def contrastive_loss(encoding: torch.Tensor, positive_sample: torch.Tensor,
negative_samples: torch.Tensor) -> torch.Tensor:
"""
Implements the contrastive loss defined in the CMC paper
Defined for batch size N, embedding size D, and num_negative_samples K
:param encoding: The encoded vector to contrast against samples (N x 1 x D)
:param positive_sample: The positive sample to contrast against (N x 1 X D)
:param negative_samples: The negative sample to contrast against (N x K x D)
:return: The contrastive loss (softmax with correct label in positive location)
"""
assert len(encoding.size()) == 3, "Expecting encoding shape: (N x 1 x D)"
assert len(positive_sample.size()) == 3, "Expecting positive sample shape: (N x 1 x D)"
assert len(negative_samples.size()) == 3, "Expecting negative sample shape: (N x K x D)"
# Stack positive sample on top of negative
all_samples = torch.cat([positive_sample, negative_samples], dim=1)
# Compute the "critic" scores from 3.2 Implementing the Critic
scores = torch.bmm(all_samples, torch.transpose(encoding, 1, 2)).squeeze(-1)
# Compute the contrastive loss
targets = torch.zeros(scores.size()[0]).long().to(util.get_project_device())
return F.cross_entropy(scores, targets)
def __init__(self,
model: nn.Module,
loss_function: Callable[[torch.Tensor, torch.Tensor],
torch.Tensor],
train_data: DataLoader,
validation_data: Optional[DataLoader] = None,
optimizer: Optional = None,
per_epoch_callbacks: List[Callable] = [],
device: torch.device = None,
checkpoint_file: str = ""):
"""
Trains a model given the data, loss function, and possibly and optimizer
:param model: The model to train
:param loss_function: The loss function to evaluate for training
:param train_data: The dataloader to use for training the model
:param validation_data: The dataloader used for model validation
:param optimizer: [Optional] The optimizer used to update weights of the model
:param per_epoch_callbacks: A set of callbacks to be called every epoch. Will be passed
the model, data loaders, optimizer, and wandb run. May be used for (e.g. logging images)
:param device: The device ((C/G/T)PU) to train and validate on
:param checkpoint_file: The location of the saved run to load the model and optimizer from
"""
# Create optimizer if none passed
if optimizer is None:
self.optimizer = torch.optim.Adam(model.parameters(recurse=True),
lr=0.001)
else:
self.optimizer = optimizer
if device is None:
self.device = util.get_project_device()
else:
self.device = device
util.set_project_device(self.device)
# Create a learning rate scheduler
self.scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
self.optimizer)
# Save the other class parameters
self.model = model.to(self.device)
self.loss_function = loss_function
self.train_data = train_data
self.validation_data = validation_data
self.callbacks = per_epoch_callbacks
# Load the model if checkpoint is given
if checkpoint_file != "":
model_save_dict, optimizer_save_dict = model_io.load_model_checkpoint(
checkpoint_file)
self.model.load_state_dict(model_save_dict)
if optimizer_save_dict is not None:
self.optimizer.load_state_dict(optimizer_save_dict)
# Initialize wandb and watch model
self.wandb_run = wandb.init(project='RosettaCV',
entity='cal-launchpad')
wandb.watch(self.model)
def get_cmc_loss_on_dataloader(model: nn.Module, dataloader: DataLoader, loss_fn: Callable) -> torch.Tensor:
"""
Gets the loss on a dataloader with a different assumed loss signature to that in util.util
:param model: The model to evaluate on
:param dataloader: The dataloader to evaluate on
:param loss_fn: The loss metric to use
:return: The loss function applied to the dataloader
"""
# The device the model is on
device = util.get_project_device()
total_loss = torch.FloatTensor([0]).to(device)
# Get all the encodings
for batch in dataloader:
# Send to GPU if available
for i, view in enumerate(batch):
if isinstance(view, torch.Tensor):
batch[i] = view.to(device)
# Get the positive and negative samples from the batch
encodings = model(batch, no_cache=True)
enc, pos, neg = get_positive_and_negative_samples(encodings, model)
total_loss += loss_fn(enc, pos, neg)
return total_loss / len(dataloader)
def decoding_loss(self, inputs: List[torch.Tensor],
encodings: List[torch.Tensor]) -> torch.Tensor:
"""
Samples decoding pathways and adds decoding loss to contrastive loss
:param inputs: The inputs to the CMC encoders
:param encodings: The encodings that were produced by the model
:return: A loss on the decodings sampled
"""
# Sample decoding pathways
decodings = random.sample(self.encode_decode_pairs,
k=self.num_decodings)
# Perform all relevant decodings
decoding_loss = torch.Tensor([0]).to(util.get_project_device())
for encode_view_ind, decode_view_ind in decodings:
decoded_view = self.model.views[decode_view_ind]
# decoded_view_out = decoded_view.decode(encodings[encode_view_ind])
decoded_view_out = decoded_view.decode(encodings[encode_view_ind],
inputs[decode_view_ind])
reconstruction_loss = decoded_view.reconstruction_loss(
decoded_view_out, inputs[decode_view_ind])
# Report this reconstruction loss
self.wandb_run.log({
f"{self.model.views[encode_view_ind].get_id()} -> "
f"{self.model.views[decode_view_ind].get_id()} Reconstruction Loss":
reconstruction_loss
})
decoding_loss += reconstruction_loss
return decoding_loss
def forward(self, x):
# Tokenize the text
tokenized = self.tokenizer(x, padding=True,
return_tensors="pt")['input_ids']
tokenized = tokenized.to(util.get_project_device())
# Forward pass
model_outputs = self.model(tokenized).last_hidden_state
# Average over sequence dimension
model_outputs = torch.mean(model_outputs, dim=-2)
model_outputs = model_outputs.view(model_outputs.shape[0], -1)
return F.relu(self.linear(model_outputs))
def language_reconstruction_loss(predicted_logits: torch.Tensor, ground_truth_caption: Union[str, List]) -> torch.Tensor:
"""
The language reconstruction loss, to compare a generated caption to a ground truth
:param predicted_logits: The logits over the vocabulary predicted by the language decoder model
:param ground_truth_caption: The actual caption for the image
:return: The cross entropy loss of the caption and the predicted values
"""
# Tokenize the caption into BERT's vocabulary
if type(ground_truth_caption) == str:
ground_truth_tok = tokenizer.encode(ground_truth_caption, padding=True, return_tensors="pt")
else:
ground_truth_tok = tokenizer.batch_encode_plus(ground_truth_caption, padding=True, return_tensors="pt")['input_ids']
# Reshape ground truth and predicted
predicted_logits = predicted_logits.view(-1, predicted_logits.shape[-1]) # Shape[-1] is the vocab size
ground_truth_tok = ground_truth_tok.view(-1).to(util.get_project_device())
# Cross entropy ignoring locations that are padded
return F.cross_entropy(predicted_logits, ground_truth_tok, ignore_index=PAD_TOKEN)
def forward(self, latent_encoding: torch.Tensor,
decoder_inputs: torch.Tensor) -> torch.Tensor:
# Change dimension of the input to match cross-attention in BERT
latent_encoding = F.relu(self.linear(latent_encoding))
# Tokenize the inputs
decoder_inputs = self.tokenizer(decoder_inputs,
return_tensors="pt",
padding=True).input_ids
decoder_inputs = decoder_inputs.to(util.get_project_device())
# Replicate the latent embedding to mimic an encoder output
sequence_length = decoder_inputs.size()[1]
latent_encoding = latent_encoding.unsqueeze(1)
encoder_output = torch.tile(latent_encoding, (1, sequence_length, 1))
# Generate logits for prediction
return self.decoder_model(decoder_inputs,
encoder_hidden_states=encoder_output).logits
def __init__(self,
views: List[View],
latent_dim: int,
memory_bank_size: int = 200):
super(WrapperModel, self).__init__()
assert len(views) >= 2, "Must specify at least 2 views!"
# Assign views and register submodules
self.views = views
self.view_encoders = nn.ModuleList([view.encoder for view in views])
self.view_decoders = nn.ModuleList(
[view.decoder for view in views if view.decoder is not None])
# Build the memory bank to sample from
model_device = util.get_project_device()
self.memory_bank = deque()
self.memory_bank.extend([
rand_vec.view(-1, latent_dim)
for rand_vec in torch.randn((memory_bank_size,
latent_dim)).to(model_device)
])
from torchvision import models, datasets
from torchvision.transforms import Compose, ToTensor, Resize
from data_loader.MultiviewDatasets import MultiviewDataset, identity_view, get_coco_captions, get_grayscale_view, get_inverted_view, get_complementary_view, get_saturated_view
from torch.utils.data import DataLoader
resnet_feature_size = 512
if __name__ == "__main__":
from models.CMC.ResNetEncoder import ResNetEncoder
from models.CMC.Decoder import Decoder
from models.CMC.language_models import TextEncoder, TextDecoder
# Ensure aspect ratio is 0.75
image_size = (480, 640)
device = util.get_project_device()
latent_dim = 512
# Define encoders, decoders, and views
image_encoder = ResNetEncoder(device, latent_dim=latent_dim)
image_decoder = Decoder(*image_size)
image_view = View(image_encoder,
image_decoder,
"Image",
reconstruction_loss=l2_reconstruction_loss)
# grayscale_encoder = ResNetEncoder(device, latent_dim=latent_dim)
# grayscale_decoder = Decoder(*image_size)
# grayscale_view = View(grayscale_encoder, grayscale_decoder, "Grayscale", reconstruction_loss=l2_reconstruction_loss)
inverted_encoder = ResNetEncoder(device, latent_dim=latent_dim)
