From d43df488d81aa6d787b8167733565989c615b2be Mon Sep 17 00:00:00 2001
From: Leonard Hackel <l.hackel@tu-berlin.de>
Date: Wed, 3 May 2023 08:18:45 +0200
Subject: [PATCH] Adding TransformerRSVQA code and update to configilm v0.3.0
to make code runable
---
conda_env.yml | 2 +-
fusion.py | 148 +++++++++++++++++++++
train_rsvqa.py | 350 +++++++++++++++++++++++++++++++++++++++++++++++++
3 files changed, 499 insertions(+), 1 deletion(-)
create mode 100644 fusion.py
create mode 100644 train_rsvqa.py
diff --git a/conda_env.yml b/conda_env.yml
index 49e1f6a..b92f7df 100644
--- a/conda_env.yml
+++ b/conda_env.yml
@@ -57,7 +57,7 @@ dependencies:
- pip:
- attrs==23.1.0
- colorama==0.4.6
- - configilm==0.2.0
+ - configilm==0.3.0
- cycler==0.11.0
- fonttools==4.39.3
- kiwisolver==1.4.4
diff --git a/fusion.py b/fusion.py
new file mode 100644
index 0000000..f215eeb
--- /dev/null
+++ b/fusion.py
@@ -0,0 +1,148 @@
+# from https://github.com/Cadene/vqa.pytorch/blob/master/vqa/models/fusion.py
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.autograd import Variable
+
+
+class AbstractFusion(nn.Module):
+
+ def __init__(self, opt={}):
+ super(AbstractFusion, self).__init__()
+ self.opt = opt
+
+ def forward(self, input_v, input_q):
+ raise NotImplementedError
+
+
+class MLBFusion(AbstractFusion):
+
+ def __init__(self, opt):
+ super(MLBFusion, self).__init__(opt)
+ # Modules
+ if 'dim_v' in self.opt:
+ self.linear_v = nn.Linear(self.opt['dim_v'], self.opt['dim_h'])
+ else:
+ print('Warning fusion.py: no visual embedding before fusion')
+
+ if 'dim_q' in self.opt:
+ self.linear_q = nn.Linear(self.opt['dim_q'], self.opt['dim_h'])
+ else:
+ print('Warning fusion.py: no question embedding before fusion')
+
+ def forward(self, input_v, input_q):
+ # visual (cnn features)
+ if 'dim_v' in self.opt:
+ x_v = F.dropout(input_v, p=self.opt['dropout_v'], training=self.training)
+ x_v = self.linear_v(x_v)
+ if 'activation_v' in self.opt:
+ x_v = getattr(F, self.opt['activation_v'])(x_v)
+ else:
+ x_v = input_v
+ # question (rnn features)
+ if 'dim_q' in self.opt:
+ x_q = F.dropout(input_q, p=self.opt['dropout_q'], training=self.training)
+ x_q = self.linear_q(x_q)
+ if 'activation_q' in self.opt:
+ x_q = getattr(F, self.opt['activation_q'])(x_q)
+ else:
+ x_q = input_q
+ # hadamard product
+ x_mm = torch.mul(x_q, x_v)
+ return x_mm
+
+
+class MutanFusion(AbstractFusion):
+
+ def __init__(self, opt, visual_embedding=True, question_embedding=True):
+ super(MutanFusion, self).__init__(opt)
+ self.visual_embedding = visual_embedding
+ self.question_embedding = question_embedding
+ # Modules
+ if self.visual_embedding:
+ self.linear_v = nn.Linear(self.opt['dim_v'], self.opt['dim_hv'])
+ else:
+ print('Warning fusion.py: no visual embedding before fusion')
+
+ if self.question_embedding:
+ self.linear_q = nn.Linear(self.opt['dim_q'], self.opt['dim_hq'])
+ else:
+ print('Warning fusion.py: no question embedding before fusion')
+
+ self.list_linear_hv = nn.ModuleList([
+ nn.Linear(self.opt['dim_hv'], self.opt['dim_mm'])
+ for i in range(self.opt['R'])])
+
+ self.list_linear_hq = nn.ModuleList([
+ nn.Linear(self.opt['dim_hq'], self.opt['dim_mm'])
+ for i in range(self.opt['R'])])
+
+ def forward(self, input_v, input_q):
+ if input_v.dim() != input_q.dim() and input_v.dim() != 2:
+ raise ValueError
+ batch_size = input_v.size(0)
+
+ if self.visual_embedding:
+ x_v = F.dropout(input_v, p=self.opt['dropout_v'], training=self.training)
+ x_v = self.linear_v(x_v)
+ if 'activation_v' in self.opt:
+ x_v = getattr(F, self.opt['activation_v'])(x_v)
+ else:
+ x_v = input_v
+
+ if self.question_embedding:
+ x_q = F.dropout(input_q, p=self.opt['dropout_q'], training=self.training)
+ x_q = self.linear_q(x_q)
+ if 'activation_q' in self.opt:
+ x_q = getattr(F, self.opt['activation_q'])(x_q)
+ else:
+ x_q = input_q
+
+ x_mm = []
+ for i in range(self.opt['R']):
+
+ x_hv = F.dropout(x_v, p=self.opt['dropout_hv'], training=self.training)
+ x_hv = self.list_linear_hv[i](x_hv)
+ if 'activation_hv' in self.opt:
+ x_hv = getattr(F, self.opt['activation_hv'])(x_hv)
+
+ x_hq = F.dropout(x_q, p=self.opt['dropout_hq'], training=self.training)
+ x_hq = self.list_linear_hq[i](x_hq)
+ if 'activation_hq' in self.opt:
+ x_hq = getattr(F, self.opt['activation_hq'])(x_hq)
+
+ x_mm.append(torch.mul(x_hq, x_hv))
+
+ x_mm = torch.stack(x_mm, dim=1)
+ x_mm = x_mm.sum(1).view(batch_size, self.opt['dim_mm'])
+
+ if 'activation_mm' in self.opt:
+ x_mm = getattr(F, self.opt['activation_mm'])(x_mm)
+
+ return x_mm
+
+
+class MutanFusion2d(MutanFusion):
+
+ def __init__(self, opt, visual_embedding=True, question_embedding=True):
+ super(MutanFusion2d, self).__init__(opt,
+ visual_embedding,
+ question_embedding)
+
+ def forward(self, input_v, input_q):
+ if input_v.dim() != input_q.dim() and input_v.dim() != 3:
+ raise ValueError
+ batch_size = input_v.size(0)
+ weight_height = input_v.size(1)
+ dim_hv = input_v.size(2)
+ dim_hq = input_q.size(2)
+ if not input_v.is_contiguous():
+ input_v = input_v.contiguous()
+ if not input_q.is_contiguous():
+ input_q = input_q.contiguous()
+ x_v = input_v.view(batch_size * weight_height, self.opt['dim_hv'])
+ x_q = input_q.view(batch_size * weight_height, self.opt['dim_hq'])
+ x_mm = super().forward(x_v, x_q)
+ x_mm = x_mm.view(batch_size, weight_height, self.opt['dim_mm'])
+ return x_mm
diff --git a/train_rsvqa.py b/train_rsvqa.py
new file mode 100644
index 0000000..48e92c1
--- /dev/null
+++ b/train_rsvqa.py
@@ -0,0 +1,350 @@
+# import packages
+import pytorch_lightning as pl
+import torch
+import torch.nn.functional as F
+from torch import optim
+from tqdm import tqdm
+
+from configilm import ConfigILM
+from configilm.ConfigILM import ILMConfiguration, ILMType
+from configilm.ConfigILM import get_hf_model as get_huggingface_model
+from configilm.extra.RSVQAxBEN_DataModule_LMDB_Encoder import RSVQAxBENDataModule
+from configilm.extra.BEN_lmdb_utils import resolve_ben_data_dir
+import typer
+import os
+from os.path import isfile
+import wandb
+from pytorch_lightning.loggers.wandb import WandbLogger
+from pytorch_lightning.callbacks import ModelCheckpoint
+from pytorch_lightning.callbacks import EarlyStopping
+from pytorch_lightning.callbacks import LearningRateMonitor
+from sklearn.metrics import accuracy_score
+from torchmetrics.classification import MultilabelF1Score
+from LinWarCosAnLR import LinearWarmupCosineAnnealingLR
+
+from fusion import MutanFusion
+
+__author__ = "Leonard Hackel - BIFOLD/RSiM TU Berlin"
+os.environ["WANDB_START_METHOD"] = "thread"
+wandb_api_key = os.environ["WANDB_API_KEY"]
+
+
+class LitVisionEncoder(pl.LightningModule):
+ """
+ Wrapper around a pytorch module, allowing this module to be used in automatic
+ training with pytorch lightning.
+ Among other things, the wrapper allows us to do automatic training and removes the
+ need to manage data on different devices (e.g. GPU and CPU).
+ """
+
+ def __init__(
+ self,
+ config: ConfigILM.ILMConfiguration,
+ lr: float = 1e-3,
+ ):
+ super().__init__()
+ self.lr = lr
+ self.config = config
+ self.model = ConfigILM.ConfigILM(config)
+
+ def _disassemble_batch(self, batch):
+ images, questions, labels = batch
+ # transposing tensor, needed for Huggingface-Dataloader combination
+ questions = torch.tensor(
+ [x.tolist() for x in questions], device=self.device
+ ).T.int()
+ return (images, questions), labels
+
+ def training_step(self, batch, batch_idx):
+ x, y = self._disassemble_batch(batch)
+ x_hat = self.model(x)
+ loss = F.binary_cross_entropy_with_logits(x_hat, y)
+ self.log("train/loss", loss)
+ return {"loss": loss}
+
+ def configure_optimizers(self):
+ optimizer = optim.AdamW(self.parameters(), lr=self.lr, weight_decay=0.01)
+
+ # these are steps if interval is set to step
+ max_intervals = int(self.trainer.max_epochs *
+ len(self.trainer.datamodule.train_ds) /
+ self.trainer.datamodule.batch_size)
+ warmup = 10000 if max_intervals > 10000 else 100 if max_intervals > 100 else 0
+
+ print(f"Optimizing for {max_intervals} steps with warmup for {warmup} steps")
+
+ lr_scheduler = {
+ 'scheduler': LinearWarmupCosineAnnealingLR(
+ optimizer,
+ warmup_epochs=warmup,
+ max_epochs=max_intervals,
+ warmup_start_lr=self.lr / 10,
+ eta_min=self.lr / 10
+ ),
+ 'name': 'learning_rate',
+ 'interval': "step",
+ 'frequency': 1
+ }
+ return [optimizer], [lr_scheduler]
+
+ def validation_step(self, batch, batch_idx):
+ x, y = self._disassemble_batch(batch)
+ x_hat = self.model(x)
+ loss = F.binary_cross_entropy_with_logits(x_hat, y)
+ return {"loss": loss, "outputs": x_hat, "labels": y}
+
+ def validation_epoch_end(self, outputs):
+ metrics = self.get_metrics(outputs)
+
+ self.log("val/loss", metrics["avg_loss"])
+ self.log("val/f1", metrics["avg_f1_score"])
+ self.log("val/Accuracy (LULC)", metrics["accuracy"]["LULC"])
+ self.log("val/Accuracy (Yes-No)", metrics["accuracy"]["Yes/No"])
+ self.log("val/Accuracy (Overall)", metrics["accuracy"]["Overall"])
+ self.log("val/Accuracy (Average)", metrics["accuracy"]["Average"])
+
+ def test_step(self, batch, batch_idx):
+ x, y = self._disassemble_batch(batch)
+ x_hat = self.model(x)
+ loss = F.binary_cross_entropy_with_logits(x_hat, y)
+ return {"loss": loss, "outputs": x_hat, "labels": y}
+
+ def test_epoch_end(self, outputs):
+ metrics = self.get_metrics(outputs)
+
+ self.log("test/loss", metrics["avg_loss"])
+ self.log("test/f1", metrics["avg_f1_score"])
+ self.log("test/Accuracy (LULC)", metrics["accuracy"]["LULC"])
+ self.log("test/Accuracy (Yes-No)", metrics["accuracy"]["Yes/No"])
+ self.log("test/Accuracy (Overall)", metrics["accuracy"]["Overall"])
+ self.log("test/Accuracy (Average)", metrics["accuracy"]["Average"])
+
+ def forward(self, batch):
+ # because we are a wrapper, we call the inner function manually
+ return self.model(batch)
+
+ def get_metrics(self, outputs):
+ avg_loss = torch.stack([x["loss"] for x in outputs]).mean()
+ logits = torch.cat([x["outputs"].cpu() for x in outputs], 0)
+ labels = torch.cat(
+ [x["labels"].cpu() for x in outputs], 0
+ ) # Tensor of size (#samples x classes)
+
+ selected_answers = self.trainer.datamodule.selected_answers
+
+ argmax_out = torch.argmax(logits, dim=1)
+ argmax_lbl = torch.argmax(labels, dim=1)
+
+ # get answers and predictions per type
+ yn_preds = []
+ yn_gts = []
+ lulc_preds = []
+ lulc_gts = []
+
+ for i, ans in enumerate(tqdm(argmax_lbl, desc="Counting answers")):
+ # Yes/No question
+ if selected_answers[ans] in ["yes", "no"]:
+
+ # stored for global Yes/No
+ yn_preds.append(argmax_out[i])
+ yn_gts.append(ans)
+
+ # LC question
+ else:
+ # stored for global LC
+ lulc_preds.append(argmax_out[i])
+ lulc_gts.append(ans)
+
+ acc_yn = accuracy_score(yn_gts, yn_preds)
+ acc_lulc = accuracy_score(lulc_gts, lulc_preds)
+
+ accuracy_dict = {
+ "Yes/No": acc_yn,
+ "LULC": acc_lulc,
+ "Overall": accuracy_score(
+ argmax_lbl, argmax_out
+ ), # micro average on classes
+ "Average": (acc_yn + acc_lulc) / 2, # macro average on types
+ }
+
+ f1_score = MultilabelF1Score(num_labels=self.config.classes, average=None).to(
+ logits.device
+ )(logits, labels)
+
+ avg_f1_score = float(
+ torch.sum(f1_score) / self.config.classes
+ ) # macro average f1 score
+
+ return {
+ "avg_loss": avg_loss,
+ "avg_f1_score": avg_f1_score,
+ "accuracy": accuracy_dict,
+ }
+
+
+def overwrite_vision_weights(model, vision_checkpoint):
+ if vision_checkpoint is None:
+ return model
+ if not isfile(vision_checkpoint):
+ print("Pretrained vision model not available, cannot load checkpoint")
+ return model
+ # load weights
+ # get model and pretrained state dicts
+ if torch.cuda.is_available():
+ pretrained_dict = torch.load(vision_checkpoint)
+ else:
+ pretrained_dict = torch.load(
+ vision_checkpoint, map_location=torch.device("cpu")
+ )
+ model_dict = model.state_dict()
+
+ # filter out unnecessary keys
+ # this allows to load lightning or pytorch model loading
+ if "pytorch-lightning_version" in pretrained_dict.keys():
+ # checkpoint is a Pytorch-Lightning Checkpoint
+ pretrained_dict = {
+ k: v
+ for k, v in pretrained_dict["state_dict"].items()
+ if k in model_dict
+ }
+ else:
+ pretrained_dict = {
+ k: v for k, v in pretrained_dict.items() if k in model_dict
+ }
+
+ # filter keys that have a size mismatch
+ mismatch_keys = [
+ x
+ for x in pretrained_dict.keys()
+ if pretrained_dict[x].shape != model_dict[x].shape
+ ]
+ for key in mismatch_keys:
+ del pretrained_dict[key]
+ print(f"Key '{key}' size mismatch, removing from loading")
+
+ # overwrite entries in the existing state dict
+ model_dict.update(pretrained_dict)
+
+ # load the new state dict
+ model.load_state_dict(model_dict)
+ print("Vision Model checkpoint loaded")
+ return model
+
+
+def mutan(fusion_in: int, fusion_out: int):
+ opt = {
+ 'dim_hv': fusion_in,
+ 'dim_hq': fusion_in,
+ 'dim_mm': fusion_out,
+ 'dropout_hv': 0.1,
+ 'dropout_hq': 0.1,
+ 'R': 1
+ }
+ device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+ f = MutanFusion(visual_embedding=False,
+ question_embedding=False,
+ opt=opt).to(device)
+ return f
+
+
+def main(
+ vision_model: str = "resnet152",
+ text_model: str = "bert-base-uncased",
+ lr: float = 1e-3,
+ epochs: int = 100,
+ batch_size: int = 32,
+ seed: int = 42,
+ data_dir: str = None,
+ test_run: bool = False,
+ num_workers_dataloader: int = 4,
+ vision_checkpoint: str = None
+):
+ if test_run:
+ max_img_index = 10 * batch_size
+ epochs = 10
+ else:
+ max_img_index = -1
+
+ pl.seed_everything(seed, workers=True)
+
+ img_size = 120
+ channels = 10
+
+ fusion_in = 1200
+ fusion_out = 360
+ model_config = ILMConfiguration(
+ timm_model_name=vision_model,
+ hf_model_name=text_model,
+ classes=1000,
+ image_size=img_size,
+ channels=channels,
+ network_type=ILMType.VQA_CLASSIFICATION,
+ visual_features_out=2048,
+ fusion_in=fusion_in,
+ fusion_out=fusion_out,
+ fusion_method=mutan(fusion_in=fusion_in, fusion_out=fusion_out),
+ fusion_hidden=256
+ )
+
+ # Key is available by wandb, project name can be chosen at will
+ wandb.login(key=wandb_api_key)
+
+ tags = ["Training", vision_model, text_model]
+ if test_run:
+ tags += ["Test Run"]
+ wandb_logger = WandbLogger(project=f"LiT4RSVQA",
+ log_model=True,
+ tags=tags, # keyword arg directly to wandb.init()
+ )
+
+ monitor = "val/f1"
+ monitor_str = "F1_score"
+ # checkpointing
+ checkpoint_callback = ModelCheckpoint(
+ monitor="val/f1",
+ dirpath="./checkpoints",
+ filename=f"{wandb_logger.experiment.name}-{vision_model}-{text_model}-seed=" +
+ str(seed) + "-epoch={epoch:03d}-" + f"{monitor_str}" + "={" +
+ f"{monitor}" + ":.3f}",
+ auto_insert_metric_name=False,
+ save_top_k=1,
+ mode="max",
+ save_last=True
+ )
+ early_stopping_callback = EarlyStopping(monitor=monitor, min_delta=0.00,
+ patience=25, verbose=False, mode="max")
+ lr_monitor = LearningRateMonitor(logging_interval='step')
+
+ trainer = pl.Trainer(
+ max_epochs=epochs,
+ accelerator="auto",
+ log_every_n_steps=5,
+ logger=wandb_logger,
+ check_val_every_n_epoch=5,
+ callbacks=[checkpoint_callback, early_stopping_callback, lr_monitor],
+
+ )
+
+ model = LitVisionEncoder(config=model_config, lr=lr)
+ model = overwrite_vision_weights(model, vision_checkpoint)
+
+ hf_tokenizer, _ = get_huggingface_model(
+ model_name=text_model, load_pretrained_if_available=False
+ )
+ dm = RSVQAxBENDataModule(
+ data_dir=resolve_ben_data_dir(data_dir=data_dir, force_mock=True),
+ img_size=(channels, img_size, img_size),
+ num_workers_dataloader=num_workers_dataloader,
+ batch_size=batch_size,
+ max_img_idx=max_img_index,
+ )
+
+ trainer.fit(model=model, datamodule=dm)
+ trainer.test(model=model, datamodule=dm, ckpt_path="best")
+
+ wandb.finish()
+ print("=== Training finished ===")
+
+
+if __name__ == "__main__":
+ typer.run(main)
--
GitLab