RuntimeError: CUDA out of memory.

[taruntiwarihp]

(mdvc) root@ever:~/MDVC# python main.py --device_ids 0
log_path: ./log/0606173845
model_checkpoint_path: ./log/0606173845
Preparing dataset for train
Preparing dataset for val_1
Preparing dataset for val_2
Preparing dataset for val_1
using SubsAudioVideoGeneratorConcatLinearDoutLinear
initialization: xavier
Param Num: 178749320
17:42:32 train (0): 0%| | 1/1221 [00:01<39:18, 1.93s/it]
Traceback (most recent call last):
File "main.py", line 573, in
main(cfg)
File "main.py", line 276, in main
cfg.modality, cfg.use_categories
File "/root/MDVC/epoch_loop/run_epoch.py", line 308, in training_loop
pred = model(feature_stacks, caption_idx, masks)
File "/root/miniconda3/envs/mdvc/lib/python3.7/site-packages/torch/nn/modules/module.py", line 547, in call
result = self.forward(*input, **kwargs)
File "/root/miniconda3/envs/mdvc/lib/python3.7/site-packages/torch/nn/parallel/data_parallel.py", line 150, in forward
return self.module(*inputs[0], **kwargs[0])
File "/root/miniconda3/envs/mdvc/lib/python3.7/site-packages/torch/nn/modules/module.py", line 547, in call
result = self.forward(*input, **kwargs)
File "/root/MDVC/model/transformer.py", line 328, in forward
memory_video = self.encoder_video(src_video, src_mask)
File "/root/miniconda3/envs/mdvc/lib/python3.7/site-packages/torch/nn/modules/module.py", line 547, in call
result = self.forward(*input, **kwargs)
File "/root/MDVC/model/transformer.py", line 202, in forward
x = layer(x, src_mask)
File "/root/miniconda3/envs/mdvc/lib/python3.7/site-packages/torch/nn/modules/module.py", line 547, in call
result = self.forward(*input, **kwargs)
File "/root/MDVC/model/transformer.py", line 189, in forward
x = self.res_layers[0](x, sublayer0)
File "/root/miniconda3/envs/mdvc/lib/python3.7/site-packages/torch/nn/modules/module.py", line 547, in call
result = self.forward(*input, **kwargs)
File "/root/MDVC/model/transformer.py", line 152, in forward
res = sublayer(res)
File "/root/MDVC/model/transformer.py", line 186, in
sublayer0 = lambda x: self.self_att(x, x, x, src_mask)
File "/root/miniconda3/envs/mdvc/lib/python3.7/site-packages/torch/nn/modules/module.py", line 547, in call
result = self.forward(*input, **kwargs)
File "/root/MDVC/model/transformer.py", line 137, in forward
att = attention(Q, K, V, mask) # (B, H, seq_len, d_k)
File "/root/MDVC/model/transformer.py", line 101, in attention
sm_input = sm_input.masked_fill(mask == 0, -float('inf'))
RuntimeError: CUDA out of memory. Tried to allocate 16.00 MiB (GPU 0; 3.94 GiB total capacity; 3.43 GiB already allocated; 12.94 MiB free; 71.07 MiB cached)

[v-iashin]

The set of default hyper-parameters require a 12 GB GPU (1080Ti/2080Ti)

[taruntiwarihp]

@v-iashin
Any way to reduce the requirement of GPU in parameters, I have only 4GB GPU.
Thanks