Tiresias/transfer/transfer.py

import os
import cv2
import torch
import numpy as np
import random
import matplotlib.pyplot as plt
from torch import nn
from transformers import AutoModelForVideoClassification, AutoProcessor, TrainingArguments, Trainer
from sklearn.model_selection import train_test_split
from torch.utils.data import Dataset
from PIL import Image

dataset_folder = os.getenv("dataset_folder", "/app/dataset_folder")
output_folder = os.getenv("output_folder", "/app/output_folder")
os.makedirs(dataset_folder, exist_ok=True)
os.makedirs(f"{output_folder}/base_model", exist_ok=True)
os.makedirs(f"{output_folder}/tuned", exist_ok=True)


model = AutoModelForVideoClassification.from_pretrained(f"{output_folder}/base_model")
processor = AutoProcessor.from_pretrained(f"{output_folder}/base_model")

# for param in model.base_model.parameters():
#     param.requires_grad = False

# Replace classifier with binary output
model.classifier = nn.Linear(model.classifier.in_features, 2)

class VideoNPZDataset(Dataset):
    def __init__(self, files, labels, processor):
        self.files = files
        self.labels = labels
        self.processor = processor

    def __len__(self):
        return len(self.files)

    def __getitem__(self, idx):
        file = self.files[idx]
        label = self.labels[idx]


        data = np.load(file)
        frames = [data[key] for key in sorted(data.files)]

        # Debug print
        # for i, frame in enumerate(frames):
        #     print(f"  Frame {i} shape: {frame.shape}, dtype: {frame.dtype}")

        # Convert to RGB (assumes frames are in BGR format)
        frames_rgb = [Image.fromarray(frame[:, :, ::-1]) for frame in frames]

        # for i, frame in enumerate(frames_rgb):
        #     print(f"  Frame {i} post-RGB shape: {frame.shape}")

        # Process the video frames
        pixel_values = self.processor(images=frames_rgb, return_tensors="pt")["pixel_values"][0]  # shape: [16, 3, 224, 224]
        label_values = torch.tensor(label, dtype=torch.long)
        # print(f"[{file}]")
        # print(f"  Processor output shape (before permute): {pixel_values.shape}")
        # print(f"  Processor label shape (before permute): {label_values.shape}")

        return {
            "pixel_values": pixel_values,  # [3, 16, 224, 224]
            "labels": label_values  # scalar tensor
        }

class myTrainer(Trainer):
    def compute_loss(self, model, inputs, return_outputs=False):
        print(f"[DEBUG] pixel_values: {inputs['pixel_values'].shape}")
        print(f"[DEBUG] labels: {inputs['labels'].shape}")

        return super().compute_loss(model, inputs, return_outputs)

def collate_fn(batch):
    pixel_values = []
    labels = []

    for item in batch:
        video = item["pixel_values"]  # shape: (16, 3, 224, 224)
        label = item["labels"]

        pixel_values.append(video)
        labels.append(label)

    pixel_values = torch.stack(pixel_values)  # (batch_size, 3, 16, 224, 224)
    labels_values = torch.tensor(labels, dtype=torch.long)  # (batch_size, 3, 16, 224, 224)

    return {
        "pixel_values": pixel_values,
        "labels": labels_values,
    }


def load_dataset_from_npz(root_dir):
    files = []
    labels = []
    groups = []
    label_map = {}  # for inference class map
    group_map = {}
    group_counter = 0
    for i, class_name in enumerate(sorted(os.listdir(root_dir))):
        label_map[i] = class_name
        class_dir = os.path.join(root_dir, class_name)
        for file in os.listdir(class_dir):
            if file.endswith(".npz"):
                group_id = f"{file.split('_')[0]}_{i}"
                if group_id not in group_map:
                    group_counter += 1
                    group_map[group_id] = group_counter

                groups.append(group_map[group_id])
                files.append(os.path.join(class_dir, file))
                labels.append(i)
    return files, labels, label_map, groups

files, labels, label_map, groups = load_dataset_from_npz(dataset_folder)

# for file, label in zip(files, labels):
#     print(f"{file} {label} {label_map[label]}")

print(f" files: {len(files)}")
print(f" labels: {len(labels)}")

# import torch
# print("CUDA available:", torch.cuda.is_available())
# print("Device count:", torch.cuda.device_count())
# print("Current device:", torch.cuda.current_device())
# print("Device name:", torch.cuda.get_device_name(0))

train_files, val_files, train_labels, val_labels = train_test_split(files, labels, test_size=0.2, stratify=groups, random_state=random.randint(1,5000))

train_dataset = VideoNPZDataset(train_files, train_labels, processor)
val_dataset = VideoNPZDataset(val_files, val_labels, processor)

def compute_metrics(eval_pred):
    logits, labels = eval_pred
    preds = torch.sigmoid(torch.tensor(logits)).numpy() > 0.5
    accuracy = (preds.flatten() == labels).mean()
    return {"accuracy": accuracy}

training_args = TrainingArguments(
    output_dir="./results",
    evaluation_strategy="epoch",
    save_strategy="epoch",
    per_device_train_batch_size=8,
    per_device_eval_batch_size=8,
    num_train_epochs=3,
    logging_dir="./logs",
    logging_steps=2,
    save_total_limit=2,
    load_best_model_at_end=True,
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    eval_dataset=val_dataset,
    data_collator=collate_fn,
)


trainer.train()

logs = trainer.state.log_history
train_loss = [x["loss"] for x in logs if "loss" in x]
eval_loss = [x["eval_loss"] for x in logs if "eval_loss" in x]

plt.plot(train_loss, label="Train Loss")
plt.plot(eval_loss, label="Eval Loss")
plt.xlabel("Log Steps")
plt.ylabel("Loss")
plt.legend()
plt.title("Loss Curve")
plt.savefig(f"{output_folder}/tuned/loss_curve.png")

trainer.save_model(f"{output_folder}/tuned")
processor.save_pretrained(f"{output_folder}/tuned")