer_embedding
+from mmt.utils import config as utilconf
+from mmt import _repopath_ as mmt_repopath
+
+# Configs
+#---------
+xp_name = "test_if_it_runs"
+mlulcconfig, _ = utilconf.get_config_from_json(
+    os.path.join(
+        mmt_repopath,
+        "configs",
+        "universal_embedding.json",
+    )
+)
+n_labels = 12
+
+baseline_model = universal_embedding.UnivEmb(
+    in_channels = n_labels + 1,
+    n_classes = n_labels + 1,
+    softpos = mlulcconfig.softpos,
+    number_feature_map = mlulcconfig.number_of_feature_map,
+    embedding_dim = mlulcconfig.embedding_dim[0],
+    memory_monger = mlulcconfig.memory_monger,
+    up_mode = mlulcconfig.up_mode,
+    num_groups = mlulcconfig.group_norm,
+    decoder_depth = mlulcconfig.decoder_depth,
+    mode = mlulcconfig.mode,
+    resize = 6,
+    cat=False,
+    pooling_factors = mlulcconfig.pooling_factors,
+    decoder_atrou = mlulcconfig.decoder_atrou,
+)
+baseline_model = torch.nn.Sequential(
+    baseline_model.encoder,
+    baseline_model.decoder,
+)
+candidate_model = transformer_embedding.TransformerEmbedding(
+    in_channels = n_labels + 1,
+    n_classes = n_labels + 1,
+    softpos = mlulcconfig.softpos,
+    number_feature_map = mlulcconfig.number_of_feature_map,
+    embedding_dim = mlulcconfig.embedding_dim[0],
+    memory_monger = mlulcconfig.memory_monger,
+    up_mode = mlulcconfig.up_mode,
+    num_groups = mlulcconfig.group_norm,
+    decoder_depth = mlulcconfig.decoder_depth,
+    mode = mlulcconfig.mode,
+    resize = 6,
+    cat=False,
+    pooling_factors = mlulcconfig.pooling_factors,
+    decoder_atrou = mlulcconfig.decoder_atrou,
+)
+candidate_model = torch.nn.Sequential(
+    candidate_model.encoder,
+    candidate_model.decoder,
+)
+
+x = torch.rand(mlulcconfig.train_batch_size, n_labels + 1, 100, 100)
+
+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")