| I have conducted some simple tests for qiqi, including the following: | |
| First, I added three sets of lora: | |
| 1. Same training set, removed regular datasets. | |
| 2. Same training set, removed regular datasets + modified dim to 32dim. | |
| 3. Introduced qiqi lora that I trained myself last time as a control group (250 images + 1000 steps * 8bs, definitely enough training). | |
| I first tested the original version, with a weight of 0.8. | |
| Then, I replaced the weight with 1 for testing on epoch 000018. | |
| Finally, I added "ofuda" at the beginning of the prompts. | |
| Please refer to the Excel file for details. | |
| In summary, I believe the reasons for this phenomenon consist of multiple components. | |
| 1.4dim itself is not easy to overfit. If a weight less than 1 is used simultaneously, it will cause the model show underfit. | |
| 2.The number of epochs given in the range of 1000 to 2000 images may be too small. | |
| In the case of no regular set, the best epoch maybe is 18/20, which is already at a critical point, and it is uncertain whether it is a one-time outlier. | |
| 3.Under 2, a large number of regular datasets were added, which slows down the fitting speed and requires more epochs. However, the epoch empirical formula is based on training without reg datasets. | |
| 4.The "ofuda" label should be a core tag, but it was not deleted, so if the ofuda prompt is not used, the ccip score will decrease. | |
| The results are a result of the combined effects of 1, 2, 3, 4. | |
| To address this issue, I suggest: | |
| 1.Change the testing weight to 1 | |
| 2.Increase the number of epochs for 1000 to 2000 images. Consider increasing it to 24 or even 26. | |
| 3.Further increase the number of epochs when introducing regular datasets. The specific amount needs to be tested (I have not extensively used regular training datasets, so I cannot be sure). | |
| 4.I don't think 4dim are insufficient, but if necessary, a slight increase to 6dim/8dim can be considered. | |